Merge remote-tracking branch 'origin/dev' into feature/ai-difficulty

Every game with computer opponents gains a setup choice (Easy, Medium, Hard,
Mixed), carried in TableConfig.aiDifficulty and through the host lobby. Each
computer seat gets a level and a hidden style (balanced, aggressive, cautious,
spoiler) derived from the table seed, so saves and hosted tables reproduce it.
Medium plays exactly as each AI did before. Easy plays weaker and slips
plausibly; Hard searches deeper or counts cards, never using hidden
information. Covers Chess, Checkers, Morris, Backgammon, Pirates, Katan (base,
Cities & Knights, Seafarers), Euchre, Kaiser, Wizard, Hearts, President, Crazy
Eights, Golf, Cribbage, Sequence, Tayu and Yahtzee.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01TWwLCrXa9aFVg8dZsCdLBZ

app/src/main/java/org/prolibertate/games/game/backgammon/BackgammonAi.kt

@@ -1,6 +1,9 @@
 package org.prolibertate.games.game.backgammon
 
 import kotlin.random.Random
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
@@ -22,8 +25,19 @@ import org.prolibertate.games.game.engine.GameAi
  * out and its first step returned; the rest follows on the moves after, because
  * the same calculation reaches the same turn.
  */
-class BackgammonAi(private val level: BackgammonLevel? = null) :
-    GameAi<BackgammonState, BackgammonMove> {
+class BackgammonAi(
+    private val profiles: AiProfiles = AiProfiles.DEFAULT,
+    private val level: BackgammonLevel? = null,
+) : GameAi<BackgammonState, BackgammonMove> {
+
+    /**
+     * The turn an easy seat has settled on, so that a deliberate slip is made
+     * once per roll rather than re-rolled on every checker (which would stitch
+     * halves of different turns together).
+     */
+    private var plan: List<BackgammonMove> = emptyList()
+    private var planned = 0
+    private var planRoll = -1
 
     override fun chooseMove(
         state: BackgammonState,
@@ -31,48 +45,104 @@ class BackgammonAi(private val level: BackgammonLevel? = null) :
         legal: List<BackgammonMove>,
     ): BackgammonMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
-        val setting = level ?: state.options.level
         if (legal.size == 1) return legal.first()
+        val profile = profiles.of(seat)
+        val base = level ?: state.options.level
+        val setting = when (profile.level) {
+            AiLevel.MEDIUM -> base
+            // Easy never looks at the dice to come.
+            AiLevel.EASY -> if (base == BackgammonLevel.STRONG) BackgammonLevel.CLUB else base
+            AiLevel.HARD -> BackgammonLevel.STRONG
+        }
+        val tuning = tuningFor(profile)
 
         val random = Random(state.rollsMade * 31L + state.moveLog.size)
 
         if (setting == BackgammonLevel.CASUAL) {
             // Not thoughtless: it still prefers to hit, to come in off the bar
             // and to bear off, which is most of what a beginner knows.
-            val scored = legal.map { move -> move to rough(state, seat, move) }
+            val scored = legal.map { move -> move to rough(state, seat, move, profile) }
             val best = scored.maxOf { it.second }
             val equal = scored.filter { it.second == best }
             return equal[random.nextInt(equal.size)].first
         }
 
+        val easy = profile.level == AiLevel.EASY
+        if (easy && planRoll == state.rollsMade) {
+            val next = plan.getOrNull(planned)
+            if (next != null && next in legal) {
+                planned++
+                return next
+            }
+        }
+
         val sequences = BackgammonRules.deepestSequences(state, seat)
         if (sequences.isEmpty()) return legal.first()
 
-        var best = sequences.first()
-        var bestScore = -Double.MAX_VALUE
+        val scored = ArrayList<Pair<List<BackgammonMove>, Double>>(sequences.size)
         for (sequence in sequences) {
             val after = BackgammonRules.playSequence(state, seat, sequence)
-            val score = if (setting == BackgammonLevel.STRONG) {
-                // What the position is worth once the dice have had their say.
-                evaluate(after, seat) - averageReply(after, other(seat))
-            } else {
-                evaluate(after, seat)
-            }
-            if (score > bestScore) {
-                bestScore = score
-                best = sequence
-            }
+            scored += sequence to evaluate(after, seat, tuning)
         }
-        return best.first()
+
+        if (setting == BackgammonLevel.STRONG) {
+            // What the position is worth once the dice have had their say. A hard
+            // seat only spends the twenty-one-roll average on the turns that
+            // already look best, which keeps a phone's thinking time bounded.
+            val finalists = if (profile.level == AiLevel.HARD && scored.size > FINALISTS) {
+                scored.sortedByDescending { it.second }.take(FINALISTS)
+            } else {
+                scored
+            }
+            val rescored = finalists.map { (sequence, _) ->
+                val after = BackgammonRules.playSequence(state, seat, sequence)
+                sequence to (evaluate(after, seat, tuning) - averageReply(after, other(seat)))
+            }
+            return chooseFrom(rescored, profile, state, legal)
+        }
+        return chooseFrom(scored, profile, state, legal)
     }
 
+    /** Best sequence first move; an easy seat sometimes takes a worse (but better-half) turn. */
+    private fun chooseFrom(
+        scored: List<Pair<List<BackgammonMove>, Double>>,
+        profile: AiProfile,
+        state: BackgammonState,
+        legal: List<BackgammonMove>,
+    ): BackgammonMove {
+        var best = scored.first()
+        for (entry in scored) if (entry.second > best.second) best = entry
+        var chosen = best.first
+        if (profile.level == AiLevel.EASY) {
+            val ranked = scored.sortedByDescending { it.second }.map { it.first }
+            chosen = profile.pick(ranked)
+            plan = chosen
+            planned = 1
+            planRoll = state.rollsMade
+        }
+        val first = chosen.first()
+        return if (first in legal) first else legal.first()
+    }
+
+    private fun tuningFor(profile: AiProfile): Tuning {
+        val style = profile.style
+        var blot = 1.0 - 0.3 * style.boldness
+        if (profile.level == AiLevel.EASY) blot *= 0.6 // overlooks some of the danger
+        return Tuning(blot, style.boldness, style.spite)
+    }
+
+    /** What a seat's temperament does to the evaluation; [NEUTRAL] is the standard player. */
+    private class Tuning(val blot: Double, val bold: Double, val spite: Double)
+
     /** A quick opinion of one checker, for the level that does not think. */
-    private fun rough(state: BackgammonState, seat: Int, move: BackgammonMove): Int {
+    private fun rough(state: BackgammonState, seat: Int, move: BackgammonMove, profile: AiProfile): Int {
         val landing = BackgammonRules.destinationOf(seat, move.from, move.die)
         var score = 0
         if (move.from == BAR) score += 50
         if (landing == null) score += 40
-        if (landing != null && state.countOn(landing, other(seat)) == 1) score += 60
+        if (landing != null && state.countOn(landing, other(seat)) == 1) {
+            score += 60 + ((profile.style.boldness + profile.style.spite) * 20).toInt()
+        }
         if (landing != null && state.countOn(landing, seat) >= 1) score += 15
         return score
     }
@@ -98,8 +168,8 @@ class BackgammonAi(private val level: BackgammonLevel? = null) :
                 val rolled = state.copy(turn = opponent, dice = dice, roll = listOf(high, low))
                 val replies = BackgammonRules.deepestSequences(rolled, opponent)
                 val bestReply = replies.maxOfOrNull { sequence ->
-                    evaluate(BackgammonRules.playSequence(rolled, opponent, sequence), opponent)
-                } ?: evaluate(rolled, opponent)
+                    evaluate(BackgammonRules.playSequence(rolled, opponent, sequence), opponent, NEUTRAL)
+                } ?: evaluate(rolled, opponent, NEUTRAL)
                 total += chance * bestReply
                 weight += chance
             }
@@ -117,7 +187,7 @@ class BackgammonAi(private val level: BackgammonLevel? = null) :
      * checkers are, and a made point in your own home board is worth having
      * because it is one more number your opponent cannot come in on.
      */
-    private fun evaluate(state: BackgammonState, seat: Int): Double {
+    private fun evaluate(state: BackgammonState, seat: Int, tuning: Tuning): Double {
         val enemy = other(seat)
 
         if (state.phase == BackgammonPhase.GAME_OVER) {
@@ -135,6 +205,8 @@ class BackgammonAi(private val level: BackgammonLevel? = null) :
         score -= state.off[enemy] * BORNE_OFF
         score -= state.bar[seat] * ON_BAR
         score += state.bar[enemy] * ON_BAR
+        // A gambler and a spoiler both like having an opponent on the bar.
+        score += state.bar[enemy] * (tuning.bold * 4.0 + tuning.spite * 8.0)
 
         for (point in 0 until POINTS) {
             val mine = state.countOn(point, seat)
@@ -142,6 +214,8 @@ class BackgammonAi(private val level: BackgammonLevel? = null) :
 
             if (mine >= 2) {
                 score += POINT_MADE
+                // A cautious seat likes covered points a little more, a gambler less.
+                score -= tuning.bold * 1.5
                 // A point in your own home board is a door shut in the face of
                 // anyone you send to the bar.
                 if (point in homeOf(seat)) score += HOME_POINT
@@ -151,7 +225,7 @@ class BackgammonAi(private val level: BackgammonLevel? = null) :
                 // A blot, weighted by how easily it can be reached: something
                 // six away is in real danger, something twenty away is not.
                 val distance = shortestEnemyDistance(state, seat, point)
-                score -= when {
+                score -= tuning.blot * when {
                     distance <= 0 -> 0.0
                     distance <= 6 -> BLOT_NEAR
                     distance <= 12 -> BLOT_FAR
@@ -160,7 +234,7 @@ class BackgammonAi(private val level: BackgammonLevel? = null) :
             }
         }
 
-        score += primeLength(state, seat) * PRIME
+        score += primeLength(state, seat) * (PRIME + tuning.spite * 3.0)
         return score
     }
 
@@ -199,6 +273,8 @@ class BackgammonAi(private val level: BackgammonLevel? = null) :
 
     private companion object {
         const val WIN = 10_000.0
+        const val FINALISTS = 8
+        val NEUTRAL = Tuning(1.0, 0.0, 0.0)
 
         const val PIP = 1.0
         const val BORNE_OFF = 12.0

app/src/main/java/org/prolibertate/games/game/catan/CatanAi.kt

@@ -1,5 +1,8 @@
 package org.prolibertate.games.game.catan
 
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
@@ -17,39 +20,48 @@ import org.prolibertate.games.game.engine.GameAi
  * dev-card counts, knights played — plus its own hand, and never looks
  * into the deck, another seat's development cards, or the seed beyond the
  * roll the rules derive from it.
+ *
+ * Each seat's [AiProfiles] entry shapes how it plays. The default profile
+ * (medium, balanced) is the AI described above. Easy settles for worse
+ * spots, trades and builds, now and then ends its turn early, and takes
+ * slightly bad trades. Hard weighs setup spots by balance and port synergy,
+ * values cards by what its next build is missing, races for Longest Road,
+ * keeps the robber off its own hexes and off the table's non-leaders, and
+ * is choosier about whom it trades with. Style nudges the tradeoffs: bold
+ * seats expand, cautious seats spend down before a seven can bite, spiteful
+ * seats go after the leader.
  */
-class CatanAi : GameAi<CatanState, CatanMove> {
+class CatanAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) : GameAi<CatanState, CatanMove> {
 
     /** Plays the Cities & Knights tables; this class keeps the base game. */
-    private val ck = CatanCkAi(this)
+    private val ck = CatanCkAi(this, profiles)
 
     override fun chooseMove(state: CatanState, seat: Int, legal: List<CatanMove>): CatanMove {
+        val profile = profiles.of(seat)
         require(legal.isNotEmpty() || state.phase == CatanPhase.PLAYING) {
             "No legal move for seat $seat"
         }
         if (state.isCk) return ck.choose(state, seat, legal)
         return when (state.phase) {
             CatanPhase.SETUP_FIRST, CatanPhase.SETUP_SECOND ->
-                chooseSetup(state, seat, legal.filterIsInstance<SetupPlace>())
+                chooseSetup(state, seat, legal.filterIsInstance<SetupPlace>(), profile)
 
             CatanPhase.AWAIT_ROLL -> legal.first { it is RollDice }
 
             // Gold: take what the seat wants most.
-            CatanPhase.GOLD -> legal.filterIsInstance<PickGold>().maxByOrNull { pick ->
-                pick.cards.sumOf { want(it) }
-            } ?: legal.first()
+            CatanPhase.GOLD -> chooseGold(state, seat, legal, profile)
 
-            CatanPhase.DISCARDING -> chooseDiscard(state, seat, legal.filterIsInstance<Discard>())
+            CatanPhase.DISCARDING -> chooseDiscard(state, seat, legal.filterIsInstance<Discard>(), profile)
 
-            CatanPhase.ROBBER -> chooseRobber(state, seat, legal.filterIsInstance<MoveRobber>())
+            CatanPhase.ROBBER -> chooseRobber(state, seat, legal.filterIsInstance<MoveRobber>(), profile)
 
-            CatanPhase.FREE_ROADS -> chooseFreeRoad(state, seat, legal)
+            CatanPhase.FREE_ROADS -> chooseFreeRoad(state, seat, legal, profile)
 
-            CatanPhase.TRADE_PENDING -> chooseResponse(state, seat, legal)
+            CatanPhase.TRADE_PENDING -> chooseResponse(state, seat, legal, profile)
 
-            CatanPhase.TRADE_COUNTERS -> chooseCounter(state, legal)
+            CatanPhase.TRADE_COUNTERS -> chooseCounter(state, seat, legal, profile)
 
-            CatanPhase.PLAYING -> choosePlaying(state, seat, legal)
+            CatanPhase.PLAYING -> choosePlaying(state, seat, legal, profile)
 
             CatanPhase.CHORE, CatanPhase.GAME_OVER -> legal.firstOrNull()
                 ?: error("Nothing to choose in phase ${state.phase}")
@@ -60,12 +72,68 @@ class CatanAi : GameAi<CatanState, CatanMove> {
     // Setup
     // -----------------------------------------------------------------------
 
-    internal fun chooseSetup(state: CatanState, seat: Int, places: List<SetupPlace>): SetupPlace {
+    internal fun chooseSetup(
+        state: CatanState,
+        seat: Int,
+        places: List<SetupPlace>,
+        profile: AiProfile = AiProfile.DEFAULT,
+    ): SetupPlace {
         require(places.isNotEmpty()) { "No legal settlement placement" }
-        return places.maxByOrNull { place ->
-            spotValue(state, place.vertex) +
-                0.25 * placeDirection(state, place)
-        } ?: places.first()
+        // A bold seat cares more where its road leads; a cautious one less.
+        val direction = 0.25 + 0.1 * profile.style.boldness
+        return when (profile.level) {
+            // Easy looks only at the spot itself, and sometimes settles for the runner-up.
+            AiLevel.EASY -> profile.pick(places.sortedByDescending { spotValue(state, it.vertex) })
+
+            AiLevel.HARD -> places.maxByOrNull { place ->
+                hardSpot(state, seat, place.vertex) + direction * placeDirection(state, place)
+            } ?: places.first()
+
+            AiLevel.MEDIUM -> places.maxByOrNull { place ->
+                spotValue(state, place.vertex) + direction * placeDirection(state, place)
+            } ?: places.first()
+        }
+    }
+
+    /**
+     * A spot as a seat that plans would see it: the plain production value,
+     * plus resources the seat does not yet draw, less piling onto one it
+     * already has plenty of, and ports worth what the seat will produce to
+     * trade through them.
+     */
+    internal fun hardSpot(state: CatanState, seat: Int, vertex: Int): Double {
+        if (state.buildings[vertex] != null) return -1.0
+        val topo = state.topo
+        val owned = HashMap<Resource, Double>()
+        for (v in state.buildings.indices) {
+            val b = state.buildings[v] ?: continue
+            if (b.owner != seat) continue
+            val mult = if (b.type == BuildingType.CITY) 2.0 else 1.0
+            for (h in topo.VERTEX_HEXES[v]) {
+                val res = state.terrains[h].produces ?: continue
+                owned[res] = (owned[res] ?: 0.0) + mult * CatanBoard.numberOdds(state.numbers[h]) / 36.0
+            }
+        }
+        val here = HashMap<Resource, Double>()
+        for (h in topo.VERTEX_HEXES[vertex]) {
+            val res = state.terrains[h].produces ?: continue
+            here[res] = (here[res] ?: 0.0) + CatanBoard.numberOdds(state.numbers[h]) / 36.0
+        }
+        var value = spotValue(state, vertex)
+        value += 0.15 * here.keys.count { (owned[it] ?: 0.0) <= 0.0 }
+        for ((res, p) in here) {
+            if ((owned[res] ?: 0.0) > 0.15) value -= 0.3 * p
+        }
+        val total = owned.values.sum() + here.values.sum()
+        for (e in topo.VERTEX_EDGES[vertex]) {
+            val port = state.ports.firstOrNull { it.edge == e } ?: continue
+            value += if (port.resource == null) {
+                0.05 + 0.15 * total
+            } else {
+                0.9 * ((owned[port.resource] ?: 0.0) + (here[port.resource] ?: 0.0))
+            }
+        }
+        return value
     }
 
     /**
@@ -87,48 +155,127 @@ class CatanAi : GameAi<CatanState, CatanMove> {
     // Dice, discards and the robber
     // -----------------------------------------------------------------------
 
-    private fun chooseDiscard(state: CatanState, seat: Int, discards: List<Discard>): Discard {
+    private fun chooseDiscard(
+        state: CatanState,
+        seat: Int,
+        discards: List<Discard>,
+        profile: AiProfile,
+    ): Discard {
         require(discards.isNotEmpty()) { "No legal discard" }
         // Throw away what it wants least; grain and ore are the city, wool
         // is mostly a dev-card tax.
-        return discards.minByOrNull { discard ->
-            discard.cards.sumOf { want(it) }
-        } ?: discards.first()
+        return when (profile.level) {
+            AiLevel.EASY -> profile.pick(discards.sortedBy { d -> d.cards.sumOf { want(it) } })
+
+            // Keep the hand that is closest to paying for something.
+            AiLevel.HARD -> {
+                val hand = state.resources[seat]
+                discards.maxByOrNull { d ->
+                    val left = hand.toMutableMap()
+                    for (c in d.cards) left[c] = (left[c] ?: 0) - 1
+                    progress(left) - 0.01 * d.cards.sumOf { want(it) }
+                } ?: discards.first()
+            }
+
+            AiLevel.MEDIUM -> discards.minByOrNull { discard ->
+                discard.cards.sumOf { want(it) }
+            } ?: discards.first()
+        }
     }
 
-    private fun chooseRobber(state: CatanState, seat: Int, moves: List<MoveRobber>): MoveRobber {
+    /** How close a hand is to the nearest of the things worth buying, a card count over a price, plus one if it pays. */
+    private fun progress(hand: Map<Resource, Int>): Double = SHOPPING.maxOf { cost ->
+        val have = cost.entries.sumOf { (r, n) -> minOf(maxOf(hand[r] ?: 0, 0), n) }
+        have.toDouble() / cost.values.sum() + if (covers(hand, cost)) 1.0 else 0.0
+    }
+
+    private fun chooseGold(state: CatanState, seat: Int, legal: List<CatanMove>, profile: AiProfile): CatanMove {
+        val picks = legal.filterIsInstance<PickGold>()
+        if (picks.isEmpty()) return legal.first()
+        return when (profile.level) {
+            AiLevel.EASY -> profile.pick(picks.sortedByDescending { p -> p.cards.sumOf { want(it) } })
+            AiLevel.HARD -> picks.maxByOrNull { p -> p.cards.sumOf { need(state, seat, it) } } ?: picks.first()
+            AiLevel.MEDIUM -> picks.maxByOrNull { pick -> pick.cards.sumOf { want(it) } } ?: legal.first()
+        }
+    }
+
+    private fun chooseRobber(
+        state: CatanState,
+        seat: Int,
+        moves: List<MoveRobber>,
+        profile: AiProfile,
+    ): MoveRobber {
         require(moves.isNotEmpty()) { "No legal robber move" }
         val points = CatanRules.publicVictoryPoints(state)
-        return moves.maxByOrNull { move ->
+        val hard = profile.level == AiLevel.HARD
+        val spite = profile.style.spite
+        val leader = (0 until state.playerCount).filter { it != seat }.maxByOrNull { points[it] }
+        val victimWeight = 0.3 + 0.1 * profile.style.boldness
+        fun score(move: MoveRobber): Double {
             val hex = state.topo.hexIndex.getValue(move.hex)
             var score = 0.0
             for (v in state.topo.HEX_CORNERS[hex]) {
                 val building = state.buildings[v] ?: continue
-                if (building.owner == seat) continue
                 val worth = if (building.type == BuildingType.CITY) 2 else 1
-                score += worth * CatanBoard.numberOdds(state.numbers[hex]) / 36.0
+                val raw = worth * CatanBoard.numberOdds(state.numbers[hex])
+                if (building.owner == seat) {
+                    // A hard seat will not sit the robber on its own income.
+                    if (hard) score -= 0.8 * raw / 36.0
+                    continue
+                }
+                score += raw / 36.0
+                if (building.owner == leader) score += spite * raw / 36.0
             }
             if (move.victim != null) {
-                score += 0.15 * state.resourceCount(move.victim) +
-                    0.3 * points[move.victim]
+                val cards = state.resourceCount(move.victim)
+                if (!hard || cards > 0) {
+                    score += 0.15 * cards + victimWeight * points[move.victim]
+                    if (move.victim == leader) score += 0.4 * spite
+                }
             }
-            score
-        } ?: moves.first()
+            return score
+        }
+        return if (profile.level == AiLevel.EASY) {
+            profile.pick(moves.sortedByDescending { score(it) })
+        } else {
+            moves.maxByOrNull { score(it) } ?: moves.first()
+        }
     }
 
     // -----------------------------------------------------------------------
     // The main phase
     // -----------------------------------------------------------------------
 
-    private fun choosePlaying(state: CatanState, seat: Int, legal: List<CatanMove>): CatanMove {
-        // A city doubles the best production the seat already owns.
-        legal.filterIsInstance<BuildCity>().maxByOrNull {
-            spotValue(state, it.vertex)
-        }?.let { return it }
+    private fun choosePlaying(state: CatanState, seat: Int, legal: List<CatanMove>, profile: AiProfile): CatanMove {
+        val easy = profile.level == AiLevel.EASY
+        val hard = profile.level == AiLevel.HARD
+        val boldness = profile.style.boldness
+        val endTurn = legal.lastOrNull { it is EndTurn } ?: legal.first()
 
-        legal.filterIsInstance<BuildSettlement>().maxByOrNull {
-            spotValue(state, it.vertex)
-        }?.let { return it }
+        // Easy loses the thread now and then and just passes.
+        if (easy && profile.slips() && profile.random.nextBoolean()) return endTurn
+
+        val cities = legal.filterIsInstance<BuildCity>()
+        val settlements = legal.filterIsInstance<BuildSettlement>()
+        if (hard && (cities.isNotEmpty() || settlements.isNotEmpty())) {
+            // A city doubles what the seat has; a settlement spreads it and
+            // takes a spot nobody else will. Bold seats lean to the spread.
+            val city = cities.maxByOrNull { spotValue(state, it.vertex) }
+            val settlement = settlements.maxByOrNull { hardSpot(state, seat, it.vertex) }
+            val cityScore = if (city != null) spotValue(state, city.vertex) + 0.15 else Double.NEGATIVE_INFINITY
+            val settleScore = if (settlement != null) {
+                hardSpot(state, seat, settlement.vertex) + 0.3 + 0.1 * boldness
+            } else {
+                Double.NEGATIVE_INFINITY
+            }
+            if (city != null && cityScore >= settleScore) return city
+            if (settlement != null) return settlement
+        } else {
+            // A city doubles the best production the seat already owns.
+            pickBest(cities, profile) { spotValue(state, it.vertex) }?.let { return it }
+
+            pickBest(settlements, profile) { spotValue(state, it.vertex) }?.let { return it }
+        }
 
         // Seafarers: a ship toward an island nobody has landed on is worth more than a card.
         legal.filterIsInstance<BuildShip>()
@@ -138,11 +285,15 @@ class CatanAi : GameAi<CatanState, CatanMove> {
 
         // A knight is worth playing whenever one is playable: it blocks an
         // opponent's best hex and marches toward Largest Army.
-        legal.filterIsInstance<PlayKnight>().firstOrNull()?.let { return it }
+        if (!(easy && profile.slips())) {
+            legal.filterIsInstance<PlayKnight>().firstOrNull()?.let { return it }
+        }
 
-        yearOfPlenty(state, legal)?.let { return it }
+        if (!(easy && profile.slips())) {
+            yearOfPlenty(state, seat, legal, profile)?.let { return it }
 
-        monopoly(state, legal)?.let { return it }
+            monopoly(state, legal)?.let { return it }
+        }
 
         // Road Building is worth two free segments whenever there is
         // somewhere to put them.
@@ -152,6 +303,14 @@ class CatanAi : GameAi<CatanState, CatanMove> {
             return legal.filterIsInstance<PlayRoadBuilding>().first()
         }
 
+        // Hard seat in a Longest Road race puts the road down before spending on cards.
+        val roadThreshold = 0.1 - 0.04 * boldness
+        if (hard && roadRace(state, seat)) {
+            legal.filterIsInstance<BuildRoad>()
+                .maxByOrNull { roadValue(state, it) }
+                ?.let { return it }
+        }
+
         legal.filterIsInstance<BuyDevCard>().firstOrNull()?.let { return it }
 
         // Seafarers: a ship that heads for a settlement spot worth having.
@@ -162,7 +321,7 @@ class CatanAi : GameAi<CatanState, CatanMove> {
 
         // A paid road only when it reaches toward a spot worth settling.
         legal.filterIsInstance<BuildRoad>()
-            .filter { roadValue(state, it) > 0.1 }
+            .filter { roadValue(state, it) > roadThreshold }
             .maxByOrNull { roadValue(state, it) }
             ?.let { return it }
 
@@ -173,16 +332,94 @@ class CatanAi : GameAi<CatanState, CatanMove> {
             .maxByOrNull { CatanRules.bestPortRatio(state, seat, it.give) * -1.0 }
             ?.let { return it }
 
-        offer(state, seat)?.let { return it }
+        if (hard) {
+            // Pile of one card, short of another: convert before it is stuck in hand.
+            legal.filterIsInstance<MaritimeTrade>()
+                .filter { surplusTrade(state, seat, it) }
+                .minByOrNull { CatanRules.bestPortRatio(state, seat, it.give) }
+                ?.let { return it }
+        }
 
-        return legal.lastOrNull { it is EndTurn } ?: legal.first()
+        // Stay under the seven: a cautious (or careful) seat sells down a big hand.
+        if ((hard || boldness < 0) && state.resourceCount(seat) > DISCARD_LIMIT) {
+            legal.filterIsInstance<MaritimeTrade>()
+                .maxByOrNull { t ->
+                    (state.resources[seat][t.give] ?: 0) * 10.0 -
+                        CatanRules.bestPortRatio(state, seat, t.give) * 3.0 + need(state, seat, t.get)
+                }
+                ?.let { return it }
+        }
+
+        if (!(easy && profile.slips())) {
+            offer(state, seat, profile)?.let { return it }
+        }
+
+        return endTurn
+    }
+
+    /** Picks the best of [items] by [score]; easy sometimes takes a near miss. */
+    private fun <T> pickBest(items: List<T>, profile: AiProfile, score: (T) -> Double): T? {
+        if (items.isEmpty()) return null
+        if (profile.level != AiLevel.EASY) return items.maxByOrNull(score)
+        return profile.pick(items.sortedByDescending(score))
+    }
+
+    /** The cheapest things worth saving for, as the shopping list a hard seat weighs hands against. */
+    private val SHOPPING = listOf(
+        CatanRules.CITY_COST,
+        CatanRules.SETTLEMENT_COST,
+        CatanRules.DEV_COST,
+        CatanRules.ROAD_COST,
+    )
+
+    /** The shopping list entry the hand is closest to paying for, among a city, a settlement and a card. */
+    private fun nearestTarget(hand: Map<Resource, Int>): Map<Resource, Int> =
+        listOf(CatanRules.CITY_COST, CatanRules.SETTLEMENT_COST, CatanRules.DEV_COST)
+            .minByOrNull { cost -> cost.entries.sumOf { (r, n) -> maxOf(0, n - (hand[r] ?: 0)) } }
+            ?: CatanRules.CITY_COST
+
+    /** What one card of [res] is worth to this seat right now: more if the next build lacks it, less if it is piled up. */
+    private fun need(state: CatanState, seat: Int, res: Resource): Double {
+        val hand = state.resources[seat]
+        var v = want(res)
+        if ((hand[res] ?: 0) < (nearestTarget(hand)[res] ?: 0)) v += 0.6
+        if ((hand[res] ?: 0) >= 4) v -= 0.4
+        return v
+    }
+
+    /** Whether this bank trade turns a pile (beyond any price) into a card the next build lacks. */
+    private fun surplusTrade(state: CatanState, seat: Int, move: MaritimeTrade): Boolean {
+        val hand = state.resources[seat]
+        val ratio = CatanRules.bestPortRatio(state, seat, move.give)
+        val keep = SHOPPING.maxOf { it[move.give] ?: 0 }
+        val target = nearestTarget(hand)
+        return (hand[move.give] ?: 0) - keep >= ratio &&
+            (hand[move.get] ?: 0) < (target[move.get] ?: 0)
+    }
+
+    /** Whether [seat] is in the hunt for Longest Road: close enough to the holder, or the road is unclaimed. */
+    private fun roadRace(state: CatanState, seat: Int): Boolean {
+        val holder = state.longestRoadOwner
+        if (holder == seat) return false
+        val mine = CatanRules.longestRoadFor(state, seat)
+        if (mine < 3) return false
+        return holder == null || mine + 1 >= CatanRules.longestRoadFor(state, holder)
+    }
+
+    /** The opponent with the most public points, if any. */
+    private fun leaderOf(state: CatanState, seat: Int): Int? {
+        val points = CatanRules.publicVictoryPoints(state)
+        return (0 until state.playerCount).filter { it != seat }.maxByOrNull { points[it] }
     }
 
     /** Year of Plenty: two of whatever the nearest build is shortest of. */
-    private fun yearOfPlenty(state: CatanState, legal: List<CatanMove>): PlayYearOfPlenty? {
+    private fun yearOfPlenty(state: CatanState, seat: Int, legal: List<CatanMove>, profile: AiProfile): PlayYearOfPlenty? {
         val takes = legal.filterIsInstance<PlayYearOfPlenty>()
         if (takes.isEmpty()) return null
         val shortage = shortage(state) ?: return null
+        if (profile.level == AiLevel.HARD) {
+            return takes.maxByOrNull { take -> take.take.sumOf { need(state, seat, it) } }
+        }
         return takes.maxByOrNull { take ->
             take.take.count { it == shortage } * 2 + take.take.size
         }
@@ -234,7 +471,7 @@ class CatanAi : GameAi<CatanState, CatanMove> {
      * At most one offer a turn: a declined offer changes nothing, so
      * re-offering it would only bounce between the same two seats forever.
      */
-    private fun offer(state: CatanState, seat: Int): OfferTrade? {
+    private fun offer(state: CatanState, seat: Int, profile: AiProfile): OfferTrade? {
         if (offeredThisTurn(state, seat)) return null
         val seatHand = state.resources[seat]
         val builds = listOf(CatanRules.CITY_COST, CatanRules.SETTLEMENT_COST)
@@ -247,11 +484,16 @@ class CatanAi : GameAi<CatanState, CatanMove> {
                 .filter { (seatHand[it] ?: 0) > (cost[it] ?: 0) }
                 .sortedByDescending { seatHand[it] ?: 0 }
             if (surplus.isEmpty()) continue
-            val target = (0 until state.playerCount)
+            val holders = (0 until state.playerCount)
                 .filter { it != seat }
                 .filter { (state.resources[it][need] ?: 0) > 0 }
-                .maxByOrNull { state.resources[it][need] ?: 0 }
-                ?: continue
+            val target = (if (profile.level == AiLevel.HARD) {
+                // Do not feed the leader when someone else can make the swap.
+                val points = CatanRules.publicVictoryPoints(state)
+                holders.maxByOrNull { (state.resources[it][need] ?: 0) - 0.5 * points[it] }
+            } else {
+                holders.maxByOrNull { state.resources[it][need] ?: 0 }
+            }) ?: continue
             val give = listOf(surplus.first(), surplus.first())
             val trade = OfferTrade(to = target, give = give, get = listOf(need))
             if (CatanRules.isLegalMove(state, seat, trade)) return trade
@@ -274,38 +516,73 @@ class CatanAi : GameAi<CatanState, CatanMove> {
     // Free roads and trade responses
     // -----------------------------------------------------------------------
 
-    private fun chooseFreeRoad(state: CatanState, seat: Int, legal: List<CatanMove>): CatanMove {
-        val roads = legal.filterIsInstance<BuildRoad>()
-            .filter { roadValue(state, it) > 0.05 }
-            .maxByOrNull { roadValue(state, it) }
-        return roads ?: legal.firstOrNull { it is EndFreeRoads }
-            ?: legal.filterIsInstance<BuildRoad>().firstOrNull()
-            ?: legal.first()
+    private fun chooseFreeRoad(state: CatanState, seat: Int, legal: List<CatanMove>, profile: AiProfile): CatanMove {
+        val roads = legal.filterIsInstance<BuildRoad>().filter { roadValue(state, it) > 0.05 }
+        val road = if (profile.level == AiLevel.EASY) {
+            profile.pick(roads.sortedByDescending { roadValue(state, it) }.ifEmpty { return freeRoadFallback(legal) })
+        } else {
+            roads.maxByOrNull { roadValue(state, it) }
+        }
+        return road ?: freeRoadFallback(legal)
     }
 
-    private fun chooseResponse(state: CatanState, seat: Int, legal: List<CatanMove>): CatanMove {
+    private fun freeRoadFallback(legal: List<CatanMove>): CatanMove =
+        legal.firstOrNull { it is EndFreeRoads }
+            ?: legal.filterIsInstance<BuildRoad>().firstOrNull()
+            ?: legal.first()
+
+    /**
+     * How much extra the incoming cards must be worth to take a swap: nothing
+     * for the standard seat, a hard or spiteful seat wants more from the
+     * table's leader, and an easy seat is not too fussy.
+     */
+    private fun tradeMargin(state: CatanState, seat: Int, other: Int, profile: AiProfile): Double {
+        var margin = 0.0
+        if (profile.level == AiLevel.EASY) margin -= 0.6
+        val leader = leaderOf(state, seat)
+        if (leader != null && other == leader) {
+            if (profile.level == AiLevel.HARD) margin += 0.4
+            margin += 0.5 * profile.style.spite
+        }
+        return margin
+    }
+
+    private fun tradeValue(state: CatanState, seat: Int, res: Resource, profile: AiProfile): Double =
+        if (profile.level == AiLevel.HARD) need(state, seat, res) else want(res)
+
+    private fun chooseResponse(state: CatanState, seat: Int, legal: List<CatanMove>, profile: AiProfile): CatanMove {
         val accept = legal.filterIsInstance<AcceptTrade>().firstOrNull()
         val offer = state.pendingTrade
         if (accept == null || offer == null) return legal.first { it is DeclineTrade }
         // Accept when what arrives (the offerer's give) is worth more than what
         // leaves (the offerer's get): a mild table, not a shark.
-        val incoming = offer.give.sumOf { want(it) }
-        val outgoing = offer.get.sumOf { want(it) }
-        return if (incoming > outgoing) accept else legal.first { it is DeclineTrade }
+        val incoming = offer.give.sumOf { tradeValue(state, seat, it, profile) }
+        val outgoing = offer.get.sumOf { tradeValue(state, seat, it, profile) }
+        return if (incoming > outgoing + tradeMargin(state, seat, offer.by, profile)) {
+            accept
+        } else {
+            legal.first { it is DeclineTrade }
+        }
     }
 
     /**
      * Picks the counter-offer that gains the most by the same valuation, or
      * turns them all down. The AI never counters itself: it only answers.
      */
-    private fun chooseCounter(state: CatanState, legal: List<CatanMove>): CatanMove {
+    private fun chooseCounter(state: CatanState, seat: Int, legal: List<CatanMove>, profile: AiProfile): CatanMove {
         val offer = state.pendingTrade ?: return legal.first()
+        fun gain(c: TradeCounter) = c.give.sumOf { tradeValue(state, seat, it, profile) } -
+            c.get.sumOf { tradeValue(state, seat, it, profile) }
         val best = legal.filterIsInstance<AcceptCounter>()
             .map { choice -> choice to offer.counters.first { it.by == choice.seat } }
-            .maxByOrNull { (_, c) -> c.give.sumOf { want(it) } - c.get.sumOf { want(it) } }
+            .maxByOrNull { (_, c) -> gain(c) }
         if (best != null) {
             val (choice, c) = best
-            if (c.give.sumOf { want(it) } > c.get.sumOf { want(it) }) return choice
+            if (c.give.sumOf { tradeValue(state, seat, it, profile) } >
+                c.get.sumOf { tradeValue(state, seat, it, profile) } + tradeMargin(state, seat, c.by, profile)
+            ) {
+                return choice
+            }
         }
         return legal.first { it is DismissCounters }
     }
@@ -392,4 +669,9 @@ class CatanAi : GameAi<CatanState, CatanMove> {
         Resource.BRICK -> 0.95
         Resource.WOOL -> 0.85 // abundant and only half of a dev card
     }
+
+    private companion object {
+        /** A seven makes a seat holding more than this discard half. */
+        const val DISCARD_LIMIT = 7
+    }
 }

app/src/main/java/org/prolibertate/games/game/catan/CatanCkAi.kt

@@ -1,5 +1,9 @@
 package org.prolibertate.games.game.catan
 
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
+
 /**
  * The computer opponent for Cities & Knights, reached through [CatanAi].
  *
@@ -14,26 +18,40 @@ package org.prolibertate.games.game.catan
  * hired, and are activated when the barbarians draw near), never plays the
  * progress cards whose value depends on a plan (Inventor, Alchemist, Diplomat,
  * Deserter, Intrigue), and makes no domestic offers, though it answers them.
+ *
+ * The seat's profile shapes it as in [CatanAi]: easy plays on a whim and
+ * discards badly, hard sizes its knights to the barbarian threat, values cards
+ * by what its improvements still need and keeps its robber off its own hexes,
+ * and the style leans it toward knights and expansion, toward spending down
+ * a big hand, or toward the leader.
  */
-internal class CatanCkAi(private val ai: CatanAi) {
+internal class CatanCkAi(
+    private val ai: CatanAi,
+    private val profiles: AiProfiles = AiProfiles.DEFAULT,
+) {
 
-    fun choose(state: CatanState, seat: Int, legal: List<CatanMove>): CatanMove = when (state.phase) {
+    fun choose(state: CatanState, seat: Int, legal: List<CatanMove>): CatanMove {
+        val profile = profiles.of(seat)
+        return choose(state, seat, legal, profile)
+    }
+
+    private fun choose(state: CatanState, seat: Int, legal: List<CatanMove>, profile: AiProfile): CatanMove = when (state.phase) {
         CatanPhase.SETUP_FIRST, CatanPhase.SETUP_SECOND ->
-            ai.chooseSetup(state, seat, legal.filterIsInstance<SetupPlace>())
+            ai.chooseSetup(state, seat, legal.filterIsInstance<SetupPlace>(), profile)
 
         CatanPhase.AWAIT_ROLL -> legal.first { it is RollDice }
 
-        CatanPhase.CHORE -> chooseChore(state, seat, legal)
+        CatanPhase.CHORE -> chooseChore(state, seat, legal, profile)
 
-        CatanPhase.ROBBER -> chooseRobber(state, seat, legal.filterIsInstance<MoveRobber>())
+        CatanPhase.ROBBER -> chooseRobber(state, seat, legal.filterIsInstance<MoveRobber>(), profile)
 
-        CatanPhase.FREE_ROADS -> chooseFreeRoad(state, seat, legal)
+        CatanPhase.FREE_ROADS -> chooseFreeRoad(state, seat, legal, profile)
 
-        CatanPhase.TRADE_PENDING -> chooseResponse(state, seat, legal)
+        CatanPhase.TRADE_PENDING -> chooseResponse(state, seat, legal, profile)
 
         CatanPhase.TRADE_COUNTERS -> chooseCounter(state, legal)
 
-        CatanPhase.PLAYING -> choosePlaying(state, seat, legal)
+        CatanPhase.PLAYING -> choosePlaying(state, seat, legal, profile)
 
         CatanPhase.DISCARDING, CatanPhase.GOLD, CatanPhase.GAME_OVER -> legal.firstOrNull()
             ?: error("Nothing to choose in phase ${state.phase}")
@@ -43,18 +61,25 @@ internal class CatanCkAi(private val ai: CatanAi) {
     // Chores
     // -----------------------------------------------------------------------
 
-    private fun chooseChore(state: CatanState, seat: Int, legal: List<CatanMove>): CatanMove {
+    private fun chooseChore(state: CatanState, seat: Int, legal: List<CatanMove>, profile: AiProfile): CatanMove {
         val ck = state.ck!!
         return when (val chore = ck.chores.first()) {
             is DiscardChore -> {
                 val n = minOf(chore.count, CatanCk.goodsCount(state, seat))
-                DiscardGoods(CatanCk.goodsList(state, seat).sortedBy { worth(state, seat, it) }.take(n))
+                val hand = CatanCk.goodsList(state, seat)
+                // Easy now and then throws away whatever comes to hand.
+                val order = if (profile.slips()) {
+                    hand.shuffled(profile.random)
+                } else {
+                    hand.sortedBy { worth(state, seat, it, profile) }
+                }
+                DiscardGoods(order.take(n))
             }
 
             is GiveChore -> {
                 val pool = CatanCk.goodsList(state, seat)
                     .filter { !chore.commoditiesOnly || it.isCommodity }
-                    .sortedBy { worth(state, seat, it) }
+                    .sortedBy { worth(state, seat, it, profile) }
                 GiveGoods(pool.take(minOf(chore.count, pool.size)))
             }
 
@@ -98,32 +123,57 @@ internal class CatanCkAi(private val ai: CatanAi) {
     // The robber
     // -----------------------------------------------------------------------
 
-    private fun chooseRobber(state: CatanState, seat: Int, moves: List<MoveRobber>): MoveRobber {
+    private fun chooseRobber(state: CatanState, seat: Int, moves: List<MoveRobber>, profile: AiProfile): MoveRobber {
         require(moves.isNotEmpty()) { "No legal robber move" }
         val points = CatanRules.publicVictoryPoints(state)
-        return moves.maxByOrNull { move ->
+        val hard = profile.level == AiLevel.HARD
+        val spite = profile.style.spite
+        val leader = (0 until state.playerCount).filter { it != seat }.maxByOrNull { points[it] }
+        val victimWeight = 0.3 + 0.1 * profile.style.boldness
+        fun score(move: MoveRobber): Double {
             val hex = CatanBoard.hexIndex.getValue(move.hex)
             var score = 0.0
             for (v in CatanBoard.HEX_CORNERS[hex]) {
                 val building = state.buildings[v] ?: continue
                 val worth = if (building.type == BuildingType.CITY) 2 else 1
                 val odds = CatanBoard.numberOdds(state.numbers[hex]) / 36.0
-                score += if (building.owner == seat) -worth * odds else worth * odds
+                score += if (building.owner == seat) {
+                    -worth * odds * (if (hard) 1.5 else 1.0)
+                } else {
+                    worth * odds + if (building.owner == leader) spite * worth * odds else 0.0
+                }
             }
             if (move.victim != null) {
-                score += 0.15 * CatanCk.goodsCount(state, move.victim) + 0.3 * points[move.victim]
+                val cards = CatanCk.goodsCount(state, move.victim)
+                if (!hard || cards > 0) {
+                    score += 0.15 * cards + victimWeight * points[move.victim]
+                    if (move.victim == leader) score += 0.4 * spite
+                }
             }
-            score
-        } ?: moves.first()
+            return score
+        }
+        return if (profile.level == AiLevel.EASY) {
+            profile.pick(moves.sortedByDescending { score(it) })
+        } else {
+            moves.maxByOrNull { score(it) } ?: moves.first()
+        }
     }
 
     // -----------------------------------------------------------------------
     // The main phase
     // -----------------------------------------------------------------------
 
-    private fun choosePlaying(state: CatanState, seat: Int, legal: List<CatanMove>): CatanMove {
+    private fun choosePlaying(state: CatanState, seat: Int, legal: List<CatanMove>, profile: AiProfile): CatanMove {
         val ck = state.ck!!
-        val soon = ck.barbarians <= 3
+        val easy = profile.level == AiLevel.EASY
+        val hard = profile.level == AiLevel.HARD
+        val boldness = profile.style.boldness
+        // A cautious seat starts getting ready a step earlier.
+        val soon = ck.barbarians <= (if (boldness < 0) 4 else 3)
+        val endTurn = legal.lastOrNull { it is EndTurn }
+
+        // Easy loses the thread now and then and just passes.
+        if (easy && endTurn != null && profile.slips() && profile.random.nextBoolean()) return endTurn
 
         // Chase the robber off a hex it is squatting on.
         if (ck.attacked) {
@@ -138,11 +188,29 @@ internal class CatanCkAi(private val ai: CatanAi) {
             .maxByOrNull { production(state, it.vertex!!) }
             ?.let { return it }
 
-        legal.filterIsInstance<BuildCity>().maxByOrNull { production(state, it.vertex) }
-            ?.let { return it }
+        val cities = legal.filterIsInstance<BuildCity>()
+        val settlements = legal.filterIsInstance<BuildSettlement>()
+        if (hard) {
+            val city = cities.maxByOrNull { production(state, it.vertex) }
+            val settlement = settlements.maxByOrNull { ai.hardSpot(state, seat, it.vertex) }
+            val cityScore = if (city != null) production(state, city.vertex) + 0.15 else Double.NEGATIVE_INFINITY
+            val settleScore = if (settlement != null) {
+                ai.hardSpot(state, seat, settlement.vertex) + 0.3 + 0.1 * boldness
+            } else {
+                Double.NEGATIVE_INFINITY
+            }
+            if (city != null && cityScore >= settleScore) return city
+            if (settlement != null) return settlement
+        } else if (easy) {
+            if (cities.isNotEmpty()) return profile.pick(cities.sortedByDescending { production(state, it.vertex) })
+            if (settlements.isNotEmpty()) {
+                return profile.pick(settlements.sortedByDescending { ai.spotValue(state, it.vertex) })
+            }
+        } else {
+            cities.maxByOrNull { production(state, it.vertex) }?.let { return it }
 
-        legal.filterIsInstance<BuildSettlement>().maxByOrNull { ai.spotValue(state, it.vertex) }
-            ?.let { return it }
+            settlements.maxByOrNull { ai.spotValue(state, it.vertex) }?.let { return it }
+        }
 
         // Improvements spend cards nothing else needs. Favour the track furthest along.
         legal.filterIsInstance<BuyImprovement>()
@@ -150,16 +218,16 @@ internal class CatanCkAi(private val ai: CatanAi) {
             ?.let { return it }
 
         // Knights: stand ready by the time the barbarians land.
-        defence(state, seat, legal, soon)?.let { return it }
+        if (!(easy && profile.slips())) defence(state, seat, legal, soon, profile)?.let { return it }
 
-        progress(state, seat, legal)?.let { return it }
+        if (!(easy && profile.slips())) progress(state, seat, legal)?.let { return it }
 
         legal.filterIsInstance<BuildWall>().firstOrNull()?.let {
-            if (CatanCk.goodsCount(state, seat) >= 7) return it
+            if (CatanCk.goodsCount(state, seat) >= (if (boldness < 0) 5 else 7)) return it
         }
 
         legal.filterIsInstance<BuildRoad>()
-            .filter { ai.roadValue(state, it) > 0.1 }
+            .filter { ai.roadValue(state, it) > 0.1 - 0.04 * boldness }
             .maxByOrNull { ai.roadValue(state, it) }
             ?.let { return it }
 
@@ -168,18 +236,43 @@ internal class CatanCkAi(private val ai: CatanAi) {
             .minByOrNull { CatanCk.bankRatio(state, seat, it.give) }
             ?.let { return it }
 
-        return legal.lastOrNull { it is EndTurn } ?: legal.first()
+        // Stay under the seven: a cautious (or careful) seat sells down a big hand.
+        if ((hard || boldness < 0) && CatanCk.goodsCount(state, seat) > 7) {
+            legal.filterIsInstance<GoodsBankTrade>()
+                .maxByOrNull { t ->
+                    held(state, seat, t.give) * 10.0 - CatanCk.bankRatio(state, seat, t.give) * 3.0 +
+                        worth(state, seat, t.get, profile)
+                }
+                ?.let { return it }
+        }
+
+        return endTurn ?: legal.first()
     }
 
     /** Hire, promote and activate so the knights are standing when the barbarians land. */
-    private fun defence(state: CatanState, seat: Int, legal: List<CatanMove>, soon: Boolean): CatanMove? {
+    private fun defence(
+        state: CatanState,
+        seat: Int,
+        legal: List<CatanMove>,
+        soon: Boolean,
+        profile: AiProfile,
+    ): CatanMove? {
         val mine = CatanCk.knightsOf(state, seat)
         val grainLeft = state.resources[seat][Resource.GRAIN] ?: 0
 
+        // Hard sizes the defence to the threat: the barbarians bring one strength
+        // per city on the board, and once the table's active knights outweigh them
+        // and this seat pulls its weight, the rest of its goods go elsewhere.
+        if (profile.level == AiLevel.HARD && soon && CatanCk.strength(state, seat) >= 1) {
+            val threat = state.buildings.count { it?.type == BuildingType.CITY }
+            val table = (0 until state.playerCount).sumOf { CatanCk.strength(state, it) }
+            if (table > threat) return null
+        }
+
         if (soon) {
             legal.filterIsInstance<ActivateKnight>().firstOrNull()?.let { return it }
         }
-        val want = if (soon) 4 else 2
+        val want = if (soon) 4 else if (profile.style.boldness > 0.5) 3 else 2
         if (mine.size < want) {
             legal.filterIsInstance<HireKnight>().maxByOrNull { ai.spotValue(state, it.vertex) }?.let { hire ->
                 // Keep a grain back for waking it, once the ship is close.
@@ -289,26 +382,44 @@ internal class CatanCkAi(private val ai: CatanAi) {
         }
     }
 
-    private fun chooseFreeRoad(state: CatanState, seat: Int, legal: List<CatanMove>): CatanMove =
-        legal.filterIsInstance<BuildRoad>()
-            .filter { ai.roadValue(state, it) > 0.05 }
-            .maxByOrNull { ai.roadValue(state, it) }
+    private fun chooseFreeRoad(state: CatanState, seat: Int, legal: List<CatanMove>, profile: AiProfile): CatanMove {
+        val roads = legal.filterIsInstance<BuildRoad>().filter { ai.roadValue(state, it) > 0.05 }
+        val road = when {
+            roads.isEmpty() -> null
+            profile.level == AiLevel.EASY -> profile.pick(roads.sortedByDescending { ai.roadValue(state, it) })
+            else -> roads.maxByOrNull { ai.roadValue(state, it) }
+        }
+        return road
             ?: legal.firstOrNull { it is EndFreeRoads }
             ?: legal.first()
+    }
 
-    private fun chooseResponse(state: CatanState, seat: Int, legal: List<CatanMove>): CatanMove {
+    /** Extra the incoming cards must be worth: easy is not fussy, hard and spiteful seats are tougher on the leader. */
+    private fun tradeMargin(state: CatanState, seat: Int, other: Int, profile: AiProfile): Double {
+        var margin = 0.0
+        if (profile.level == AiLevel.EASY) margin -= 0.6
+        val points = CatanRules.publicVictoryPoints(state)
+        val leader = (0 until state.playerCount).filter { it != seat }.maxByOrNull { points[it] }
+        if (leader != null && other == leader) {
+            if (profile.level == AiLevel.HARD) margin += 0.4
+            margin += 0.5 * profile.style.spite
+        }
+        return margin
+    }
+
+    private fun chooseResponse(state: CatanState, seat: Int, legal: List<CatanMove>, profile: AiProfile): CatanMove {
         val decline = legal.first { it is DeclineTrade }
         val accept = legal.firstOrNull { it is AcceptTrade } ?: return decline
         val offer = state.ck?.goodsOffer
         if (offer != null) {
-            val incoming = offer.give.sumOf { worth(state, seat, it) }
-            val outgoing = offer.get.sumOf { worth(state, seat, it) }
-            return if (incoming > outgoing) accept else decline
+            val incoming = offer.give.sumOf { worth(state, seat, it, profile) }
+            val outgoing = offer.get.sumOf { worth(state, seat, it, profile) }
+            return if (incoming > outgoing + tradeMargin(state, seat, offer.by, profile)) accept else decline
         }
         val plain = state.pendingTrade ?: return decline
         val incoming = plain.give.sumOf { ai.want(it) }
         val outgoing = plain.get.sumOf { ai.want(it) }
-        return if (incoming > outgoing) accept else decline
+        return if (incoming > outgoing + tradeMargin(state, seat, plain.by, profile)) accept else decline
     }
 
     /** Counters to a human's... the AI never makes an offer, so this only guards the phase: take none. */
@@ -320,12 +431,22 @@ internal class CatanCkAi(private val ai: CatanAi) {
     // -----------------------------------------------------------------------
 
     /** What a card is worth to this seat: resources as the base AI sees them, commodities by what they buy. */
-    private fun worth(state: CatanState, seat: Int, good: Good): Double =
-        good.resource?.let { ai.want(it) } ?: when (good) {
+    private fun worth(state: CatanState, seat: Int, good: Good, profile: AiProfile = AiProfile.DEFAULT): Double {
+        val base = good.resource?.let { ai.want(it) } ?: when (good) {
             Good.PAPER -> 0.95
             Good.COIN -> 0.9
             else -> 0.85
         }
+        if (profile.level != AiLevel.HARD) return base
+        // Hard keeps the commodity an improvement is about to need.
+        var bonus = if (good.resource == Resource.ORE || good.resource == Resource.GRAIN) 0.1 else 0.0
+        for (track in Track.entries) {
+            if (Good.of(track.commodity) != good) continue
+            val level = CatanCk.improvement(state, seat, track)
+            if (level < IMPROVEMENT_MAX && held(state, seat, good) <= level + 1) bonus += 0.5
+        }
+        return base + bonus
+    }
 
     /** Which track to lean toward when two are level: politics for the knights, then science. */
     private fun trackLean(track: Track): Int = when (track) {

app/src/main/java/org/prolibertate/games/game/checkers/CheckersAi.kt

@@ -1,6 +1,9 @@
 package org.prolibertate.games.game.checkers
 
 import kotlin.random.Random
+import kotlin.math.roundToInt
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
@@ -17,22 +20,49 @@ import org.prolibertate.games.game.engine.GameAi
  * recapture as its only legal move, so the search plays it whether it wanted to
  * or not.
  */
-class CheckersAi(private val level: CheckersLevel? = null) :
-    GameAi<CheckersState, CheckersMove> {
+class CheckersAi(
+    private val profiles: AiProfiles = AiProfiles.DEFAULT,
+    private val level: CheckersLevel? = null,
+) : GameAi<CheckersState, CheckersMove> {
 
     private var nodeBudget = 0
     private var nodesUsed = 0
 
+    // Evaluation weights and viewpoint for the move being chosen. At the default
+    // (balanced) style they equal the plain constants, so Medium is unchanged.
+    private var rootSeat = 0
+    private var wAdvance = ADVANCE
+    private var wEdge = EDGE
+    private var wBackRow = BACK_ROW
+    private var wTheirMobility = MOBILITY
+
     override fun chooseMove(
         state: CheckersState,
         seat: Int,
         legal: List<CheckersMove>,
     ): CheckersMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
+        val profile = profiles.of(seat)
         val setting = level ?: state.options.level
         if (legal.size == 1) return legal.first()
 
-        nodeBudget = NODE_BUDGET
+        val bold = profile.style.boldness
+        val spite = profile.style.spite
+        rootSeat = seat
+        // Bold seats push men forward and care less about the edge and back row;
+        // cautious ones do the reverse. Spiteful seats weigh the opponent's
+        // freedom of movement more heavily.
+        wAdvance = ADVANCE + (bold * 2).roundToInt()
+        wEdge = EDGE - (bold * 2).roundToInt()
+        wBackRow = BACK_ROW - (bold * 3).roundToInt()
+        wTheirMobility = MOBILITY + (spite * 2).roundToInt()
+
+        val maxDepth = when (profile.level) {
+            AiLevel.EASY -> maxOf(2, setting.depth - 2)
+            AiLevel.MEDIUM -> setting.depth
+            AiLevel.HARD -> setting.depth + 2
+        }
+        nodeBudget = if (profile.level == AiLevel.HARD) HARD_NODE_BUDGET else NODE_BUDGET
         nodesUsed = 0
 
         val random = Random(state.moveLog.size * 31L + state.positionKey().hashCode())
@@ -40,7 +70,7 @@ class CheckersAi(private val level: CheckersLevel? = null) :
         var best = legal.first()
         var scored = legal.map { it to 0 }
 
-        for (depth in 1..setting.depth) {
+        for (depth in 1..maxDepth) {
             val results = mutableListOf<Pair<CheckersMove, Int>>()
             val ordered = order(legal, preferred = best)
             for (move in ordered) {
@@ -61,7 +91,16 @@ class CheckersAi(private val level: CheckersLevel? = null) :
 
         val bestScore = scored.maxOf { it.second }
         val acceptable = scored.filter { it.second >= bestScore - setting.slack() }
-        return acceptable[random.nextInt(acceptable.size)].first
+        val chosen = acceptable[random.nextInt(acceptable.size)].first
+        if (profile.level != AiLevel.EASY) return chosen
+
+        // Easy now and then settles for another sound-looking move: only ones
+        // within a man and a half of the best, never a real blunder.
+        val ranked = listOf(chosen) + scored
+            .filter { it.first != chosen && it.second >= bestScore - EASY_TOLERANCE }
+            .sortedByDescending { it.second }
+            .map { it.first }
+        return profile.pick(ranked)
     }
 
     private fun CheckersLevel.slack(): Int = when (this) {
@@ -126,29 +165,32 @@ class CheckersAi(private val level: CheckersLevel? = null) :
      */
     private fun evaluate(state: CheckersState): Int {
         val us = state.turn
-        val them = other(us)
         var score = 0
+        val me = rootSeat
 
         for (square in 0 until SQUARES) {
             val piece = state.board[square] ?: continue
-            val sign = if (piece.seat == us) 1 else -1
+            val sign = if (piece.seat == me) 1 else -1
             score += sign * if (piece.king) KING else MAN
 
             if (!piece.king) {
                 // How many rows it has come, out of the seven it needs.
                 val advanced = if (piece.seat == BLACK_SEAT) rowOf(square) else 7 - rowOf(square)
-                score += sign * advanced * ADVANCE
+                score += sign * advanced * wAdvance
             }
 
             // Sitting on the edge cannot be jumped, which is worth a little in
             // a game whose entire tactic is being jumped.
-            if (columnOf(square) == 0 || columnOf(square) == 7) score += sign * EDGE
+            if (columnOf(square) == 0 || columnOf(square) == 7) score += sign * wEdge
 
-            if (rowOf(square) == crownRowOf(other(piece.seat))) score += sign * BACK_ROW
+            if (rowOf(square) == crownRowOf(other(piece.seat))) score += sign * wBackRow
         }
 
-        score += MOBILITY * (mobility(state, us) - mobility(state, them))
-        return score
+        val mine = mobility(state, me)
+        val theirs = mobility(state, other(me))
+        score += MOBILITY * mine - wTheirMobility * theirs
+        // Scored from the root seat's side; flip for the side to move.
+        return if (us == me) score else -score
     }
 
     /**
@@ -178,6 +220,8 @@ class CheckersAi(private val level: CheckersLevel? = null) :
         const val INFINITY = 1_000_000
         const val WIN = 100_000
         const val NODE_BUDGET = 200_000
+        const val HARD_NODE_BUDGET = 450_000
+        const val EASY_TOLERANCE = 150
 
         const val MAN = 100
         const val KING = 175

app/src/main/java/org/prolibertate/games/game/chess/ChessAi.kt

@@ -2,6 +2,9 @@ package org.prolibertate.games.game.chess
 
 import kotlin.math.abs
 import kotlin.random.Random
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
@@ -16,7 +19,14 @@ import org.prolibertate.games.game.engine.GameAi
  * so the setup screen sets the strength without the caller having to thread it
  * through.
  */
-class ChessAi(private val level: ChessLevel? = null) : GameAi<ChessState, ChessMove> {
+class ChessAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) : GameAi<ChessState, ChessMove> {
+
+    /** Fixed strength that overrides the table's option (used by tests and tools). */
+    private var level: ChessLevel? = null
+
+    constructor(level: ChessLevel) : this(AiProfiles.DEFAULT) {
+        this.level = level
+    }
 
     /**
      * Ceiling on how much thinking any one move gets. A phone should not stall
@@ -28,10 +38,22 @@ class ChessAi(private val level: ChessLevel? = null) : GameAi<ChessState, ChessM
 
     override fun chooseMove(state: ChessState, seat: Int, legal: List<ChessMove>): ChessMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
+        val profile = profiles.of(seat)
         val setting = level ?: state.options.level
         if (legal.size == 1) return legal.first()
 
-        nodeBudget = NODE_BUDGET
+        // Easy looks one move less far ahead; Hard looks one further and gets
+        // roughly twice the node budget (still a couple of seconds on a phone).
+        val maxDepth = when (profile.level) {
+            AiLevel.EASY -> maxOf(1, setting.depth - 1)
+            AiLevel.MEDIUM -> setting.depth
+            AiLevel.HARD -> setting.depth + 1
+        }
+        nodeBudget = if (profile.level == AiLevel.HARD) HARD_NODE_BUDGET else NODE_BUDGET
+        aiWhite = state.whiteToMove
+        styleActive = profile.style.boldness != 0.0 || profile.style.spite != 0.0
+        styleBoldness = profile.style.boldness
+        styleSpite = profile.style.spite
         nodesUsed = 0
 
         // Deterministic per position, so a host and a client replaying the same
@@ -42,7 +64,7 @@ class ChessAi(private val level: ChessLevel? = null) : GameAi<ChessState, ChessM
         var best = legal.first()
         var scored = legal.map { it to 0 }
 
-        for (depth in 1..setting.depth) {
+        for (depth in 1..maxDepth) {
             val results = mutableListOf<Pair<ChessMove, Int>>()
             // Search the previous iteration's best first, so the sub-searches
             // below it have a decent bound to prune against.
@@ -74,9 +96,56 @@ class ChessAi(private val level: ChessLevel? = null) : GameAi<ChessState, ChessM
         // makes them beatable without making them play nonsense.
         val bestScore = scored.maxOf { it.second }
         val acceptable = scored.filter { it.second >= bestScore - setting.slack() }
-        return acceptable[random.nextInt(acceptable.size)].first
+        val chosen = acceptable[random.nextInt(acceptable.size)].first
+        if (profile.level != AiLevel.EASY) return chosen
+
+        // Easy sometimes settles for a worse move, but only one within a modest
+        // margin of the best, so it never throws a piece away or walks into mate.
+        val ranked = scored.filter { it.second >= bestScore - setting.slack() - EASY_EXTRA_SLACK }
+            .sortedByDescending { it.second }.map { it.first }
+        val ordered = listOf(chosen) + ranked.filter { it != chosen }
+        return profile.pick(ordered)
     }
 
+    // Per-move style state, set at the start of chooseMove.
+    private var aiWhite = true
+    private var styleActive = false
+    private var styleBoldness = 0.0
+    private var styleSpite = 0.0
+
+    /**
+     * A small temperament term (well under a pawn) from the mover's point of
+     * view. Aggressive seats like pieces near the enemy king, cautious seats
+     * like a sheltered own king, spiteful seats like the enemy being badly
+     * placed. Zero for the default style, so default play is unchanged.
+     */
+    private fun styleTerm(state: ChessState): Int {
+        val enemyKing = state.kingSquare(white = !aiWhite)
+        val myKing = state.kingSquare(white = aiWhite)
+        var attackers = 0
+        var shield = 0
+        var enemyPlacement = 0
+        for (square in 0 until 64) {
+            val piece = state.board[square] ?: continue
+            if (piece.white == aiWhite) {
+                if (piece.kind == PieceKind.KING) continue
+                if (enemyKing != null && piece.kind != PieceKind.PAWN && distance(square, enemyKing) <= 3) attackers++
+                if (myKing != null && distance(square, myKing) <= 1) shield++
+            } else {
+                val table = tableFor(piece.kind)
+                enemyPlacement += table[if (piece.white) square else square xor 56]
+            }
+        }
+        var term = 0.0
+        if (styleBoldness > 0) term += styleBoldness * 4 * attackers
+        if (styleBoldness < 0) term += -styleBoldness * 5 * shield
+        term -= styleSpite * 0.1 * enemyPlacement
+        val own = term.toInt().coerceIn(-40, 40)
+        return if (state.whiteToMove == aiWhite) own else -own
+    }
+
+    private fun distance(a: Int, b: Int): Int = maxOf(abs(fileOf(a) - fileOf(b)), abs(rankOf(a) - rankOf(b)))
+
     private fun ChessLevel.slack(): Int = when (this) {
         ChessLevel.CASUAL -> 70
         ChessLevel.CLUB -> 20
@@ -116,7 +185,7 @@ class ChessAi(private val level: ChessLevel? = null) : GameAi<ChessState, ChessM
      */
     private fun quiescence(state: ChessState, alpha: Int, beta: Int, ply: Int): Int {
         nodesUsed++
-        val standPat = evaluate(state)
+        val standPat = evaluate(state) + if (styleActive) styleTerm(state) else 0
         if (standPat >= beta) return beta
         var localAlpha = maxOf(alpha, standPat)
         if (ply >= MAX_QUIESCENCE_PLY) return localAlpha
@@ -205,6 +274,8 @@ class ChessAi(private val level: ChessLevel? = null) : GameAi<ChessState, ChessM
         const val INFINITY = 1_000_000
         const val MATE = 100_000
         const val NODE_BUDGET = 120_000
+        const val HARD_NODE_BUDGET = 240_000
+        const val EASY_EXTRA_SLACK = 60
         const val MAX_QUIESCENCE_PLY = 24
 
         /**

app/src/main/java/org/prolibertate/games/game/crazyeights/CrazyEightsAi.kt

@@ -1,52 +1,135 @@
 package org.prolibertate.games.game.crazyeights
 
+import org.prolibertate.games.game.cards.Card
+import org.prolibertate.games.game.cards.Rank
 import org.prolibertate.games.game.cards.Suit
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
- * A heuristic Crazy 8s opponent.
+ * A Crazy 8s opponent whose strength and temperament come from the seat's profile.
  *
- * It sheds expensive cards first, keeps its eights back for when it is stuck,
- * and when it does play one it names the suit it holds most of.
+ * MEDIUM (the default) sheds expensive cards first, keeps its eights back for
+ * when it is stuck, and when it does play one it names the suit it holds most of.
+ *
+ * EASY plays the same way but sometimes sheds a cheaper card or wastes an eight.
+ * HARD counts the cards on the discard pile to know what is still out, plays to
+ * keep a long suit it can run, and when someone is close to going out it turns
+ * the game to a suit that is nearly exhausted so they cannot follow.
  */
-class CrazyEightsAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
+class CrazyEightsAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) :
     GameAi<CrazyEightsState, CrazyEightsMove> {
 
-    enum class Difficulty(
-        /** Play an eight anyway once the hand is down to this size. */
-        val eightEndgame: Int,
-    ) {
-        EASY(eightEndgame = 99),
-        NORMAL(eightEndgame = 3),
-        CAUTIOUS(eightEndgame = 2),
-    }
-
     override fun chooseMove(
         state: CrazyEightsState,
         seat: Int,
         legal: List<CrazyEightsMove>,
     ): CrazyEightsMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
+        val profile = profiles.of(seat)
         val plays = legal.filterIsInstance<PlayCard>()
         if (plays.isEmpty()) return legal.first()
 
+        val move: CrazyEightsMove = when (profile.level) {
+            AiLevel.MEDIUM -> standardRanking(state, seat, plays, EIGHT_ENDGAME).first()
+            AiLevel.EASY -> profile.pick(standardRanking(state, seat, plays, EASY_EIGHT_ENDGAME))
+            AiLevel.HARD -> hard(state, seat, plays, profile)
+        }
+        return if (move in legal) move else legal.first()
+    }
+
+    /**
+     * Ordinary cards first, most expensive leading, then eights with the suit
+     * this hand holds most of named first.
+     */
+    private fun standardRanking(
+        state: CrazyEightsState,
+        seat: Int,
+        plays: List<PlayCard>,
+        eightEndgame: Int,
+    ): List<PlayCard> {
         val hand = state.hands[seat]
         val ordinary = plays.filter { !isWild(it.card) }
 
         // An eight is the only card that always plays, so it is worth more held
         // than spent — unless the hand is nearly gone, or nothing else works.
-        val useEight = ordinary.isEmpty() || hand.size <= difficulty.eightEndgame
+        val useEight = ordinary.isEmpty() || hand.size <= eightEndgame
         val pool = if (useEight) plays else ordinary
 
-        val best = pool
-            .filter { !isWild(it.card) }
-            .maxByOrNull { crazyEightsPenalty(it.card) }
-
-        if (best != null) return best
+        val ranked = pool.filter { !isWild(it.card) }.sortedByDescending { crazyEightsPenalty(it.card) }
+        if (ranked.isNotEmpty() && !useEight) return ranked
 
         // Playing an eight: switch the game to whatever we hold most of.
-        val strongest = Suit.entries.maxByOrNull { suit -> hand.count { it.suit == suit } }
-            ?: Suit.SPADES
-        return pool.firstOrNull { it.nominatedSuit == strongest } ?: pool.first()
+        val suits = Suit.entries.sortedByDescending { suit -> hand.count { it.suit == suit } }
+        val wilds = suits.flatMap { suit -> pool.filter { isWild(it.card) && it.nominatedSuit == suit } }
+        return ranked + wilds.ifEmpty { pool.filter { isWild(it.card) } }
+    }
+
+    // -----------------------------------------------------------------------
+    // Hard
+    // -----------------------------------------------------------------------
+
+    private fun hard(
+        state: CrazyEightsState,
+        seat: Int,
+        plays: List<PlayCard>,
+        profile: AiProfile,
+    ): CrazyEightsMove {
+        val hand = state.hands[seat]
+        val options = state.options
+
+        // Cards this seat can account for: its own hand and everything on the pile.
+        val known = HashSet<Card>().also { it.addAll(hand); it.addAll(state.discard) }
+        val unseen = FULL_DECK.filter { it !in known }
+        fun unseenIn(suit: Suit): Int = unseen.count { it.suit == suit }
+
+        val next = (seat + 1) % options.playerCount
+        val nextCount = state.handCounts[next]
+        val smallestOpponent = (0 until options.playerCount).filter { it != seat }
+            .minOf { state.handCounts[it] }
+        // How hard to squeeze the suits: more when somebody is about to go out.
+        val urgency = 1.0 + profile.style.spite +
+            (if (nextCount <= 2) 1.0 else 0.0) + (if (smallestOpponent <= 2) 0.5 else 0.0)
+
+        /** Value of leaving the game showing [suit], with [remaining] the cards still held. */
+        fun suitValue(suit: Suit, remaining: List<Card>): Double =
+            3.0 * remaining.count { it.suit == suit } - 0.7 * urgency * unseenIn(suit)
+
+        val ordinary = plays.filter { !isWild(it.card) }
+        val wilds = plays.filter { isWild(it.card) }
+
+        // Aggressive seats spend eights earlier; cautious ones keep them as insurance.
+        val endgame = (EIGHT_ENDGAME + profile.style.boldness * 2.0)
+        val blocking = smallestOpponent <= 1 || (nextCount <= 2 && urgency >= 2.0)
+        val useEight = ordinary.isEmpty() || hand.size <= endgame || (blocking && wilds.isNotEmpty())
+
+        if (hand.size == 1) {
+            plays.firstOrNull()?.let { return it }
+        }
+
+        if (!useEight || wilds.isEmpty()) {
+            if (ordinary.isNotEmpty()) {
+                // Cautious seats shed liabilities harder, bold ones care more about position.
+                val shedWeight = 1.0 - 0.3 * profile.style.boldness
+                return ordinary.maxBy { play ->
+                    val remaining = hand - play.card
+                    crazyEightsPenalty(play.card) * shedWeight + suitValue(play.card.suit, remaining)
+                }
+            }
+        }
+
+        // An eight: name the suit that is best for this hand and worst for the others.
+        return wilds.maxBy { play ->
+            val remaining = hand - play.card
+            suitValue(play.nominatedSuit ?: play.card.suit, remaining)
+        }
+    }
+
+    private companion object {
+        const val EIGHT_ENDGAME = 3
+        const val EASY_EIGHT_ENDGAME = 99
+        val FULL_DECK: List<Card> = Suit.entries.flatMap { suit -> Rank.standard.map { Card(it, suit) } }
     }
 }

app/src/main/java/org/prolibertate/games/game/cribbage/CribbageAi.kt

@@ -3,8 +3,12 @@ package org.prolibertate.games.game.cribbage
 import org.prolibertate.games.game.cards.Card
 import org.prolibertate.games.game.cards.Decks
 import org.prolibertate.games.game.cards.Rank
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 import kotlin.math.abs
+import kotlin.random.Random
 
 /**
  * A cribbage opponent that counts rather than guesses.
@@ -18,8 +22,14 @@ import kotlin.math.abs
  *
  * The play is a heuristic: take the points on offer, and try not to leave the
  * count where a ten-card collects fifteen or thirty-one.
+ *
+ * Per seat, from [profiles]: Easy settles for a worse lay-away or card now and
+ * then; Hard also samples the crib (what the unseen cards might add to it) and
+ * weighs what a reply could score off each card it plays. Boldness trades the
+ * two counts-to-avoid and crib ownership; spite weighs the other side's crib.
  */
-class CribbageAi : GameAi<CribbageState, CribbageMove> {
+class CribbageAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) :
+    GameAi<CribbageState, CribbageMove> {
 
     override fun chooseMove(
         state: CribbageState,
@@ -27,30 +37,78 @@ class CribbageAi : GameAi<CribbageState, CribbageMove> {
         legal: List<CribbageMove>,
     ): CribbageMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
+        val profile = profiles.of(seat)
         val layAways = legal.filterIsInstance<LayAway>()
-        if (layAways.isNotEmpty()) return chooseLayAway(state, seat, layAways)
-        return choosePeg(state, legal.filterIsInstance<PegCard>())
+        if (layAways.isNotEmpty()) return chooseLayAway(state, seat, layAways, profile)
+        val pegs = legal.filterIsInstance<PegCard>()
+        if (pegs.isNotEmpty()) return choosePeg(state, seat, pegs, profile)
+        return legal.first()
     }
 
     private fun chooseLayAway(
         state: CribbageState,
         seat: Int,
         choices: List<LayAway>,
+        profile: AiProfile,
     ): LayAway {
         val hand = state.hands[seat]
         val held = hand.toSet()
         val unseen = Decks.standard52().filter { it !in held }
         // A partner's crib is your crib. That is the whole of four-handed play.
         val ownCrib = state.options.teamOf(seat) == state.options.teamOf(state.dealer)
+        val boldness = profile.style.boldness
+        // Bold seats back their own crib; careful ones fear the other side's.
+        val cribSign = if (ownCrib) {
+            1.0 + 0.25 * boldness
+        } else {
+            -(1.0 - 0.25 * boldness + 0.3 * profile.style.spite)
+        }
+        val trials = if (profile.level == AiLevel.HARD) sampleCribs(hand, unseen, state.options) else null
 
-        return choices.maxByOrNull { choice ->
+        val scored = choices.map { choice ->
             val given = choice.cards.toSet()
             val keep = hand.filter { it !in given }
             val expected = unseen.sumOf { starter ->
                 CribbageScoring.show(keep, starter, isCrib = false).sumOf { it.points }
             }.toDouble() / unseen.size
-            expected + cribValue(choice.cards) * (if (ownCrib) 1.0 else -1.0)
-        } ?: choices.first()
+            val crib = if (trials != null) {
+                // Heuristic and sampled estimates agree on direction; averaging
+                // them smooths the sampling noise without losing the structure.
+                (sampledCribValue(choice.cards, trials) + cribValue(choice.cards)) / 2.0
+            } else {
+                cribValue(choice.cards)
+            }
+            choice to expected + crib * cribSign
+        }
+        // Stable sort: ties resolve to the first choice, as maxByOrNull did.
+        val ranked = scored.sortedByDescending { it.second }.map { it.first }
+        return profile.pick(ranked)
+    }
+
+    /** Sets of unseen cards a crib might receive, plus a cut, for Hard to average over. */
+    private fun sampleCribs(
+        hand: List<Card>,
+        unseen: List<Card>,
+        options: CribbageOptions,
+    ): List<List<Card>> {
+        val random = Random(hand.sumOf { it.rank.ordinal * 4 + it.suit.ordinal }.toLong())
+        // Whatever the crib is short of after our own cards, plus the cut.
+        val others = (CRIB_SIZE - options.layAwaySize).coerceAtLeast(0)
+        return List(CRIB_SAMPLES) {
+            val picked = LinkedHashSet<Card>()
+            while (picked.size < others + 1) picked += unseen[random.nextInt(unseen.size)]
+            picked.toList()
+        }
+    }
+
+    private fun sampledCribValue(given: List<Card>, trials: List<List<Card>>): Double {
+        var total = 0
+        for (trial in trials) {
+            val starter = trial.last()
+            val crib = given + trial.dropLast(1)
+            total += CribbageScoring.show(crib, starter, isCrib = true).sumOf { it.points }
+        }
+        return total.toDouble() / trials.size
     }
 
     /**
@@ -83,22 +141,63 @@ class CribbageAi : GameAi<CribbageState, CribbageMove> {
         return value
     }
 
-    private fun choosePeg(state: CribbageState, choices: List<PegCard>): PegCard {
+    private fun choosePeg(
+        state: CribbageState,
+        seat: Int,
+        choices: List<PegCard>,
+        profile: AiProfile,
+    ): PegCard {
         val count = state.count
         val onTheTable = state.series.map { it.card }
+        val boldness = profile.style.boldness
+        // Bold seats risk leaving five or twenty-one; careful ones really avoid it.
+        val countPenalty = 3.0 - boldness * 1.5
+        val leadFivePenalty = 4.0 - boldness * 2.0
+        val hard = profile.level == AiLevel.HARD
+        // Cards a reply could come from: nothing this seat holds, has seen played or was cut.
+        val unseen = if (hard) {
+            val known = (state.remaining(seat) + state.played.map { it.card } + listOfNotNull(state.starter)).toSet()
+            Decks.standard52().filter { it !in known && it !in state.hands[seat] }
+        } else {
+            emptyList()
+        }
 
-        return choices.maxByOrNull { choice ->
-            val scored = CribbageScoring.peg(onTheTable + choice.card).sumOf { it.points }
+        val scored = choices.map { choice ->
+            val points = CribbageScoring.peg(onTheTable + choice.card).sumOf { it.points }
             val after = count + pipValue(choice.card)
-            var value = scored * 10.0
+            var value = points * 10.0
             // Five and twenty-one are the two counts to avoid leaving: any of
             // the sixteen ten-cards takes fifteen or thirty-one off them.
-            if (after == 5 || after == 21) value -= 3.0
+            if (after == 5 || after == 21) value -= countPenalty
             // Leading a five is the same present, offered first.
-            if (count == 0 && choice.card.rank == Rank.FIVE) value -= 4.0
+            if (count == 0 && choice.card.rank == Rank.FIVE) value -= leadFivePenalty
             // Big cards early. Low ones are what let you keep playing at the
             // awkward end of a series, where the last card is worth a point.
-            value + pipValue(choice.card) * 0.1
-        } ?: choices.first()
+            value += pipValue(choice.card) * 0.1
+            if (hard && unseen.isNotEmpty() && after < CRIBBAGE_LIMIT) {
+                value -= replyThreat(onTheTable + choice.card, after, unseen) * 8.0
+            }
+            choice to value
+        }
+        val ranked = scored.sortedByDescending { it.second }.map { it.first }
+        return profile.pick(ranked)
+    }
+
+    /** Average points a reply scores off [series], over the cards still unseen. */
+    private fun replyThreat(series: List<Card>, after: Int, unseen: List<Card>): Double {
+        var total = 0
+        var playable = 0
+        for (card in unseen) {
+            if (after + pipValue(card) > CRIBBAGE_LIMIT) continue
+            playable++
+            total += CribbageScoring.peg(series + card).sumOf { it.points }
+        }
+        // Nothing playable is a go for them, worth a point to whoever last played.
+        return if (playable == 0) 0.0 else total.toDouble() / unseen.size
+    }
+
+    private companion object {
+        const val CRIB_SIZE = 4
+        const val CRIB_SAMPLES = 60
     }
 }

app/src/main/java/org/prolibertate/games/game/engine/AiProfile.kt

@@ -0,0 +1,123 @@
+package org.prolibertate.games.game.engine
+
+import kotlinx.serialization.Serializable
+import kotlin.random.Random
+
+/** How well one computer seat plays. */
+enum class AiLevel {
+    /** Makes frequent, plausible mistakes and plays without much foresight. */
+    EASY,
+
+    /** The game's standard computer opponent. */
+    MEDIUM,
+
+    /** Plays carefully: more foresight, no deliberate mistakes. */
+    HARD,
+}
+
+/** What the person at setup chooses for the table's computer seats. */
+@Serializable
+enum class AiDifficulty(val label: String) {
+    EASY("Easy"),
+    MEDIUM("Medium"),
+    HARD("Hard"),
+
+    /** Each computer seat gets its own level, picked at random. */
+    MIXED("Mixed"),
+}
+
+/**
+ * A computer seat's temperament. Never shown to the human players: it only
+ * shapes how the AI trades off the options in front of it.
+ */
+enum class AiStyle(
+    /** -1 plays safe, +1 gambles; 0 is the game's standard balance. */
+    val boldness: Double,
+    /** How much the seat cares about stopping opponents rather than building its own position. */
+    val spite: Double,
+) {
+    BALANCED(0.0, 0.0),
+    AGGRESSIVE(0.8, 0.3),
+    CAUTIOUS(-0.8, 0.0),
+    SPOILER(0.0, 0.8),
+}
+
+/**
+ * One computer seat's level and style, and a source of randomness of its own.
+ *
+ * [DEFAULT] is the game's standard opponent: [AiLevel.MEDIUM], [AiStyle.BALANCED],
+ * and it never slips, so an AI built without a profile plays exactly as it did
+ * before profiles existed.
+ */
+class AiProfile(
+    val level: AiLevel = AiLevel.MEDIUM,
+    val style: AiStyle = AiStyle.BALANCED,
+    val random: Random = Random(0),
+) {
+    /**
+     * Whether this seat should throw away its best option this time. Only
+     * [AiLevel.EASY] does, and only some of the time, so it still plays like a
+     * person who is not concentrating rather than like a random-move generator.
+     */
+    fun slips(): Boolean = level == AiLevel.EASY && random.nextDouble() < EASY_SLIP_CHANCE
+
+    /**
+     * Picks from [ranked] (best first). Easy sometimes settles for a worse choice,
+     * drawn from the better half so it stays plausible; other levels take the best.
+     */
+    fun <T> pick(ranked: List<T>): T {
+        require(ranked.isNotEmpty()) { "Nothing to pick from" }
+        if (!slips() || ranked.size == 1) return ranked.first()
+        val pool = ranked.take(maxOf(2, (ranked.size + 1) / 2))
+        return pool[random.nextInt(pool.size)]
+    }
+
+    /** [base] nudged by this seat's boldness: a gambler raises it, a cautious seat lowers it. */
+    fun bolder(base: Double, strength: Double = 1.0): Double = base + style.boldness * strength
+
+    companion object {
+        const val EASY_SLIP_CHANCE = 0.3
+        val DEFAULT = AiProfile()
+    }
+}
+
+/**
+ * The profile for every computer seat at a table, worked out from the table's
+ * difficulty setting and its seed alone.
+ *
+ * Derived rather than stored, so a saved game, a rematch from the same config
+ * and a guest's copy of a hosted table all give the same seat the same
+ * personality without it travelling anywhere.
+ */
+class AiProfiles(
+    private val difficulty: AiDifficulty,
+    private val seed: Long,
+) {
+    private val cache = HashMap<Int, AiProfile>()
+
+    constructor(config: TableConfig) : this(config.aiDifficulty, config.seed)
+
+    @Synchronized
+    fun of(seat: Int): AiProfile = cache.getOrPut(seat) {
+        val picker = Random(seed * 31 + seat * 7919L + 17)
+        val level = when (difficulty) {
+            AiDifficulty.EASY -> AiLevel.EASY
+            AiDifficulty.MEDIUM -> AiLevel.MEDIUM
+            AiDifficulty.HARD -> AiLevel.HARD
+            AiDifficulty.MIXED -> AiLevel.entries[picker.nextInt(AiLevel.entries.size)]
+        }
+        // The standard opponent is balanced by definition; the others each get a
+        // temperament, so no two computer seats feel alike.
+        val style = if (difficulty == AiDifficulty.MEDIUM) {
+            AiStyle.BALANCED
+        } else {
+            AiStyle.entries[picker.nextInt(AiStyle.entries.size)]
+        }
+        AiProfile(level, style, Random(picker.nextLong()))
+    }
+
+    companion object {
+        /** Standard opponents, for a caller that has no table to ask. */
+        val DEFAULT = AiProfiles(AiDifficulty.MEDIUM, 0L)
+    }
+}

app/src/main/java/org/prolibertate/games/game/engine/Engine.kt

@@ -43,6 +43,8 @@ data class TableConfig(
     val optionsJson: String,
     /** Shared shuffle seed so a hand is reproducible from its state. */
     val seed: Long,
+    /** How hard the computer seats play. Old saves and old hosts omit it and get the standard opponent. */
+    val aiDifficulty: AiDifficulty = AiDifficulty.MEDIUM,
 )
 
 /**

app/src/main/java/org/prolibertate/games/game/euchre/EuchreAi.kt

@@ -3,6 +3,10 @@ package org.prolibertate.games.game.euchre
 import org.prolibertate.games.game.cards.Card
 import org.prolibertate.games.game.cards.Rank
 import org.prolibertate.games.game.cards.Suit
+import org.prolibertate.games.game.cards.Decks
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
@@ -13,7 +17,7 @@ import org.prolibertate.games.game.engine.GameAi
  * That is enough to be a reasonable table opponent without pretending to be a
  * solver.
  */
-class EuchreAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
+class EuchreAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) :
     GameAi<EuchreState, EuchreMove> {
 
     /**
@@ -22,25 +26,50 @@ class EuchreAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
      * means every seat bids and the card is never turned down — see
      * EuchreBiddingRateTest, which pins the resulting order-up rate.
      */
-    enum class Difficulty(val bidThreshold: Int, val aloneThreshold: Int) {
-        /** Bids loosely and plays straightforwardly. */
-        EASY(bidThreshold = 21, aloneThreshold = 32),
+    private companion object {
+        const val BID_THRESHOLD = 24.0
+        const val ALONE_THRESHOLD = 32.0
+    }
 
-        NORMAL(bidThreshold = 24, aloneThreshold = 32),
+    /** Cards this AI has watched played in the current hand, per seat it plays. */
+    private val memory = HashMap<Int, SeenCards>()
 
-        /** Bids tighter and is harder to sneak a trick past. */
-        CAUTIOUS(bidThreshold = 27, aloneThreshold = 31),
+    private class SeenCards {
+        var key: Long = Long.MIN_VALUE
+        val cards = HashSet<Card>()
+    }
+
+    /** Public information only: cards that have appeared on the table this hand. */
+    private fun observe(state: EuchreState, seat: Int): Set<Card> = synchronized(memory) {
+        val seen = memory.getOrPut(seat) { SeenCards() }
+        val key = state.seed * 1_000_003L + state.handNumber
+        if (key != seen.key || state.phase == EuchrePhase.BID_ROUND_1) {
+            seen.key = key
+            seen.cards.clear()
+        }
+        state.trick.forEach { seen.cards.add(it.card) }
+        state.completedTrick.forEach { seen.cards.add(it.card) }
+        state.turnedDown?.let { seen.cards.add(it) }
+        seen.cards.removeAll(state.hands[seat].toSet())
+        seen.cards.toSet()
     }
 
     override fun chooseMove(state: EuchreState, seat: Int, legal: List<EuchreMove>): EuchreMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
-        return when (state.phase) {
-            EuchrePhase.BID_ROUND_1 -> chooseBidRound1(state, seat, legal)
-            EuchrePhase.BID_ROUND_2 -> chooseBidRound2(state, seat, legal)
-            EuchrePhase.DEALER_DISCARD -> chooseDiscard(state, legal)
-            EuchrePhase.PLAYING -> choosePlay(state, seat, legal)
-            else -> legal.first()
+        val profile = profiles.of(seat)
+        val seen = if (profile.level == AiLevel.HARD) observe(state, seat) else emptySet()
+        val move = try {
+            when (state.phase) {
+                EuchrePhase.BID_ROUND_1 -> chooseBidRound1(state, seat, legal, profile)
+                EuchrePhase.BID_ROUND_2 -> chooseBidRound2(state, seat, legal, profile)
+                EuchrePhase.DEALER_DISCARD -> chooseDiscard(state, legal, profile)
+                EuchrePhase.PLAYING -> choosePlay(state, seat, legal, profile, seen)
+                else -> legal.first()
+            }
+        } catch (e: RuntimeException) {
+            legal.first()
         }
+        return if (move in legal) move else legal.first()
     }
 
     // -----------------------------------------------------------------------
@@ -75,27 +104,61 @@ class EuchreAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
         return value
     }
 
+    /** The five-card hand left after the dealer picks up [upCard] and discards the least useful card. */
+    private fun bestAfterPickup(hand: List<Card>, upCard: Card, trump: Suit): Int {
+        val six = hand + upCard
+        return six.maxOf { dropped -> handValue(six - dropped, trump) }
+    }
+
+    /** A hand worth playing without a partner: real trump length with a bower to lead it. */
+    private fun soundAlone(hand: List<Card>, trump: Suit): Boolean {
+        val trumps = hand.count { isTrumpCard(it, trump) }
+        val bower = hand.any { isRightBower(it, trump) || isLeftBower(it, trump) }
+        return trumps >= 3 && bower
+    }
+
+    /** Easy misjudges how good its hand is. */
+    private fun misjudge(value: Int, profile: AiProfile): Int =
+        if (profile.level == AiLevel.EASY) value + profile.random.nextInt(-5, 6) else value
+
     private fun chooseBidRound1(
         state: EuchreState,
         seat: Int,
         legal: List<EuchreMove>,
+        profile: AiProfile,
     ): EuchreMove {
         val upCard = state.upCard ?: return Pass
         val trump = upCard.suit
-        var hand = state.hands[seat]
+        val isDealer = seat == state.dealer
+        val partnerDealer = partnerOf(seat) == state.dealer
+        val hard = profile.level == AiLevel.HARD
 
         // If our partner is the dealer they gain the turn card, and if we are
         // the dealer we gain it ourselves — both make the hand stronger than it
         // looks. Ordering up into an opponent dealer instead hands them a card.
-        if (seat == state.dealer) hand = hand + upCard
-        var value = handValue(hand, trump)
-        if (partnerOf(seat) == state.dealer) value += 2
-        if (seat != state.dealer && partnerOf(seat) != state.dealer) value -= 2
+        val owned = if (isDealer) state.hands[seat] + upCard else state.hands[seat]
+        var value = when {
+            isDealer && hard -> bestAfterPickup(state.hands[seat], upCard, trump) + 3
+            else -> handValue(owned, trump)
+        }
+        if (partnerDealer) value += 2
+        if (!isDealer && !partnerDealer) {
+            value -= 2
+            // An opposing dealer who picks up a jack gets a bower out of our bid.
+            if (hard && upCard.rank == Rank.JACK) value -= 2
+        }
+        value = misjudge(value, profile)
+
+        val bidThreshold = profile.bolder(BID_THRESHOLD, -3.0)
+        var aloneThreshold = profile.bolder(ALONE_THRESHOLD, -3.0)
+        if (profile.level == AiLevel.EASY) aloneThreshold += 4.0
 
         val alone = OrderUp(alone = true)
-        if (value >= difficulty.aloneThreshold && legal.contains(alone)) return alone
+        if (value >= aloneThreshold && legal.contains(alone) && (!hard || soundAlone(owned, trump))) {
+            return alone
+        }
         val order = OrderUp(alone = false)
-        if (value >= difficulty.bidThreshold && legal.contains(order)) return order
+        if (value >= bidThreshold && legal.contains(order)) return order
         return if (legal.contains(Pass)) Pass else legal.first()
     }
 
@@ -103,37 +166,64 @@ class EuchreAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
         state: EuchreState,
         seat: Int,
         legal: List<EuchreMove>,
+        profile: AiProfile,
     ): EuchreMove {
         val hand = state.hands[seat]
         val callable = legal.filterIsInstance<CallTrump>()
         if (callable.isEmpty()) return legal.first()
+        val plain = callable.filter { !it.alone }
 
-        val best = callable
-            .filter { !it.alone }
-            .maxByOrNull { handValue(hand, it.suit) }
-            ?: callable.first()
-        val bestValue = handValue(hand, best.suit)
+        val best: CallTrump
+        val bestValue: Int
+        if (profile.level == AiLevel.EASY) {
+            val noisy = plain.associateWith { handValue(hand, it.suit) + profile.random.nextInt(-5, 6) }
+            val ranked = plain.sortedByDescending { noisy.getValue(it) }
+            best = if (ranked.isEmpty()) callable.first() else profile.pick(ranked)
+            bestValue = noisy[best] ?: handValue(hand, best.suit)
+        } else {
+            best = plain.maxByOrNull { handValue(hand, it.suit) } ?: callable.first()
+            bestValue = handValue(hand, best.suit)
+        }
+
+        val bidThreshold = profile.bolder(BID_THRESHOLD, -3.0)
+        var aloneThreshold = profile.bolder(ALONE_THRESHOLD, -3.0)
+        if (profile.level == AiLevel.EASY) aloneThreshold += 4.0
 
         val alone = callable.firstOrNull { it.alone && it.suit == best.suit }
-        if (alone != null && bestValue >= difficulty.aloneThreshold) return alone
-        if (bestValue >= difficulty.bidThreshold) return best
+        val alonePlausible = profile.level != AiLevel.HARD || soundAlone(hand, best.suit)
+        if (alone != null && bestValue >= aloneThreshold && alonePlausible) return alone
+        if (bestValue >= bidThreshold) return best
         // Stuck as dealer with a bad hand: name the least-bad suit anyway.
         return if (legal.contains(Pass)) Pass else best
     }
 
-    private fun chooseDiscard(state: EuchreState, legal: List<EuchreMove>): EuchreMove {
+    private fun chooseDiscard(
+        state: EuchreState,
+        legal: List<EuchreMove>,
+        profile: AiProfile,
+    ): EuchreMove {
         val trump = requireNotNull(state.trump)
         val discards = legal.filterIsInstance<Discard>()
         // Prefer to void a side suit entirely; otherwise throw the lowest card.
         val hand = state.hands[state.dealer]
         val suitCounts = hand.groupingBy { effectiveSuit(it, trump) }.eachCount()
-        return discards.minByOrNull { discard ->
+        fun keepScore(discard: Discard): Int {
             val card = discard.card
             val suit = effectiveSuit(card, trump)
             val trumpPenalty = if (suit == trump) 1000 else 0
             val singletonBonus = if (suitCounts[suit] == 1 && card.rank != Rank.ACE) -20 else 0
-            trumpPenalty + singletonBonus + card.rank.order
-        } ?: discards.first()
+            return trumpPenalty + singletonBonus + card.rank.order
+        }
+        return when (profile.level) {
+            AiLevel.HARD -> discards
+                .sortedWith(
+                    compareByDescending<Discard> { handValue(hand - it.card, trump) }
+                        .thenBy { keepScore(it) },
+                )
+                .first()
+            AiLevel.EASY -> profile.pick(discards.sortedBy { keepScore(it) })
+            AiLevel.MEDIUM -> discards.minByOrNull { keepScore(it) } ?: discards.first()
+        }
     }
 
     // -----------------------------------------------------------------------
@@ -144,12 +234,40 @@ class EuchreAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
         state: EuchreState,
         seat: Int,
         legal: List<EuchreMove>,
+        profile: AiProfile,
+        seen: Set<Card>,
     ): EuchreMove {
         val trump = requireNotNull(state.trump)
         val cards = legal.filterIsInstance<PlayCard>().map { it.card }
         if (cards.size == 1) return PlayCard(cards.first())
 
-        if (state.trick.isEmpty()) return PlayCard(chooseLead(cards, trump))
+        val maker = state.maker
+        val defending = maker != null && teamOf(maker) != teamOf(seat)
+        val makerSide = maker != null && teamOf(maker) == teamOf(seat)
+        // A spoiler on the defence leads trump to cut the maker's trump down.
+        val spiteful = defending && profile.style.spite >= 0.5
+
+        val best = if (profile.level == AiLevel.HARD) {
+            hardPlay(state, seat, cards, trump, seen, makerSide, spiteful)
+        } else {
+            standardPlay(state, seat, cards, trump, spiteful, profile)
+        }
+        if (profile.level != AiLevel.EASY) return PlayCard(best)
+
+        // Easy sometimes burns a high card it should have kept.
+        val ranked = listOf(best) + (cards - best).sortedByDescending { trumpStrength(it, trump) }
+        return PlayCard(profile.pick(ranked))
+    }
+
+    private fun standardPlay(
+        state: EuchreState,
+        seat: Int,
+        cards: List<Card>,
+        trump: Suit,
+        spiteful: Boolean,
+        profile: AiProfile,
+    ): Card {
+        if (state.trick.isEmpty()) return chooseLead(cards, trump, spiteful)
 
         val ledSuit = effectiveSuit(state.trick.first().card, trump)
         val bestSoFar = state.trick.maxBy { trickStrength(it.card, trump, ledSuit) }
@@ -157,22 +275,120 @@ class EuchreAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
         val lastToPlay = state.trick.size == state.activeSeatCount - 1
 
         // Partner already has it and we are last: never waste a good card.
-        if (partnerWinning && lastToPlay) return PlayCard(lowest(cards, trump, ledSuit))
+        // (Easy sometimes forgets and overtakes anyway.)
+        if (partnerWinning && lastToPlay && !profile.slips()) {
+            return lowest(cards, trump, ledSuit)
+        }
 
         val bestStrength = trickStrength(bestSoFar.card, trump, ledSuit)
         val winners = cards.filter { trickStrength(it, trump, ledSuit) > bestStrength }
         return when {
             // Take it as cheaply as we can.
-            winners.isNotEmpty() && (!partnerWinning || !lastToPlay) ->
-                PlayCard(winners.minBy { trickStrength(it, trump, ledSuit) })
+            winners.isNotEmpty() && (!partnerWinning || !lastToPlay || profile.level == AiLevel.EASY) ->
+                winners.minBy { trickStrength(it, trump, ledSuit) }
 
-            else -> PlayCard(lowest(cards, trump, ledSuit))
+            else -> lowest(cards, trump, ledSuit)
         }
     }
 
-    private fun chooseLead(cards: List<Card>, trump: Suit): Card {
+    /**
+     * The strong player's choice. It remembers which cards have gone, so it
+     * knows when a card is the best one left, and it lets a partner who is
+     * already safe take the trick.
+     */
+    private fun hardPlay(
+        state: EuchreState,
+        seat: Int,
+        cards: List<Card>,
+        trump: Suit,
+        seen: Set<Card>,
+        makerSide: Boolean,
+        spiteful: Boolean,
+    ): Card {
+        val hand = state.hands[seat]
+        // Cards that could still be in somebody else's hand.
+        val unseen = Decks.euchre(state.options.deckSize).filter { it !in seen && it !in hand }
+        val unseenTrump = unseen.count { isTrumpCard(it, trump) }
+
+        /** No unseen card of its suit beats it; trump is ignored for a side-suit card. */
+        fun isBoss(card: Card): Boolean {
+            val led = effectiveSuit(card, trump)
+            val mine = trickStrength(card, trump, led)
+            val cardIsTrump = isTrumpCard(card, trump)
+            return unseen.none { other ->
+                (cardIsTrump || !isTrumpCard(other, trump)) && trickStrength(other, trump, led) > mine
+            }
+        }
+
+        if (state.trick.isEmpty()) {
+            val myTrump = cards.filter { isTrumpCard(it, trump) }
+            val off = cards.filter { !isTrumpCard(it, trump) }
+            val bossTrump = myTrump.filter { isBoss(it) }.maxByOrNull { trumpStrength(it, trump) }
+            if (unseenTrump > 0 && bossTrump != null && (makerSide || myTrump.size >= 2)) return bossTrump
+            if (unseenTrump > 0 && makerSide && myTrump.size >= 3) {
+                return myTrump.maxBy { trumpStrength(it, trump) }
+            }
+            if (unseenTrump > 0 && spiteful && myTrump.size >= 2) {
+                return myTrump.maxBy { trumpStrength(it, trump) }
+            }
+            val bossOff = off.filter { isBoss(it) }.maxByOrNull { it.rank.order }
+            if (bossOff != null) return bossOff
+            if (bossTrump != null) return bossTrump
+            if (off.isEmpty()) return cards.maxBy { trumpStrength(it, trump) }
+            if (makerSide) return chooseLead(cards, trump, false)
+            // On the defence with nothing sure: lead low from the shortest suit,
+            // which keeps the aces and sets up a ruff.
+            val counts = hand.groupingBy { effectiveSuit(it, trump) }.eachCount()
+            return off.minWith(
+                compareBy<Card>({ counts[effectiveSuit(it, trump)] ?: 0 }, { it.rank.order }),
+            )
+        }
+
+        val ledSuit = effectiveSuit(state.trick.first().card, trump)
+        val bestSoFar = state.trick.maxBy { trickStrength(it.card, trump, ledSuit) }
+        val partnerWinning = partnerOf(bestSoFar.seat) == seat
+        val remaining = state.activeSeatCount - state.trick.size - 1
+        val bestStrength = trickStrength(bestSoFar.card, trump, ledSuit)
+        val winners = cards.filter { trickStrength(it, trump, ledSuit) > bestStrength }
+
+        fun safe(card: Card): Boolean =
+            remaining == 0 ||
+                unseen.none { trickStrength(it, trump, ledSuit) > trickStrength(card, trump, ledSuit) }
+
+        fun throwAway(): Card {
+            // A card that neither follows nor trumps costs nothing; shed the
+            // lowest one that is not a sure winner later.
+            val junk = cards.filter { trickStrength(it, trump, ledSuit) == 0 }
+            if (junk.isNotEmpty()) {
+                return junk.minBy { it.rank.order + if (isBoss(it)) 100 else 0 }
+            }
+            return lowest(cards, trump, ledSuit)
+        }
+
+        if (partnerWinning) {
+            val partnerSecure = remaining == 0 ||
+                unseen.none { trickStrength(it, trump, ledSuit) > bestStrength }
+            if (partnerSecure) return throwAway()
+            // Partner may yet be beaten: back them with a side-suit card, but do
+            // not spend trump on top of their lead.
+            val backing = winners.filter { !isTrumpCard(it, trump) }
+                .minByOrNull { trickStrength(it, trump, ledSuit) }
+            return backing ?: throwAway()
+        }
+        if (winners.isNotEmpty()) {
+            val sure = winners.filter { safe(it) }
+            return (sure.ifEmpty { winners }).minBy { trickStrength(it, trump, ledSuit) }
+        }
+        return throwAway()
+    }
+
+    private fun chooseLead(cards: List<Card>, trump: Suit, spiteTrump: Boolean): Card {
         // Right bower up front pulls trump and protects the left.
         cards.firstOrNull { isRightBower(it, trump) }?.let { return it }
+        if (spiteTrump) {
+            val trumps = cards.filter { isTrumpCard(it, trump) }
+            if (trumps.size >= 2) return trumps.maxBy { trumpStrength(it, trump) }
+        }
         // Off-suit aces are most likely to survive early.
         cards.filter { !isTrumpCard(it, trump) && it.rank == Rank.ACE }
             .maxByOrNull { it.rank.order }

app/src/main/java/org/prolibertate/games/game/golf/GolfAi.kt

@@ -1,49 +1,78 @@
 package org.prolibertate.games.game.golf
 
+import org.prolibertate.games.game.cards.Card
+import org.prolibertate.games.game.cards.Rank
+import org.prolibertate.games.game.cards.Suit
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
- * A heuristic Golf opponent.
+ * A Golf opponent whose strength and temperament come from the seat's profile.
  *
- * It takes a card off the discard pile when it is cheap or completes a column,
- * replaces its worst known card, and turns cards over when it has nothing
- * better to do. It never peeks at a face-down card — including its own, which
- * it genuinely does not know.
+ * MEDIUM (the default) takes a card off the discard pile when it is cheap or
+ * completes a column, replaces its worst known card, and turns cards over when
+ * it has nothing better to do. It never peeks at a face-down card — including
+ * its own, which it genuinely does not know.
+ *
+ * EASY is pickier about the discard pile, closes the hole from a poor position
+ * and sometimes makes a worse play than it saw. HARD keeps count of every card
+ * it has seen (turned-over grids and the discard pile), values its unknown
+ * cards at what is still unseen, and chooses each draw and placement by how
+ * much it lowers its expected score, including the chance of completing a
+ * column and whether closing the hole would leave it ahead.
  */
-class GolfAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
+class GolfAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) :
     GameAi<GolfState, GolfMove> {
 
-    enum class Difficulty(
+    /** The tuning the standard plan runs on. */
+    private class Params(
         /** Take a discard worth this or less rather than gambling on the stock. */
         val discardAppetite: Int,
         /** Only close the hole out when the visible score is at or under this. */
         val closeOutBelow: Int,
-    ) {
-        EASY(discardAppetite = 2, closeOutBelow = 30),
-        NORMAL(discardAppetite = 5, closeOutBelow = 16),
-        CAUTIOUS(discardAppetite = 7, closeOutBelow = 12),
-    }
+    )
 
     override fun chooseMove(state: GolfState, seat: Int, legal: List<GolfMove>): GolfMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
-        return when (state.phase) {
-            GolfPhase.SETUP -> chooseReveal(state, seat, legal)
-            GolfPhase.DRAW -> chooseDraw(state, seat, legal)
-            GolfPhase.PLACE -> choosePlace(state, seat, legal)
+        val profile = profiles.of(seat)
+        val move = when (state.phase) {
+            GolfPhase.SETUP -> chooseReveal(state, seat, legal, profile)
+            GolfPhase.DRAW -> chooseDraw(state, seat, legal, profile)
+            GolfPhase.PLACE -> choosePlace(state, seat, legal, profile)
             else -> legal.first()
         }
+        return if (move in legal) move else legal.first()
     }
 
+    private fun paramsFor(level: AiLevel): Params = when (level) {
+        AiLevel.EASY -> EASY
+        else -> NORMAL
+    }
+
+    // -----------------------------------------------------------------------
+    // Setup
+    // -----------------------------------------------------------------------
+
     /**
      * Opening reveals are spread across columns. Two cards in the same column
      * tell you whether that column can be cancelled but nothing about the rest
      * of the board; one in each column leaves more of the grid legible.
      */
-    private fun chooseReveal(state: GolfState, seat: Int, legal: List<GolfMove>): GolfMove {
+    private fun chooseReveal(
+        state: GolfState,
+        seat: Int,
+        legal: List<GolfMove>,
+        profile: AiProfile,
+    ): GolfMove {
         val reveals = legal.filterIsInstance<RevealCard>()
         if (reveals.isEmpty()) return legal.first()
         val options = state.options
 
+        // A distracted player turns over whatever is nearest.
+        if (profile.slips()) return reveals[profile.random.nextInt(reveals.size)]
+
         val untouchedColumn = reveals.firstOrNull { move ->
             val column = move.index % options.cols
             (0 until options.rows).none { row ->
@@ -53,17 +82,61 @@ class GolfAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
         return untouchedColumn ?: reveals.first()
     }
 
-    private fun chooseDraw(state: GolfState, seat: Int, legal: List<GolfMove>): GolfMove {
+    // -----------------------------------------------------------------------
+    // Drawing
+    // -----------------------------------------------------------------------
+
+    private fun chooseDraw(
+        state: GolfState,
+        seat: Int,
+        legal: List<GolfMove>,
+        profile: AiProfile,
+    ): GolfMove {
+        if (profile.level == AiLevel.HARD) return hardDraw(state, seat, legal, profile)
+
+        val params = paramsFor(profile.level)
         val top = state.discard.lastOrNull()
+        var take = false
         if (top != null && legal.contains(DrawFromDiscard)) {
-            val worthTaking = golfValue(top) <= difficulty.discardAppetite ||
+            take = golfValue(top) <= params.discardAppetite ||
                 completesColumn(state, seat, top) != null
-            if (worthTaking) return DrawFromDiscard
         }
+        // Easy sometimes flips its own decision.
+        if (profile.slips() && legal.contains(DrawFromDiscard) && legal.contains(DrawFromStock)) take = !take
+        if (take && legal.contains(DrawFromDiscard)) return DrawFromDiscard
         return if (legal.contains(DrawFromStock)) DrawFromStock else legal.first()
     }
 
-    private fun choosePlace(state: GolfState, seat: Int, legal: List<GolfMove>): GolfMove {
+    // -----------------------------------------------------------------------
+    // Placing
+    // -----------------------------------------------------------------------
+
+    private fun choosePlace(
+        state: GolfState,
+        seat: Int,
+        legal: List<GolfMove>,
+        profile: AiProfile,
+    ): GolfMove {
+        if (profile.level == AiLevel.HARD) return hardPlace(state, seat, legal, profile)
+
+        val best = standardPlace(state, seat, legal, paramsFor(profile.level))
+        if (profile.level != AiLevel.EASY || !profile.slips()) return best
+
+        // A lapse: turn something over or cover a face-down card instead.
+        val alternatives = buildList<GolfMove> {
+            addAll(legal.filterIsInstance<DiscardAndFlip>())
+            addAll(legal.filterIsInstance<ReplaceCard>().filter { !state.revealed[seat][it.index] })
+        }.filter { it != best }
+        if (alternatives.isEmpty()) return best
+        return alternatives[profile.random.nextInt(alternatives.size)]
+    }
+
+    private fun standardPlace(
+        state: GolfState,
+        seat: Int,
+        legal: List<GolfMove>,
+        params: Params,
+    ): GolfMove {
         val drawn = state.drawn ?: return legal.first()
         val replacements = legal.filterIsInstance<ReplaceCard>()
         val flips = legal.filterIsInstance<DiscardAndFlip>()
@@ -90,7 +163,7 @@ class GolfAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
             val showing = state.grids[seat].indices
                 .filter { state.revealed[seat][it] }
                 .sumOf { golfValue(state.grids[seat][it]) }
-            if (showing > difficulty.closeOutBelow) return DiscardOnly
+            if (showing > params.closeOutBelow) return DiscardOnly
         }
 
         // Nothing worth swapping: throw it and turn something over instead,
@@ -106,7 +179,7 @@ class GolfAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
     private fun completesColumn(
         state: GolfState,
         seat: Int,
-        card: org.prolibertate.games.game.cards.Card,
+        card: Card,
     ): Int? {
         val options = state.options
         for (col in 0 until options.cols) {
@@ -122,4 +195,229 @@ class GolfAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
         }
         return null
     }
+
+    // -----------------------------------------------------------------------
+    // Hard
+    // -----------------------------------------------------------------------
+
+    /**
+     * What this seat knows about the hole: the cards in every turned-over cell,
+     * the discard pile and (on its own turn) the card in its hand. Everything
+     * else — face-down cells and the stock — is valued at the average of the
+     * cards not yet seen.
+     */
+    private class Board(private val state: GolfState, seat: Int) {
+        val options = state.options
+        private val counts = HashMap<Rank, Int>()
+        private val unseenTotal: Int
+        val unknownValue: Double
+        private val lines: List<List<Int>>
+
+        init {
+            val copies = if (options.playerCount * options.gridSize > 24) 2 else 1
+            Rank.standard.forEach { counts[it] = 4 * copies }
+            fun see(card: Card) {
+                counts[card.rank] = maxOf(0, (counts[card.rank] ?: 0) - 1)
+            }
+            state.grids.forEachIndexed { player, grid ->
+                grid.forEachIndexed { index, card -> if (state.revealed[player][index]) see(card) }
+            }
+            state.discard.forEach { see(it) }
+            if (seat == state.turn) state.drawn?.let { see(it) }
+
+            unseenTotal = counts.values.sum()
+            unknownValue = if (unseenTotal == 0) {
+                5.0
+            } else {
+                counts.entries.sumOf { (rank, n) -> n * valueOf(rank).toDouble() } / unseenTotal
+            }
+
+            val columns = (0 until options.cols).map { col -> (0 until options.rows).map { it * options.cols + col } }
+            val rows = if (options.rows == 3) {
+                (0 until options.rows).map { row -> (0 until options.cols).map { row * options.cols + it } }
+            } else {
+                emptyList()
+            }
+            lines = columns + rows
+        }
+
+        fun valueOf(rank: Rank): Int = golfValue(Card(rank, Suit.SPADES))
+
+        private fun chanceOf(rank: Rank): Double =
+            if (unseenTotal == 0) 0.0 else (counts[rank] ?: 0).toDouble() / unseenTotal
+
+        /** A seat's cells as far as this seat can see them; null is a face-down card. */
+        fun cellsOf(player: Int): List<Card?> =
+            state.grids[player].mapIndexed { index, card -> if (state.revealed[player][index]) card else null }
+
+        /** Expected score of a grid, counting the chance that an unknown cell completes a line. */
+        fun expected(cells: List<Card?>): Double {
+            val cancelled = BooleanArray(cells.size)
+            for (line in lines) {
+                val line0 = line.map { cells[it] }
+                if (line0.all { it != null } && line0.map { it!!.rank }.distinct().size == 1) {
+                    line.forEach { cancelled[it] = true }
+                }
+            }
+            var total = 0.0
+            for (index in cells.indices) {
+                if (cancelled[index]) continue
+                val card = cells[index]
+                total += if (card != null) golfValue(card).toDouble() else unknownValue
+            }
+            for (line in lines) {
+                if (line.any { cancelled[it] }) continue
+                val unknown = line.filter { cells[it] == null }
+                if (unknown.size != 1) continue
+                val known = line.mapNotNull { cells[it] }
+                if (known.isEmpty() || known.map { it.rank }.distinct().size != 1) continue
+                val rank = known.first().rank
+                total -= chanceOf(rank) * (known.sumOf { golfValue(it) } + valueOf(rank))
+            }
+            return total
+        }
+
+        /** Unseen ranks and how many of each remain, for averaging over a stock draw. */
+        fun unseenRanks(): List<Pair<Rank, Int>> = counts.entries.filter { it.value > 0 }.map { it.key to it.value }
+
+        /** How likely turning over [index] is to finish a line that already shows one rank. */
+        fun matchHope(cells: List<Card?>, index: Int): Double {
+            var best = 0.0
+            for (line in lines) {
+                if (index !in line) continue
+                val others = line.filter { it != index }
+                val known = others.mapNotNull { cells[it] }
+                if (known.size != others.size || known.isEmpty()) continue
+                if (known.map { it.rank }.distinct().size != 1) continue
+                val rank = known.first().rank
+                best = maxOf(best, chanceOf(rank) * (known.sumOf { golfValue(it) } + valueOf(rank)))
+            }
+            return best
+        }
+    }
+
+    private fun effectiveBoldness(state: GolfState, seat: Int, profile: AiProfile): Double {
+        // Well behind late in the match: take more chances.
+        val holesLeft = state.options.holes - state.hole - 1
+        val best = (0 until state.options.playerCount).filter { it != seat }.minOf { state.scores[it] }
+        val behind = state.scores[seat] - best
+        return profile.style.boldness + if (holesLeft <= 2 && behind > 8) 0.3 else 0.0
+    }
+
+    /** Scores one way of using the card in hand: lower grid score, more information, less risk. */
+    private class Planner(
+        val state: GolfState,
+        val seat: Int,
+        val board: Board,
+        val boldness: Double,
+        val spite: Double,
+    ) {
+        private val cells = board.cellsOf(seat)
+        private val before = board.expected(cells)
+        private val hidden = cells.count { it == null }
+        private val finalTurn = state.closedBy != null
+        private val rivalBest: Double = (0 until state.options.playerCount)
+            .filter { it != seat }
+            .minOfOrNull { board.expected(board.cellsOf(it)) } ?: before
+
+        private fun closePenalty(after: Double): Double {
+            if (finalTurn) return 0.0
+            val margin = 2.0 - 2.0 * boldness
+            return maxOf(0.0, after - (rivalBest - margin)) * 0.8
+        }
+
+        private fun handoff(drawn: Card): Double =
+            maxOf(0.0, 3.0 - golfValue(drawn)) * 0.15 * (1.0 + spite)
+
+        private fun flipBonus(index: Int): Double =
+            if (finalTurn || hidden <= 1) 0.0 else 0.35 + 0.5 * board.matchHope(cells, index)
+
+        /** Utility of [move] with [drawn] in hand; higher is better. */
+        fun utility(move: GolfMove, drawn: Card): Double = when (move) {
+            is ReplaceCard -> {
+                val after = cells.toMutableList().also { it[move.index] = drawn }
+                val closes = hidden == 1 && cells[move.index] == null
+                val afterScore = board.expected(after)
+                before - afterScore - if (closes) closePenalty(afterScore) else 0.0
+            }
+
+            is DiscardAndFlip -> {
+                val closes = hidden == 1
+                flipBonus(move.index) - handoff(drawn) - if (closes) closePenalty(before) else 0.0
+            }
+
+            is DiscardOnly -> -handoff(drawn)
+            else -> Double.NEGATIVE_INFINITY
+        }
+
+        /** The moves open to a card in hand, standing in for the legal list during look-ahead. */
+        fun candidates(fromDiscard: Boolean, putt: Boolean): List<GolfMove> = buildList {
+            for (index in cells.indices) add(ReplaceCard(index))
+            if (!fromDiscard) {
+                for (index in cells.indices) if (cells[index] == null) add(DiscardAndFlip(index))
+                if (putt && hidden == 1) add(DiscardOnly)
+            }
+        }
+    }
+
+    private fun planner(state: GolfState, seat: Int, profile: AiProfile): Planner {
+        val board = Board(state, seat)
+        return Planner(state, seat, board, effectiveBoldness(state, seat, profile), profile.style.spite)
+    }
+
+    private fun hardDraw(
+        state: GolfState,
+        seat: Int,
+        legal: List<GolfMove>,
+        profile: AiProfile,
+    ): GolfMove {
+        val top = state.discard.lastOrNull()
+        val canStock = legal.contains(DrawFromStock)
+        val canDiscard = legal.contains(DrawFromDiscard) && top != null
+        if (!canDiscard) return if (canStock) DrawFromStock else legal.first()
+        if (!canStock) return DrawFromDiscard
+
+        val plan = planner(state, seat, profile)
+        val putt = state.options.lineUpFinalPutt
+
+        // The discard has to be used, so it is only as good as its best replacement.
+        val takeUtility = plan.candidates(fromDiscard = true, putt = putt)
+            .maxOf { plan.utility(it, top!!) } +
+            profile.style.spite * 0.3 * maxOf(0, 4 - golfValue(top!!))
+
+        // The stock is a gamble on the average of what is left, but can be thrown back.
+        val ranks = plan.board.unseenRanks()
+        val total = ranks.sumOf { it.second }
+        val stockUtility = if (total == 0) {
+            Double.NEGATIVE_INFINITY
+        } else {
+            val moves = plan.candidates(fromDiscard = false, putt = putt)
+            ranks.sumOf { (rank, n) ->
+                val sample = Card(rank, Suit.SPADES)
+                n * moves.maxOf { plan.utility(it, sample) }
+            } / total
+        }
+
+        // A bold seat wants a clear edge before it passes up the unknown card.
+        val take = takeUtility - plan.boldness * 0.4 > stockUtility
+        return if (take) DrawFromDiscard else DrawFromStock
+    }
+
+    private fun hardPlace(
+        state: GolfState,
+        seat: Int,
+        legal: List<GolfMove>,
+        profile: AiProfile,
+    ): GolfMove {
+        val drawn = state.drawn ?: return legal.first()
+        val plan = planner(state, seat, profile)
+        val options = legal.filter { it is ReplaceCard || it is DiscardAndFlip || it is DiscardOnly }
+        if (options.isEmpty()) return legal.first()
+        return options.maxBy { plan.utility(it, drawn) }
+    }
+
+    private companion object {
+        private val EASY = Params(discardAppetite = 2, closeOutBelow = 30)
+        private val NORMAL = Params(discardAppetite = 5, closeOutBelow = 16)
+    }
 }

app/src/main/java/org/prolibertate/games/game/hearts/HeartsAi.kt

@@ -3,35 +3,98 @@ package org.prolibertate.games.game.hearts
 import org.prolibertate.games.game.cards.Card
 import org.prolibertate.games.game.cards.Rank
 import org.prolibertate.games.game.cards.Suit
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
- * A cautious Hearts player.
+ * A Hearts player whose strength and temperament come from the seat's profile.
  *
- * It plays to avoid points rather than to shoot the moon: it sheds the queen of
- * spades and the guards above her at the first opportunity, ducks under whoever
- * is winning a trick when it can, and takes a trick as cheaply as possible when
- * it cannot get out of the way. That is enough to make it an opponent worth
- * beating without pretending to read the table.
+ * MEDIUM (the default) avoids points rather than shooting the moon: it sheds
+ * the queen of spades and the guards above her at the first opportunity, ducks
+ * under whoever is winning a trick when it can, and takes a trick as cheaply as
+ * possible when it cannot get out of the way.
+ *
+ * EASY plays the same plan but sometimes settles for a worse card in each
+ * decision. HARD counts the penalty cards and spades that have been taken,
+ * leads by risk, sheds by what is still dangerous, shoots the moon when its
+ * hand and the table allow it and tries to break a moon somebody else is on.
+ * It only uses its own hand and what the table has shown.
  */
-object HeartsAi : GameAi<HeartsState, HeartsMove> {
+class HeartsAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) : GameAi<HeartsState, HeartsMove> {
 
     override fun chooseMove(state: HeartsState, seat: Int, legal: List<HeartsMove>): HeartsMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
-        return when (state.phase) {
-            HeartsPhase.PASSING -> choosePass(state, seat, legal)
-            else -> choosePlay(state, seat, legal)
+        val profile = profiles.of(seat)
+        val move = when (state.phase) {
+            HeartsPhase.PASSING -> choosePass(state, seat, legal, profile)
+            else -> choosePlay(state, seat, legal, profile)
         }
+        return if (move in legal) move else legal.first()
     }
 
-    private fun choosePass(state: HeartsState, seat: Int, legal: List<HeartsMove>): HeartsMove {
+    // -----------------------------------------------------------------------
+    // Passing
+    // -----------------------------------------------------------------------
+
+    private fun choosePass(
+        state: HeartsState,
+        seat: Int,
+        legal: List<HeartsMove>,
+        profile: AiProfile,
+    ): HeartsMove {
         val hand = state.hands[seat]
-        val wanted = hand.sortedByDescending { passDanger(it, hand) }.take(PASS_SIZE).toSet()
+        val wanted: Set<Card> = when (profile.level) {
+            AiLevel.MEDIUM -> hand.sortedByDescending { passDanger(it, hand) }.take(PASS_SIZE).toSet()
+            AiLevel.EASY -> easyPass(hand, profile)
+            AiLevel.HARD -> hardPass(state, hand, profile)
+        }
         // Prefer the exact bundle the heuristic asked for, but fall back rather
         // than fail: the caller's list is what the host will validate against.
         return legal.firstOrNull { it is PassCards && it.cards.toSet() == wanted } ?: legal.first()
     }
 
+    private fun easyPass(hand: List<Card>, profile: AiProfile): Set<Card> {
+        val ranked = hand.sortedByDescending { passDanger(it, hand) }
+        val chosen = ranked.take(PASS_SIZE).toMutableList()
+        if (profile.slips()) {
+            // Keeps one of the cards it should have sent.
+            val spare = ranked.drop(PASS_SIZE).take(4)
+            if (spare.isNotEmpty()) {
+                chosen[profile.random.nextInt(chosen.size)] = spare[profile.random.nextInt(spare.size)]
+            }
+        }
+        return chosen.toSet()
+    }
+
+    private fun hardPass(state: HeartsState, hand: List<Card>, profile: AiProfile): Set<Card> {
+        if (state.options.allowShootTheMoon && moonHand(hand, profile)) {
+            // Keep every high card and heart, send the three smallest.
+            val junk = hand.filter { !(it.suit == Suit.HEARTS && it.rank.order >= 10) }
+                .sortedBy { it.rank.order }
+            if (junk.size >= PASS_SIZE) return junk.take(PASS_SIZE).toSet()
+        }
+        return hand.sortedByDescending { passDanger(it, hand) + shortSuitBonus(it, hand) }
+            .take(PASS_SIZE).toSet()
+    }
+
+    /** Sending away the last cards of a short suit makes a void, which is worth a lot when discarding. */
+    private fun shortSuitBonus(card: Card, hand: List<Card>): Int {
+        if (card.suit == Suit.HEARTS || card.suit == Suit.SPADES) return 0
+        val length = hand.count { it.suit == card.suit }
+        return if (length > 3) 0 else (4 - length) * 8
+    }
+
+    /** A hand worth trying to take all twenty-six with: lots of high cards and a heart suit to run. */
+    private fun moonHand(hand: List<Card>, profile: AiProfile): Boolean {
+        val hearts = hand.count { it.suit == Suit.HEARTS }
+        if (hearts < 4) return false
+        val strength = hand.count { it.rank.order >= Rank.QUEEN.order } +
+            hand.count { it.suit == Suit.HEARTS && it.rank.order >= 9 }
+        return strength >= profile.bolder(8.0, -2.0)
+    }
+
     /**
      * How much a card wants to be somebody else's problem.
      *
@@ -52,27 +115,41 @@ object HeartsAi : GameAi<HeartsState, HeartsMove> {
         }
     }
 
-    private fun choosePlay(state: HeartsState, seat: Int, legal: List<HeartsMove>): HeartsMove {
+    // -----------------------------------------------------------------------
+    // Playing
+    // -----------------------------------------------------------------------
+
+    private fun choosePlay(
+        state: HeartsState,
+        seat: Int,
+        legal: List<HeartsMove>,
+        profile: AiProfile,
+    ): HeartsMove {
         val cards = legal.filterIsInstance<PlayCard>().map { it.card }
         if (cards.isEmpty()) return legal.first()
-        val chosen = when {
-            state.trick.isEmpty() -> lead(cards)
-            else -> follow(state, cards)
+        val chosen = when (profile.level) {
+            AiLevel.MEDIUM -> standardRanking(state, cards).first()
+            AiLevel.EASY -> profile.pick(standardRanking(state, cards))
+            AiLevel.HARD -> hardPlay(state, seat, cards, profile)
         }
-        return PlayCard(chosen)
+        return PlayCard(if (chosen in cards) chosen else cards.first())
     }
 
+    /** The standard plan's preferences, best first. */
+    private fun standardRanking(state: HeartsState, cards: List<Card>): List<Card> =
+        if (state.trick.isEmpty()) rankedLead(cards) else rankedFollow(state, cards)
+
     /**
      * Leads low, and keeps off spades while the queen is still unaccounted for —
      * leading a spade under her is how a seat ends up holding her.
      */
-    private fun lead(cards: List<Card>): Card {
-        val offSpades = cards.filter { it.suit != Suit.SPADES }
-        val pool = if (offSpades.isNotEmpty()) offSpades else cards
-        return pool.minBy { it.rank.order }
+    private fun rankedLead(cards: List<Card>): List<Card> {
+        val offSpades = cards.filter { it.suit != Suit.SPADES }.sortedBy { it.rank.order }
+        val spades = cards.filter { it.suit == Suit.SPADES }.sortedBy { it.rank.order }
+        return offSpades + spades
     }
 
-    private fun follow(state: HeartsState, cards: List<Card>): Card {
+    private fun rankedFollow(state: HeartsState, cards: List<Card>): List<Card> {
         val led = state.trick.first().card.suit
         val best = state.trick.maxOf { trickStrength(it.card, led) }
         val pot = pointsIn(state.trick.map { it.card })
@@ -82,27 +159,189 @@ object HeartsAi : GameAi<HeartsState, HeartsMove> {
         if (ducking.isNotEmpty()) {
             // Safe under the current winner, so shed the most expensive card
             // that still loses.
-            return ducking.maxBy { it.rank.order + pointsOf(it) * 2 }
+            val rest = (cards - ducking.toSet()).sortedBy { it.rank.order }
+            return ducking.sortedByDescending { it.rank.order + pointsOf(it) * 2 } + rest
         }
 
-        val void = cards.none { it.suit == led }
-        if (void) return discard(cards)
-
         // Every card left would take the trick. Playing last with nothing in the
         // pot means it costs nothing, so there is no reason to waste a high card
         // on it; otherwise take it as cheaply as possible.
-        return if (last && pot == 0) cards.maxBy { it.rank.order } else cards.minBy { it.rank.order }
+        return if (last && pot == 0) {
+            cards.sortedByDescending { it.rank.order }
+        } else {
+            cards.sortedBy { it.rank.order }
+        }
     }
 
-    /** Void in the led suit and free to throw anything, so throw the worst of it. */
-    private fun discard(cards: List<Card>): Card {
-        cards.firstOrNull { it == QUEEN_OF_SPADES }?.let { return it }
-        cards.filter { it.suit == Suit.SPADES && it.rank.order > Rank.QUEEN.order }
-            .maxByOrNull { it.rank.order }
-            ?.let { return it }
-        cards.filter { it.suit == Suit.HEARTS }
-            .maxByOrNull { it.rank.order }
-            ?.let { return it }
-        return cards.maxBy { it.rank.order }
+    // -----------------------------------------------------------------------
+    // Hard play
+    // -----------------------------------------------------------------------
+
+    /** What this seat can legitimately know: its hand, and every card taken or on the table. */
+    private class Memory(state: HeartsState, seat: Int) {
+        val hand: List<Card> = state.hands[seat]
+        val unseen: List<Card>
+
+        init {
+            val seen = HashSet<Card>()
+            state.taken.forEach { seen.addAll(it) }
+            state.trick.forEach { seen.add(it.card) }
+            seen.addAll(hand)
+            unseen = FULL_DECK.filter { it !in seen }
+        }
+
+        val queenUnseen: Boolean get() = QUEEN_OF_SPADES in unseen
+
+        fun unseenIn(suit: Suit): Int = unseen.count { it.suit == suit }
+
+        fun unseenHigher(card: Card): Int =
+            unseen.count { it.suit == card.suit && it.rank.order > card.rank.order }
+    }
+
+    private fun hardPlay(state: HeartsState, seat: Int, cards: List<Card>, profile: AiProfile): Card {
+        val mem = Memory(state, seat)
+        if (isShooting(state, seat, mem, profile)) return shootPlay(state, cards, mem)
+        return if (state.trick.isEmpty()) {
+            hardLead(cards, mem, profile)
+        } else {
+            hardFollow(state, seat, cards, mem, profile)
+        }
+    }
+
+    /** Holding every point so far, with a hand that can keep winning. */
+    private fun isShooting(state: HeartsState, seat: Int, mem: Memory, profile: AiProfile): Boolean {
+        if (!state.options.allowShootTheMoon) return false
+        val all = pointsIn(state.taken.flatten())
+        val mine = pointsIn(state.taken[seat])
+        if (mine != all) return false
+        if (mine == 0 && !moonHand(mem.hand, profile)) return false
+        if (mem.hand.isEmpty()) return false
+        val masters = mem.hand.count { mem.unseenHigher(it) == 0 }
+        return masters.toDouble() / mem.hand.size >= profile.bolder(0.6, -0.15)
+    }
+
+    private fun shootPlay(state: HeartsState, cards: List<Card>, mem: Memory): Card {
+        if (state.trick.isEmpty()) {
+            val masters = cards.filter { mem.unseenHigher(it) == 0 }
+            if (masters.isNotEmpty()) {
+                return masters.sortedWith(
+                    compareBy<Card> { it.suit == Suit.HEARTS }.thenByDescending { it.rank.order },
+                ).first()
+            }
+            return cards.maxBy { it.rank.order }
+        }
+        val led = state.trick.first().card.suit
+        val best = state.trick.maxOf { trickStrength(it.card, led) }
+        val remaining = state.options.playerCount - state.trick.size - 1
+        val winning = cards.filter { trickStrength(it, led) > best }
+        if (winning.isNotEmpty()) {
+            if (remaining == 0) return winning.minBy { it.rank.order }
+            return winning.filter { mem.unseenHigher(it) == 0 }.minByOrNull { it.rank.order }
+                ?: winning.maxBy { it.rank.order }
+        }
+        // Cannot win this one: lose it with a card that is not a point.
+        return cards.minBy { pointsOf(it) * 100 + it.rank.order }
+    }
+
+    /** Another seat that has taken every penalty card so far and looks set to run the table. */
+    private fun moonThreat(state: HeartsState, seat: Int): Int? {
+        if (!state.options.allowShootTheMoon) return null
+        val all = pointsIn(state.taken.flatten())
+        if (all < 14 && !(state.trickNumber >= 6 && all >= 8)) return null
+        return (0 until state.options.playerCount).firstOrNull {
+            it != seat && pointsIn(state.taken[it]) == all
+        }
+    }
+
+    private fun scoreLeader(state: HeartsState, seat: Int): Int? =
+        (0 until state.options.playerCount).filter { it != seat }.minByOrNull { state.scores[it] }
+
+    private fun hardLead(cards: List<Card>, mem: Memory, profile: AiProfile): Card {
+        // A cautious seat weighs a trick it might win more heavily than a bold one.
+        val weight = 1.0 - 0.25 * profile.style.boldness
+
+        fun risk(card: Card): Double {
+            val others = mem.unseenIn(card.suit)
+            val higher = mem.unseenHigher(card)
+            val winChance = if (others == 0) 1.0 else Math.pow(1.0 - higher.toDouble() / others, 3.0)
+            var points = 0.5
+            if (card.suit == Suit.HEARTS) points += 2.0
+            if (card.suit == Suit.SPADES && mem.queenUnseen) points += 13 * 0.3
+            // A suit nobody else holds lets everyone dump on whoever wins it.
+            if (others == 0) points += 3.0
+            val shedding = if (winChance < 0.15) 0.05 * card.rank.order else 0.0
+            return winChance * points * weight - shedding
+        }
+        return cards.minBy { risk(it) }
+    }
+
+    private fun hardFollow(
+        state: HeartsState,
+        seat: Int,
+        cards: List<Card>,
+        mem: Memory,
+        profile: AiProfile,
+    ): Card {
+        val trick = state.trick
+        val led = trick.first().card.suit
+        val winner = trick.maxBy { trickStrength(it.card, led) }
+        val best = trickStrength(winner.card, led)
+        val pot = pointsIn(trick.map { it.card })
+        val remaining = state.options.playerCount - trick.size - 1
+        val winning = cards.filter { trickStrength(it, led) > best }
+        val ducking = cards.filter { trickStrength(it, led) < best }
+        val threat = moonThreat(state, seat)
+
+        // Taking even a few points is far better than letting a moon through.
+        if (threat != null && pot > 0 && winning.isNotEmpty()) {
+            return winning.minBy { it.rank.order }
+        }
+
+        val void = cards.none { it.suit == led }
+        if (void) return hardDischarge(state, seat, cards, mem, profile, winner.seat, threat)
+
+        if (ducking.isNotEmpty()) {
+            return ducking.maxBy { it.rank.order + pointsOf(it) * 2 }
+        }
+        // Forced to win. Last to play, the trick is taken whatever is played,
+        // so spend the highest card; otherwise win as cheaply as possible.
+        return if (remaining == 0) {
+            cards.maxBy { it.rank.order }
+        } else {
+            cards.minBy { it.rank.order }
+        }
+    }
+
+    /** Void in the led suit: throw away whatever is still dangerous to hold. */
+    private fun hardDischarge(
+        state: HeartsState,
+        seat: Int,
+        cards: List<Card>,
+        mem: Memory,
+        profile: AiProfile,
+        winnerSeat: Int,
+        threat: Int?,
+    ): Card {
+        val spoil = profile.style.spite >= 0.5 && winnerSeat == scoreLeader(state, seat)
+
+        fun score(card: Card): Double {
+            var s: Double = when {
+                card == QUEEN_OF_SPADES -> 1000.0
+                else -> card.rank.order.toDouble()
+            }
+            // The ace and king of spades only matter while the queen is out there.
+            if (card.suit == Suit.SPADES && card.rank.order > Rank.QUEEN.order && mem.queenUnseen) s += 200
+            if (card.suit == Suit.HEARTS) s += 100
+            if (pointsOf(card) == 0 && mem.hand.count { it.suit == card.suit } <= 2) s += 10
+            if (spoil && pointsOf(card) > 0) s += 300
+            // Do not feed a seat that is running for the moon.
+            if (threat != null && winnerSeat == threat && pointsOf(card) > 0) s -= 2000
+            return s
+        }
+        return cards.maxBy { score(it) }
+    }
+
+    private companion object {
+        val FULL_DECK: List<Card> = Suit.entries.flatMap { suit -> Rank.standard.map { Card(it, suit) } }
     }
 }

app/src/main/java/org/prolibertate/games/game/kaiser/KaiserAi.kt

@@ -3,6 +3,10 @@ package org.prolibertate.games.game.kaiser
 import org.prolibertate.games.game.cards.Card
 import org.prolibertate.games.game.cards.Rank
 import org.prolibertate.games.game.cards.Suit
+import org.prolibertate.games.game.cards.Decks
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
@@ -12,25 +16,51 @@ import org.prolibertate.games.game.engine.GameAi
  * in play chases the five of hearts while trying not to be the one holding the
  * trick the three of spades lands in.
  */
-class KaiserAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
+class KaiserAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) :
     GameAi<KaiserState, KaiserMove> {
 
-    enum class Difficulty(
+    private companion object {
         /** Hand strength needed before bidding above the floor. */
-        val bidThreshold: Int,
-    ) {
-        EASY(bidThreshold = 12),
-        NORMAL(bidThreshold = 16),
-        CAUTIOUS(bidThreshold = 19),
+        const val BID_THRESHOLD = 16.0
+        const val EASY_BID_THRESHOLD = 15.0
+    }
+
+    /** Cards this AI has watched played in the current hand, per seat it plays. */
+    private val memory = HashMap<Int, SeenCards>()
+
+    private class SeenCards {
+        var key: Long = Long.MIN_VALUE
+        val cards = HashSet<Card>()
+    }
+
+    /** Public information only: cards that have appeared on the table this hand. */
+    private fun observe(state: KaiserState, seat: Int): Set<Card> = synchronized(memory) {
+        val seen = memory.getOrPut(seat) { SeenCards() }
+        val key = state.seed * 1_000_003L + state.handNumber
+        if (key != seen.key || state.phase == KaiserPhase.BIDDING) {
+            seen.key = key
+            seen.cards.clear()
+        }
+        state.trick.forEach { seen.cards.add(it.card) }
+        state.completedTrick.forEach { seen.cards.add(it.card) }
+        seen.cards.removeAll(state.hands[seat].toSet())
+        seen.cards.toSet()
     }
 
     override fun chooseMove(state: KaiserState, seat: Int, legal: List<KaiserMove>): KaiserMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
-        return when (state.phase) {
-            KaiserPhase.BIDDING -> chooseBid(state, seat, legal)
-            KaiserPhase.PLAYING -> choosePlay(state, seat, legal)
-            else -> legal.first()
+        val profile = profiles.of(seat)
+        val seen = if (profile.level == AiLevel.HARD) observe(state, seat) else emptySet()
+        val move = try {
+            when (state.phase) {
+                KaiserPhase.BIDDING -> chooseBid(state, seat, legal, profile)
+                KaiserPhase.PLAYING -> choosePlay(state, seat, legal, profile, seen)
+                else -> legal.first()
+            }
+        } catch (e: RuntimeException) {
+            legal.first()
         }
+        return if (move in legal) move else legal.first()
     }
 
     /** Trump length and top cards, plus a nod to holding the counters. */
@@ -50,41 +80,137 @@ class KaiserAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
         return value
     }
 
-    private fun chooseBid(state: KaiserState, seat: Int, legal: List<KaiserMove>): KaiserMove {
+    /**
+     * A finer rating on the same scale as [suitValue]: it rewards a run of top
+     * trump, short side suits to ruff with, and only counts the five of hearts
+     * when the hand can protect it.
+     */
+    private fun hardValue(hand: List<Card>, trump: Suit): Int {
+        val trumps = hand.filter { it.suit == trump }
+        var value = trumps.size * 3
+        for (card in trumps) {
+            value += when (card.rank) {
+                Rank.ACE -> 4
+                Rank.KING -> 3
+                Rank.QUEEN -> 1
+                else -> 0
+            }
+        }
+        // A king backed by the ace, or a queen backed by both, is a real trick.
+        val ranks = trumps.map { it.rank.order }.toSet()
+        if (Rank.ACE.order in ranks && Rank.KING.order in ranks) value += 2
+        if (Rank.ACE.order in ranks && Rank.KING.order in ranks && Rank.QUEEN.order in ranks) value += 1
+
+        for (card in hand) {
+            if (card.suit == trump) continue
+            if (card.rank == Rank.ACE) value += 2
+            else if (card.rank == Rank.KING && hand.any { it.suit == card.suit && it.rank == Rank.ACE }) value += 1
+        }
+        if (trumps.size >= 3) {
+            value += Suit.entries.filter { it != trump }.count { suit -> hand.none { it.suit == suit } } * 2
+        }
+        if (hand.contains(FIVE_OF_HEARTS)) {
+            val guarded = trump == Suit.HEARTS ||
+                hand.any { it.suit == Suit.HEARTS && it.rank == Rank.ACE }
+            value += if (guarded) 3 else -1
+        }
+        if (hand.contains(THREE_OF_SPADES)) value -= if (trump == Suit.SPADES) 1 else 2
+        return value
+    }
+
+    private fun chooseBid(
+        state: KaiserState,
+        seat: Int,
+        legal: List<KaiserMove>,
+        profile: AiProfile,
+    ): KaiserMove {
+        val hand = state.hands[seat]
+        val hard = profile.level == AiLevel.HARD
+        val easy = profile.level == AiLevel.EASY
+
         val naming = legal.filterIsInstance<NameTrump>()
         if (naming.isNotEmpty()) {
-            val hand = state.hands[seat]
-            val best = naming
-                .filter { it.suit != null }
-                .maxByOrNull { suitValue(hand, it.suit!!) }
-            return best ?: naming.first()
+            val named = naming.filter { it.suit != null }
+            return when {
+                hard -> named.maxByOrNull { hardValue(hand, it.suit!!) }
+                easy -> {
+                    val noisy = named.associateWith { suitValue(hand, it.suit!!) + profile.random.nextInt(-4, 5) }
+                    named.sortedByDescending { noisy.getValue(it) }.let { if (it.isEmpty()) null else profile.pick(it) }
+                }
+                else -> named.maxByOrNull { suitValue(hand, it.suit!!) }
+            } ?: naming.first()
         }
 
-        val hand = state.hands[seat]
-        val bestSuit = Suit.entries.maxByOrNull { suitValue(hand, it) } ?: Suit.SPADES
-        val strength = suitValue(hand, bestSuit)
+        val bestSuit: Suit
+        val strength: Int
+        if (hard) {
+            bestSuit = Suit.entries.maxBy { hardValue(hand, it) }
+            strength = hardValue(hand, bestSuit)
+        } else {
+            bestSuit = Suit.entries.maxByOrNull { suitValue(hand, it) } ?: Suit.SPADES
+            strength = suitValue(hand, bestSuit) + if (easy) profile.random.nextInt(-5, 6) else 0
+        }
 
         val bids = legal.filterIsInstance<MakeBid>().filter { !it.bid.noTrump }
-        if (strength >= difficulty.bidThreshold && bids.isNotEmpty()) {
+        var needed = profile.bolder(if (easy) EASY_BID_THRESHOLD else BID_THRESHOLD, -4.0)
+        // A spoiler steals contracts from the other side.
+        val bidder = state.highBidder
+        if (bidder != null && teamOf(bidder) != teamOf(seat) && profile.style.spite >= 0.5) needed -= 2.0
+        if (hard && bids.isNotEmpty()) {
+            val cheapest = bids.minBy { it.bid.points }
+            // A bigger contract needs a bigger hand, and a partner who is
+            // already bidding is better supported than outbid.
+            needed += 2.0 * (cheapest.bid.points - state.options.minimumBid)
+            if (bidder != null && bidder == partnerOf(seat)) needed += 4.0
+        }
+        if (strength >= needed && bids.isNotEmpty()) {
             // Bid the floor rather than reaching; the contract is what matters.
             return bids.minBy { it.bid.points }
         }
         return legal.firstOrNull { it is PassBid } ?: bids.minByOrNull { it.bid.points } ?: legal.first()
     }
 
-    private fun choosePlay(state: KaiserState, seat: Int, legal: List<KaiserMove>): KaiserMove {
+    private fun choosePlay(
+        state: KaiserState,
+        seat: Int,
+        legal: List<KaiserMove>,
+        profile: AiProfile,
+        seen: Set<Card>,
+    ): KaiserMove {
         val cards = legal.filterIsInstance<PlayCard>().map { it.card }
         if (cards.size == 1) return PlayCard(cards.first())
 
+        val bidder = state.highBidder
+        val defending = bidder != null && teamOf(bidder) != teamOf(seat)
+        val spiteful = defending && profile.style.spite >= 0.5
+        val best = if (profile.level == AiLevel.HARD) {
+            hardPlay(state, seat, cards, seen, bidder != null && !defending)
+        } else {
+            standardPlay(state, seat, cards, spiteful, profile)
+        }
+        if (profile.level != AiLevel.EASY) return PlayCard(best)
+
+        // Easy sometimes throws a high card, or a counter, at the wrong time.
+        val ranked = listOf(best) + (cards - best).sortedByDescending { it.rank.order }
+        return PlayCard(profile.pick(ranked))
+    }
+
+    private fun standardPlay(
+        state: KaiserState,
+        seat: Int,
+        cards: List<Card>,
+        spiteful: Boolean,
+        profile: AiProfile,
+    ): Card {
         if (state.trick.isEmpty()) {
             // Leading: pull with the top trump, else lead an ace.
             val trump = state.trump
             val topTrump = cards.filter { trump != null && it.suit == trump }
                 .maxByOrNull { it.rank.order }
-            if (topTrump != null && topTrump.rank.order >= Rank.KING.order) return PlayCard(topTrump)
+            if (topTrump != null && topTrump.rank.order >= Rank.KING.order) return topTrump
             val ace = cards.filter { it.rank == Rank.ACE }.maxByOrNull { it.rank.order }
-            if (ace != null) return PlayCard(ace)
-            return PlayCard(cards.minBy { it.rank.order })
+            if (ace != null) return ace
+            return cards.minBy { it.rank.order }
         }
 
         val ledSuit = state.trick.first().card.suit
@@ -96,22 +222,129 @@ class KaiserAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
         // The five of hearts is worth more than the trick it rides in on.
         val fiveOnTable = state.trick.any { it.card == FIVE_OF_HEARTS }
         val threeOnTable = state.trick.any { it.card == THREE_OF_SPADES }
+        // A spoiler on the defence will not leave a trick to a partner who may be beaten.
+        val notLast = state.trick.size < 3
+        val letPartner = partnerWinning && !threeOnTable &&
+            !(spiteful && notLast && winners.isNotEmpty()) && !profile.slips()
 
         return when {
             // Partner has it and the trick is worth taking: let them have it,
             // unless the three of spades is in there.
-            partnerWinning && !threeOnTable -> PlayCard(cards.minBy { it.rank.order })
+            letPartner -> cards.minBy { it.rank.order }
 
             fiveOnTable && winners.isNotEmpty() ->
-                PlayCard(winners.maxBy { trickStrength(it, state.trump, ledSuit) })
+                winners.maxBy { trickStrength(it, state.trump, ledSuit) }
 
             // Nobody wants the three; duck if the trick is already lost.
-            threeOnTable && winners.isEmpty() -> PlayCard(cards.minBy { it.rank.order })
+            threeOnTable && winners.isEmpty() -> cards.minBy { it.rank.order }
 
             winners.isNotEmpty() && !threeOnTable ->
-                PlayCard(winners.minBy { trickStrength(it, state.trump, ledSuit) })
+                winners.minBy { trickStrength(it, state.trump, ledSuit) }
 
-            else -> PlayCard(cards.minBy { it.rank.order })
+            else -> cards.minBy { it.rank.order }
         }
     }
+
+    /**
+     * The strong player's choice: it remembers what has been played, so it
+     * knows which of its cards are sure winners, feeds the five of hearts to a
+     * partner who is safe, and buries the three of spades under the other side.
+     */
+    private fun hardPlay(
+        state: KaiserState,
+        seat: Int,
+        cards: List<Card>,
+        seen: Set<Card>,
+        bidding: Boolean,
+    ): Card {
+        val trump = state.trump
+        val hand = state.hands[seat]
+        val unseen = Decks.kaiser().filter { it !in seen && it !in hand }
+        val counters = setOf(FIVE_OF_HEARTS, THREE_OF_SPADES)
+
+        fun isBoss(card: Card): Boolean {
+            val mine = trickStrength(card, trump, card.suit)
+            val cardIsTrump = trump != null && card.suit == trump
+            return unseen.none { other ->
+                (cardIsTrump || trump == null || other.suit != trump) &&
+                    trickStrength(other, trump, card.suit) > mine
+            }
+        }
+
+        if (state.trick.isEmpty()) {
+            val plain = cards.filter { it !in counters }.ifEmpty { return cards.minBy { it.rank.order } }
+            val myTrump = plain.filter { trump != null && it.suit == trump }
+            val unseenTrump = unseen.count { trump != null && it.suit == trump }
+            val bossTrump = myTrump.filter { isBoss(it) }.maxByOrNull { it.rank.order }
+            if (bidding && unseenTrump > 0 && bossTrump != null) return bossTrump
+            plain.filter { it !in myTrump && isBoss(it) }.maxByOrNull { it.rank.order }?.let { return it }
+            if (bidding && unseenTrump > 0 && myTrump.size >= 2) return myTrump.maxBy { it.rank.order }
+            if (bossTrump != null) return bossTrump
+            val counts = hand.groupingBy { it.suit }.eachCount()
+            val offSuit = plain.filter { it !in myTrump }.ifEmpty { plain }
+            // The bidders draw a suit out; the defenders shorten one to ruff.
+            return if (bidding) {
+                offSuit.minWith(compareBy<Card>({ -(counts[it.suit] ?: 0) }, { it.rank.order }))
+            } else {
+                offSuit.minWith(compareBy<Card>({ counts[it.suit] ?: 0 }, { it.rank.order }))
+            }
+        }
+
+        val ledSuit = state.trick.first().card.suit
+        val best = state.trick.maxBy { trickStrength(it.card, trump, ledSuit) }
+        val partnerWinning = partnerOf(best.seat) == seat
+        val remaining = 4 - state.trick.size - 1
+        val bestStrength = trickStrength(best.card, trump, ledSuit)
+        val winners = cards.filter { trickStrength(it, trump, ledSuit) > bestStrength }
+        val fiveOnTable = state.trick.any { it.card == FIVE_OF_HEARTS }
+        val threeOnTable = state.trick.any { it.card == THREE_OF_SPADES }
+
+        fun sure(card: Card): Boolean =
+            remaining == 0 ||
+                unseen.none { trickStrength(it, trump, ledSuit) > trickStrength(card, trump, ledSuit) }
+
+        fun throwAway(): Card {
+            val plain = cards.filter { it !in counters }
+            if (plain.isNotEmpty()) return plain.minBy { it.rank.order }
+            // Only counters left: the five belongs with partner, the three does not.
+            return if (partnerWinning) cards.maxBy { it.rank.order } else cards.minBy { it.rank.order }
+        }
+
+        if (partnerWinning) {
+            if (sure(best.card)) {
+                // Partner has it: feed them the five of hearts if we can.
+                if (FIVE_OF_HEARTS in cards && trickStrength(FIVE_OF_HEARTS, trump, ledSuit) <= bestStrength) {
+                    return FIVE_OF_HEARTS
+                }
+                return throwAway()
+            }
+            if (!threeOnTable) {
+                winners.filter { sure(it) }.minByOrNull { trickStrength(it, trump, ledSuit) }?.let { return it }
+            }
+            return throwAway()
+        }
+
+        // The other side is winning.
+        if (threeOnTable) return throwAway()
+        if (fiveOnTable && winners.isNotEmpty()) {
+            val safe = winners.filter { sure(it) }
+            return if (safe.isNotEmpty()) {
+                safe.minBy { trickStrength(it, trump, ledSuit) }
+            } else {
+                winners.maxBy { trickStrength(it, trump, ledSuit) }
+            }
+        }
+        // Bury the three under a trick the other side is certain to take.
+        if (THREE_OF_SPADES in cards &&
+            trickStrength(THREE_OF_SPADES, trump, ledSuit) <= bestStrength &&
+            sure(best.card)
+        ) {
+            return THREE_OF_SPADES
+        }
+        if (winners.isNotEmpty()) {
+            val safe = winners.filter { sure(it) }
+            return (safe.ifEmpty { winners }).minBy { trickStrength(it, trump, ledSuit) }
+        }
+        return throwAway()
+    }
 }

app/src/main/java/org/prolibertate/games/game/mastermind/MastermindAi.kt

@@ -1,6 +1,7 @@
 package org.prolibertate.games.game.mastermind
 
 import kotlin.random.Random
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
@@ -16,9 +17,20 @@ import org.prolibertate.games.game.engine.GameAi
  * - **Strong** is Knuth's idea one ply deep: among the codes still standing, it
  *   plays the one whose *worst* possible answer leaves the smallest number of
  *   candidates behind, so it cannot be given a reply that tells it little.
+ *
+ * The table already carries its own three-step level ([MastermindOptions.level]),
+ * chosen at setup and shown to the player, and it means exactly what an
+ * [org.prolibertate.games.game.engine.AiLevel] would here. Layering a second
+ * knob on top would make the two disagree, so the shared [profiles] are
+ * accepted for a uniform constructor but this game keeps playing at the level
+ * the table names. [level] overrides that, which is what the tests use.
  */
-class MastermindAi(private val level: MastermindLevel? = null) :
-    GameAi<MastermindState, MastermindMove> {
+class MastermindAi(
+    @Suppress("unused") private val profiles: AiProfiles = AiProfiles.DEFAULT,
+    private val level: MastermindLevel? = null,
+) : GameAi<MastermindState, MastermindMove> {
+
+    constructor(level: MastermindLevel?) : this(AiProfiles.DEFAULT, level)
 
     override fun chooseMove(
         state: MastermindState,

app/src/main/java/org/prolibertate/games/game/morris/MorrisAi.kt

@@ -1,6 +1,9 @@
 package org.prolibertate.games.game.morris
 
 import kotlin.random.Random
+import kotlin.math.roundToInt
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
@@ -19,7 +22,20 @@ import org.prolibertate.games.game.engine.GameAi
  *
  * Strength comes from the table's own options unless a level is passed in.
  */
-class MorrisAi(private val level: MorrisLevel? = null) : GameAi<MorrisState, MorrisMove> {
+class MorrisAi(
+    private val profiles: AiProfiles = AiProfiles.DEFAULT,
+    private val level: MorrisLevel? = null,
+) : GameAi<MorrisState, MorrisMove> {
+
+    // Viewpoint and weights for the move being chosen. At the default (balanced)
+    // style they equal the plain constants, so Medium is unchanged.
+    private var rootSeat = 0
+    private var wMill = MILL
+    private var wAlmost = ALMOST
+    private var wTheirAlmost = ALMOST
+    private var wMobility = MOBILITY
+    private var wTheirMobility = MOBILITY
+    private var wDanger = DANGER
 
     /**
      * Ceiling on how much thinking any one move gets, so a phone does not stall
@@ -32,10 +48,29 @@ class MorrisAi(private val level: MorrisLevel? = null) : GameAi<MorrisState, Mor
 
     override fun chooseMove(state: MorrisState, seat: Int, legal: List<MorrisMove>): MorrisMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
+        val profile = profiles.of(seat)
         val setting = level ?: state.options.level
         if (legal.size == 1) return legal.first()
 
-        nodeBudget = NODE_BUDGET
+        val bold = profile.style.boldness
+        val spite = profile.style.spite
+        rootSeat = seat
+        // Bold seats value mills and half-built mills more; cautious ones value
+        // mobility and avoiding the three-piece trap. Spiteful seats weigh the
+        // opponent's mill threats and freedom of movement more.
+        wMill = MILL + (bold * 6).roundToInt()
+        wAlmost = ALMOST + (bold * 3).roundToInt()
+        wTheirAlmost = ALMOST + (spite * 4).roundToInt() + (-bold * 2).roundToInt()
+        wMobility = MOBILITY + (-bold * 1).roundToInt()
+        wTheirMobility = MOBILITY + (spite * 2).roundToInt() + (-bold * 1).roundToInt()
+        wDanger = DANGER + (-bold * 10).roundToInt()
+
+        val maxDepth = when (profile.level) {
+            AiLevel.EASY -> maxOf(2, setting.depth - 2)
+            AiLevel.MEDIUM -> setting.depth
+            AiLevel.HARD -> setting.depth + 1
+        }
+        nodeBudget = if (profile.level == AiLevel.HARD) HARD_NODE_BUDGET else NODE_BUDGET
         nodesUsed = 0
 
         // Deterministic for a given position, so a host and a client replaying
@@ -46,7 +81,7 @@ class MorrisAi(private val level: MorrisLevel? = null) : GameAi<MorrisState, Mor
         var best = legal.first()
         var scored = legal.map { it to 0 }
 
-        for (depth in 1..setting.depth) {
+        for (depth in 1..maxDepth) {
             val results = mutableListOf<Pair<MorrisMove, Int>>()
             val ordered = order(legal, preferred = best)
             for (move in ordered) {
@@ -73,7 +108,16 @@ class MorrisAi(private val level: MorrisLevel? = null) : GameAi<MorrisState, Mor
 
         val bestScore = scored.maxOf { it.second }
         val acceptable = scored.filter { it.second >= bestScore - setting.slack() }
-        return acceptable[random.nextInt(acceptable.size)].first
+        val chosen = acceptable[random.nextInt(acceptable.size)].first
+        if (profile.level != AiLevel.EASY) return chosen
+
+        // Easy now and then settles for another sound-looking move: only ones
+        // within about a piece of the best, never a real blunder.
+        val ranked = listOf(chosen) + scored
+            .filter { it.first != chosen && it.second >= bestScore - EASY_TOLERANCE }
+            .sortedByDescending { it.second }
+            .map { it.first }
+        return profile.pick(ranked)
     }
 
     /** How far below best a move may score and still be considered. */
@@ -155,16 +199,17 @@ class MorrisAi(private val level: MorrisLevel? = null) : GameAi<MorrisState, Mor
 
     /** The position from the point of view of the side to move. */
     private fun evaluate(state: MorrisState): Int {
-        val us = state.turn
-        val them = other(us)
+        val me = rootSeat
+        val them = other(me)
 
-        var score = PIECE * (state.remaining(us) - state.remaining(them))
-        score += MILL * (state.millCount(us) - state.millCount(them))
-        score += ALMOST * (almostMills(state, us) - almostMills(state, them))
-        score += MOBILITY * (mobility(state, us) - mobility(state, them))
-        score += LADDER * (midEdgeCount(state, us) - midEdgeCount(state, them))
-        score -= DANGER * (onThree(state, us) - onThree(state, them))
-        return score
+        var score = PIECE * (state.remaining(me) - state.remaining(them))
+        score += wMill * (state.millCount(me) - state.millCount(them))
+        score += wAlmost * almostMills(state, me) - wTheirAlmost * almostMills(state, them)
+        score += wMobility * mobility(state, me) - wTheirMobility * mobility(state, them)
+        score += LADDER * (midEdgeCount(state, me) - midEdgeCount(state, them))
+        score -= wDanger * (onThree(state, me) - onThree(state, them))
+        // Scored from the root seat's side; flip for the side to move.
+        return if (state.turn == me) score else -score
     }
 
     /**
@@ -215,6 +260,8 @@ class MorrisAi(private val level: MorrisLevel? = null) : GameAi<MorrisState, Mor
         const val INFINITY = 1_000_000
         const val WIN = 100_000
         const val NODE_BUDGET = 120_000
+        const val HARD_NODE_BUDGET = 280_000
+        const val EASY_TOLERANCE = 120
 
         const val PIECE = 120
         const val MILL = 26

app/src/main/java/org/prolibertate/games/game/pirates/PiratesAi.kt

@@ -1,6 +1,10 @@
 package org.prolibertate.games.game.pirates
 
+import kotlin.math.roundToInt
 import kotlin.random.Random
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
@@ -14,17 +18,73 @@ import org.prolibertate.games.game.engine.GameAi
  * from the Bulgars' side and negated for the pirates, so the two cannot drift
  * apart and start valuing the same position differently.
  */
-class PiratesAi(private val level: PiratesLevel? = null) : GameAi<PiratesState, PiratesMove> {
+class PiratesAi(
+    private val profiles: AiProfiles = AiProfiles.DEFAULT,
+    private val level: PiratesLevel? = null,
+) : GameAi<PiratesState, PiratesMove> {
 
     private var nodeBudget = 0
     private var nodesUsed = 0
 
+    // Evaluation weights for the seat now choosing; the standard values unless
+    // a style or level bends them. Set at the top of each chooseMove.
+    private var wBulgar = BULGAR
+    private var wStronghold = IN_STRONGHOLD
+    private var wAdvance = ADVANCE
+    private var wExposed = EXPOSED
+    private var wRoom = PIRATE_ROOM
+
+    private fun tune(profile: AiProfile, seat: Int) {
+        val bold = profile.style.boldness
+        val spite = profile.style.spite
+        var bulgar = 1.0
+        var stronghold = 1.0
+        var advance = 1.0
+        var exposed = 1.0
+        var room = 1.0
+        if (seat == BULGAR_SEAT) {
+            // A bold crowd pushes forward and risks being jumped; a cautious one stays tight.
+            exposed -= 0.3 * bold
+            advance += 0.3 * bold
+            // A spoiler squeezes the pirates' room.
+            room += 0.4 * spite
+        } else {
+            // A bold pirate hunts; a cautious one keeps its room to move.
+            bulgar += 0.15 * bold
+            room -= 0.3 * bold
+            // A spoiler pays more attention to stopping the climb.
+            stronghold += 0.4 * spite
+            advance += 0.4 * spite
+        }
+        if (profile.level == AiLevel.EASY) exposed *= 0.5 // overlooks some threats
+        wBulgar = (BULGAR * bulgar).roundToInt()
+        wStronghold = (IN_STRONGHOLD * stronghold).roundToInt()
+        wAdvance = (ADVANCE * advance).roundToInt()
+        wExposed = (EXPOSED * exposed).roundToInt()
+        wRoom = (PIRATE_ROOM * room).roundToInt()
+    }
+
     override fun chooseMove(state: PiratesState, seat: Int, legal: List<PiratesMove>): PiratesMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
-        val setting = level ?: state.options.level
         if (legal.size == 1) return legal.first()
+        val profile = profiles.of(seat)
+        val base = level ?: state.options.level
+        tune(profile, seat)
 
-        nodeBudget = NODE_BUDGET
+        val hard = profile.level == AiLevel.HARD
+        val easy = profile.level == AiLevel.EASY
+        val maxDepth = when (profile.level) {
+            AiLevel.MEDIUM -> base.depth
+            AiLevel.EASY -> minOf(base.depth, 2)
+            AiLevel.HARD -> base.depth + 2
+        }
+        val slack = when (profile.level) {
+            AiLevel.MEDIUM -> base.slack()
+            AiLevel.EASY -> maxOf(base.slack(), 40)
+            AiLevel.HARD -> 0
+        }
+
+        nodeBudget = if (hard) HARD_NODE_BUDGET else NODE_BUDGET
         nodesUsed = 0
 
         val random = Random(state.moveLog.size * 31L + state.positionKey().hashCode())
@@ -32,18 +92,23 @@ class PiratesAi(private val level: PiratesLevel? = null) : GameAi<PiratesState,
         var best = legal.first()
         var scored = legal.map { it to 0 }
 
-        for (depth in 1..setting.depth) {
+        for (depth in 1..maxDepth) {
             val results = mutableListOf<Pair<PiratesMove, Int>>()
             val ordered = order(legal, preferred = best)
+            var bestSoFar = -INFINITY
             for (move in ordered) {
+                // A hard seat only needs the exact value of moves that tie or beat
+                // the best so far, so everything else is searched against a window.
+                val floor = if (hard && results.isNotEmpty()) bestSoFar - 1 else -INFINITY
                 val score = -negamax(
                     state = PiratesRules.advanced(state, move),
                     depth = depth - 1,
                     alpha = -INFINITY,
-                    beta = INFINITY,
+                    beta = -floor,
                     ply = 1,
                 )
                 results += move to score
+                if (score > bestSoFar) bestSoFar = score
                 if (nodesUsed >= nodeBudget) break
             }
             if (nodesUsed >= nodeBudget && results.size < ordered.size) break
@@ -51,9 +116,16 @@ class PiratesAi(private val level: PiratesLevel? = null) : GameAi<PiratesState,
             best = results.maxBy { it.second }.first
         }
 
+        if (easy) {
+            val ranked = scored.sortedByDescending { it.second }.map { it.first }
+            val picked = profile.pick(ranked)
+            if (picked != ranked.first() && picked in legal) return picked
+        }
+
         val bestScore = scored.maxOf { it.second }
-        val acceptable = scored.filter { it.second >= bestScore - setting.slack() }
-        return acceptable[random.nextInt(acceptable.size)].first
+        val acceptable = scored.filter { it.second >= bestScore - slack }
+        val chosen = acceptable[random.nextInt(acceptable.size)].first
+        return if (chosen in legal) chosen else legal.first()
     }
 
     private fun PiratesLevel.slack(): Int = when (this) {
@@ -115,22 +187,22 @@ class PiratesAi(private val level: PiratesLevel? = null) : GameAi<PiratesState,
     private fun evaluate(state: PiratesState): Int {
         var bulgarView = 0
 
-        bulgarView += state.count(BULGAR_SEAT) * BULGAR
-        bulgarView += state.stronghold() * IN_STRONGHOLD
+        bulgarView += state.count(BULGAR_SEAT) * wBulgar
+        bulgarView += state.stronghold() * wStronghold
 
         // How far up the board the crowd has come. The stronghold is at the
         // top, so a low row number is ground gained.
         for (point in 0 until POINTS) {
             if (state.board[point] != BULGAR_SEAT) continue
-            bulgarView += (GRID - 1 - rowOf(point)) * ADVANCE
+            bulgarView += (GRID - 1 - rowOf(point)) * wAdvance
             // A Bulgar with an empty point directly behind it on a line is one
             // a pirate can jump. Being unjumpable is most of staying alive.
-            if (isExposed(state, point)) bulgarView -= EXPOSED
+            if (isExposed(state, point)) bulgarView -= wExposed
         }
 
         // Room for the pirates to work in, which is what the Bulgars are
         // slowly taking away.
-        bulgarView -= pirateMobility(state) * PIRATE_ROOM
+        bulgarView -= pirateMobility(state) * wRoom
 
         return if (state.turn == BULGAR_SEAT) bulgarView else -bulgarView
     }
@@ -160,6 +232,7 @@ class PiratesAi(private val level: PiratesLevel? = null) : GameAi<PiratesState,
         const val INFINITY = 1_000_000
         const val WIN = 100_000
         const val NODE_BUDGET = 150_000
+        const val HARD_NODE_BUDGET = 400_000
 
         const val BULGAR = 100
         const val IN_STRONGHOLD = 55

app/src/main/java/org/prolibertate/games/game/president/PresidentAi.kt

@@ -1,63 +1,85 @@
 package org.prolibertate.games.game.president
 
+import org.prolibertate.games.game.cards.Card
+import org.prolibertate.games.game.cards.Rank
+import org.prolibertate.games.game.cards.Suit
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
- * A heuristic President opponent.
+ * A President opponent whose strength and temperament come from the seat's profile.
  *
- * It sheds low cards early, spends the cheapest set that takes the pile, and
- * holds twos back for when they are worth something. It does not count cards.
+ * MEDIUM (the default) sheds low cards early, spends the cheapest set that takes
+ * the pile, and holds twos back for when they are worth something. It does not
+ * count cards.
+ *
+ * EASY spends high cards freely when leading and sometimes plays a worse set or
+ * passes when it should have played. HARD reads the round's log to count which
+ * ranks are still out, leads so that its remaining hand is unbeatable, blocks an
+ * opponent who is about to go out, and keeps its high cards back when a cheap
+ * pile is not worth them.
  */
-class PresidentAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
+class PresidentAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) :
     GameAi<PresidentState, PresidentMove> {
 
-    enum class Difficulty(
-        /** Rank above which a card is hoarded rather than spent to lead. */
-        val hoardAbove: Int,
-    ) {
-        /** Plays whatever is cheapest, whenever it can. */
-        EASY(hoardAbove = 99),
-
-        NORMAL(hoardAbove = 13),
-
-        /** Sits on high cards until late, then dumps them. */
-        CAUTIOUS(hoardAbove = 11),
-    }
-
     override fun chooseMove(
         state: PresidentState,
         seat: Int,
         legal: List<PresidentMove>,
     ): PresidentMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
+        val profile = profiles.of(seat)
         val plays = legal.filterIsInstance<PlayCards>()
         if (plays.isEmpty()) return legal.first()
 
-        return if (state.setSize == 0) chooseLead(state, seat, plays) else chooseFollow(state, plays, legal)
+        val move: PresidentMove = when (profile.level) {
+            AiLevel.MEDIUM -> standard(state, seat, plays, legal, hoardAbove = HOARD_ABOVE)
+            AiLevel.EASY -> easy(state, seat, plays, legal, profile)
+            AiLevel.HARD -> hard(state, seat, plays, legal, profile)
+        }
+        return if (move in legal) move else legal.first()
     }
 
+    private fun standard(
+        state: PresidentState,
+        seat: Int,
+        plays: List<PlayCards>,
+        legal: List<PresidentMove>,
+        hoardAbove: Int,
+    ): PresidentMove =
+        if (state.setSize == 0) {
+            leadRanking(state, seat, plays, hoardAbove).first()
+        } else {
+            chooseFollow(state, plays, legal)
+        }
+
+    // -----------------------------------------------------------------------
+    // Standard plan
+    // -----------------------------------------------------------------------
+
     /**
      * Leading is a chance to shed. Prefer the biggest set of the lowest rank,
      * which clears junk fastest without spending anything valuable.
      */
-    private fun chooseLead(
+    private fun leadRanking(
         state: PresidentState,
         seat: Int,
         plays: List<PlayCards>,
-    ): PresidentMove {
+        hoardAbove: Int,
+    ): List<PlayCards> {
         val endgame = state.hands[seat].size <= 4
         val candidates = plays.filter { play ->
-            endgame || presidentRank(play.cards.first()) <= difficulty.hoardAbove
+            endgame || presidentRank(play.cards.first()) <= hoardAbove
         }.ifEmpty { plays }
 
-        return candidates
-            .sortedWith(
-                compareBy(
-                    { presidentRank(it.cards.first()) },
-                    { -it.cards.size },
-                )
-            )
-            .first()
+        return candidates.sortedWith(
+            compareBy<PlayCards>(
+                { presidentRank(it.cards.first()) },
+                { -it.cards.size },
+            ),
+        )
     }
 
     /**
@@ -82,4 +104,188 @@ class PresidentAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
 
         return cheapest
     }
+
+    // -----------------------------------------------------------------------
+    // Easy
+    // -----------------------------------------------------------------------
+
+    private fun easy(
+        state: PresidentState,
+        seat: Int,
+        plays: List<PlayCards>,
+        legal: List<PresidentMove>,
+        profile: AiProfile,
+    ): PresidentMove {
+        if (state.setSize == 0) {
+            // Never hoards: leads aces and twos as readily as anything else.
+            return profile.pick(leadRanking(state, seat, plays, hoardAbove = 99))
+        }
+        val best = chooseFollow(state, plays, legal)
+        val alternatives = plays.sortedBy { presidentRank(it.cards.first()) }
+        val ranked = buildList<PresidentMove> {
+            add(best)
+            addAll(alternatives.filter { it != best })
+            if (PassTurn in legal && best != PassTurn) add(PassTurn)
+        }
+        return profile.pick(ranked)
+    }
+
+    // -----------------------------------------------------------------------
+    // Hard
+    // -----------------------------------------------------------------------
+
+    /** Ranks still unaccounted for: not in this seat's hand and not played yet this round. */
+    private class Counts(state: PresidentState, seat: Int) {
+        private val unseen = HashMap<Int, Int>()
+
+        init {
+            for (rank in 3..15) unseen[rank] = 4
+            state.hands[seat].forEach { dec(presidentRank(it)) }
+            val log = state.log
+            val start = log.indexOfLast { ROUND_MARK.matches(it) } + 1
+            for (i in start until log.size) {
+                val line = log[i]
+                val at = line.indexOf(" plays ")
+                if (!line.startsWith("Seat ") || at < 0) continue
+                line.substring(at + " plays ".length).removeSuffix(".").split(' ').forEach { token ->
+                    LABELS[token]?.let { dec(presidentRank(it)) }
+                }
+            }
+        }
+
+        private fun dec(rank: Int) {
+            unseen[rank] = maxOf(0, (unseen[rank] ?: 0) - 1)
+        }
+
+        fun of(rank: Int): Int = unseen[rank] ?: 0
+
+        /** Could an opponent still hold a set of [size] that beats a set of [size] at [rank]? */
+        fun beatable(rank: Int, size: Int, options: PresidentOptions): Boolean {
+            if (options.twosClear && rank != 15 && of(15) >= size) return true
+            if (options.fourOfAKindBomb && size == 4 && (3..15).any { it != rank && of(it) == 4 }) return true
+            return (rank + 1..15).any { of(it) >= size }
+        }
+    }
+
+    private fun hard(
+        state: PresidentState,
+        seat: Int,
+        plays: List<PlayCards>,
+        legal: List<PresidentMove>,
+        profile: AiProfile,
+    ): PresidentMove {
+        val hand = state.hands[seat]
+        // Going out beats any plan.
+        plays.firstOrNull { it.cards.size == hand.size }?.let { return it }
+
+        val counts = Counts(state, seat)
+        val others = (0 until state.options.playerCount).filter { it != seat && !state.isOut(it) }
+        return if (state.setSize == 0) {
+            hardLead(state, seat, plays, counts, others, profile)
+        } else {
+            hardFollow(state, seat, plays, legal, counts, others, profile)
+        }
+    }
+
+    /** An opponent who could empty their hand with a set the size of the one on top, or is the score leader close to out. */
+    private fun outThreat(
+        state: PresidentState,
+        seat: Int,
+        others: List<Int>,
+        profile: AiProfile,
+    ): Boolean {
+        if (others.any { state.handCounts[it] <= maxOf(1, state.setSize) }) return true
+        if (profile.style.spite >= 0.5) {
+            val leader = others.maxByOrNull { state.scores[it] }
+            if (leader != null && state.handCounts[leader] <= 2) return true
+        }
+        return false
+    }
+
+    private fun hardLead(
+        state: PresidentState,
+        seat: Int,
+        plays: List<PlayCards>,
+        counts: Counts,
+        others: List<Int>,
+        profile: AiProfile,
+    ): PresidentMove {
+        val hand = state.hands[seat]
+        val options = state.options
+        val oneCardLeft = others.any { state.handCounts[it] == 1 }
+
+        fun unbeatableAfter(play: PlayCards): Boolean {
+            val remaining = hand.toMutableList().also { left -> play.cards.forEach { left.remove(it) } }
+            return remaining.groupBy { presidentRank(it) }
+                .all { (rank, cards) -> !counts.beatable(rank, cards.size, options) }
+        }
+
+        fun key(play: PlayCards): Double {
+            val rank = presidentRank(play.cards.first())
+            val size = play.cards.size
+            var k = rank - 0.9 * size
+            if (unbeatableAfter(play)) return k - 50.0
+            // Leading aces and twos spends control that is worth more later.
+            if (rank >= 14 && hand.size > 4) k += 20.0
+            // Somebody is down to one card: a single would let them out if they can beat it.
+            if (oneCardLeft && size == 1) k += 10.0 + (15 - rank) * 2.0
+            // A bold seat is happier to lead its strong sets early.
+            if (rank >= 11) k -= profile.style.boldness * 2.0
+            return k
+        }
+        return plays.minBy { key(it) }
+    }
+
+    private fun hardFollow(
+        state: PresidentState,
+        seat: Int,
+        plays: List<PlayCards>,
+        legal: List<PresidentMove>,
+        counts: Counts,
+        others: List<Int>,
+        profile: AiProfile,
+    ): PresidentMove {
+        val hand = state.hands[seat]
+        val options = state.options
+        val canPass = PassTurn in legal
+        val threat = outThreat(state, seat, others, profile)
+
+        val withoutTwos = plays.filter { play -> play.cards.none { isTwo(it) } }
+        val pool = withoutTwos.ifEmpty { plays }
+        val cheapest = pool.minBy { presidentRank(it.cards.first()) }
+
+        if (threat) {
+            // Block them: take the pile with the cheapest set nobody can beat,
+            // or failing that the cheapest set there is, twos included.
+            val sure = plays.filter { !counts.beatable(presidentRank(it.cards.first()), it.cards.size, options) }
+                .minByOrNull { presidentRank(it.cards.first()) }
+            return sure ?: plays.minBy { presidentRank(it.cards.first()) }
+        }
+
+        val spendingTwo = cheapest.cards.any { isTwo(it) }
+        if (spendingTwo) {
+            val remaining = hand.toMutableList().also { left -> cheapest.cards.forEach { left.remove(it) } }
+            val runOut = remaining.groupBy { presidentRank(it) }
+                .all { (rank, cards) -> !counts.beatable(rank, cards.size, options) }
+            val worthIt = state.setRank >= 13 || state.pile.size >= 4 || hand.size <= 4 || runOut
+            if (!worthIt && canPass) return PassTurn
+            return cheapest
+        }
+
+        // A cheap pile is not worth a high card while the hand is still big.
+        val gap = profile.bolder(6.0, 2.0)
+        val cost = presidentRank(cheapest.cards.first()) - state.setRank
+        if (canPass && hand.size > 5 && cost > gap && presidentRank(cheapest.cards.first()) >= 12) {
+            return PassTurn
+        }
+        return cheapest
+    }
+
+    private companion object {
+        const val HOARD_ABOVE = 13
+        val ROUND_MARK = Regex("^Round \\d+\\.$")
+        val LABELS: Map<String, Card> = Suit.entries.flatMap { suit ->
+            Rank.standard.map { Card(it, suit) }
+        }.associateBy { it.label }
+    }
 }

app/src/main/java/org/prolibertate/games/game/sequence/SequenceAi.kt

@@ -1,5 +1,8 @@
 package org.prolibertate.games.game.sequence
 
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
@@ -8,18 +11,39 @@ import org.prolibertate.games.game.engine.GameAi
  * It scores every legal placement by how much it advances one of its own runs
  * and how badly it spoils an opponent's, holds its jacks back until they are
  * worth spending, and never looks at anyone's hand.
+ *
+ * Per seat, from [profiles]: Easy mostly ignores your runs, burns its one-eyed
+ * jacks on anything and sometimes settles for a lesser square; Hard blocks hard,
+ * saves jacks for real threats, always takes a completed run and always stops
+ * one that is one square from done. Boldness leans towards building over
+ * blocking and spite the other way round.
  */
-class SequenceAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
+class SequenceAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) :
     GameAi<SequenceState, SequenceMove> {
 
-    enum class Difficulty(val blockWeight: Int, val jackPatience: Int) {
-        /** Builds its own runs and mostly ignores yours. */
-        EASY(blockWeight = 1, jackPatience = 0),
+    /** The numbers one seat plays with, worked out from its level and style. */
+    private class Tuning(
+        val ownWeight: Double,
+        val blockWeight: Double,
+        val jackPatience: Double,
+        val finishing: Boolean,
+    )
 
-        NORMAL(blockWeight = 3, jackPatience = 3),
-
-        /** Blocks aggressively and saves jacks for real threats. */
-        CAUTIOUS(blockWeight = 5, jackPatience = 4),
+    private fun tuningOf(profile: AiProfile): Tuning {
+        val (block, patience) = when (profile.level) {
+            AiLevel.EASY -> 1.0 to 0.0
+            AiLevel.MEDIUM -> 3.0 to 3.0
+            AiLevel.HARD -> 5.0 to 4.0
+        }
+        val style = profile.style
+        val easy = profile.level == AiLevel.EASY
+        return Tuning(
+            ownWeight = 4.0 + style.boldness,
+            blockWeight = block - style.boldness * 0.8 + style.spite * 2.0,
+            // Only a seat that is paying attention changes its mind about jacks.
+            jackPatience = if (easy) patience else patience + style.boldness - style.spite,
+            finishing = profile.level == AiLevel.HARD,
+        )
     }
 
     override fun chooseMove(
@@ -28,15 +52,30 @@ class SequenceAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
         legal: List<SequenceMove>,
     ): SequenceMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
+        val profile = profiles.of(seat)
+        val tuning = tuningOf(profile)
         val team = state.teamOf(seat)
 
         // Clearing a dead card is free value when nothing better is on offer.
         val exchange = legal.filterIsInstance<ExchangeDeadCard>().firstOrNull()
         if (exchange != null && legal.none { it is PlaceChip }) return exchange
 
-        val best = legal
+        val placements = legal
             .filterIsInstance<PlaceChip>()
-            .maxByOrNull { scorePlacement(state, it, team) }
+            .map { it to scorePlacement(state, it, team, tuning) }
+            .sortedByDescending { it.second }
+            .map { it.first }
+        val best = if (placements.isEmpty()) {
+            null
+        } else if (tuning.finishing) {
+            // A run that completes, or a run of the opponent's that would, is
+            // never something to be talked out of.
+            placements.first()
+        } else {
+            // Easy only: the others never slip. Only the better few are in play,
+            // so a slip is a weaker square rather than a wasted card.
+            profile.pick(placements.take(4))
+        }
 
         val removal = legal
             .filterIsInstance<RemoveChip>()
@@ -45,7 +84,7 @@ class SequenceAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
         // Spend a one-eyed jack only on a genuinely dangerous chip.
         if (removal != null) {
             val threat = threatAt(state, removal.cell, state.chips[removal.cell])
-            if (threat >= difficulty.jackPatience) return removal
+            if (threat >= tuning.jackPatience && !profile.slips()) return removal
         }
 
         return best ?: removal ?: exchange ?: legal.first()
@@ -56,14 +95,25 @@ class SequenceAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
      * opponent's run it interrupts. Wild jacks are discounted so they are not
      * burned on an ordinary square.
      */
-    private fun scorePlacement(state: SequenceState, move: PlaceChip, team: Int): Int {
+    private fun scorePlacement(
+        state: SequenceState,
+        move: PlaceChip,
+        team: Int,
+        tuning: Tuning,
+    ): Double {
         val ours = runPotential(state, move.cell, team)
         val theirs = (0 until state.options.teamCount)
             .filter { it != team }
             .maxOfOrNull { runPotential(state, move.cell, it) }
             ?: 0
 
-        var score = ours * 4 + theirs * difficulty.blockWeight
+        var score = ours * tuning.ownWeight + theirs * tuning.blockWeight
+        if (tuning.finishing) {
+            // A potential of five through this square means placing here
+            // completes the run (for whoever places it).
+            if (ours >= RUN_LENGTH) score += 100.0
+            if (theirs >= RUN_LENGTH) score += 60.0 else if (theirs == RUN_LENGTH - 1) score += 8.0
+        }
         // Central squares sit on more possible lines than edge squares.
         val row = SequenceBoard.rowOf(move.cell)
         val col = SequenceBoard.colOf(move.cell)

app/src/main/java/org/prolibertate/games/game/tayu/TayuAi.kt

@@ -1,5 +1,8 @@
 package org.prolibertate.games.game.tayu
 
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 
 /**
@@ -18,17 +21,53 @@ import org.prolibertate.games.game.engine.GameAi
  *
  * [level] overrides the table's own setting, which is what the tests use;
  * ordinarily it is left null and read from [TayuOptions.level].
+ *
+ * Per seat, from [profiles]: the table's setting says which terms count at all
+ * (Gentle ignores the opponent entirely and that is respected at every level).
+ * Within Full strength, Easy blocks and steers less and sometimes takes one of
+ * its better few placements rather than the best; Hard blocks and steers harder.
+ * Boldness weights its own axis against the opponent's, spite weights blocking.
  */
-class TayuAi(private val level: TayuLevel? = null) : GameAi<TayuState, TayuMove> {
+class TayuAi(
+    private val profiles: AiProfiles = AiProfiles.DEFAULT,
+    private val level: TayuLevel? = null,
+) : GameAi<TayuState, TayuMove> {
+
+    constructor(level: TayuLevel) : this(AiProfiles.DEFAULT, level)
+
+    private class Weights(val own: Double, val block: Double, val reach: Double)
+
+    private fun weightsOf(setting: TayuLevel, profile: AiProfile): Weights {
+        var own = setting.ownWeight.toDouble()
+        var block = setting.blockWeight.toDouble()
+        var reach = setting.reachWeight.toDouble()
+        // Gentle is a table rule, not a skill level: it stays as it is.
+        if (setting.blockWeight > 0) {
+            when (profile.level) {
+                AiLevel.EASY -> { block *= 0.4; reach *= 0.5 }
+                AiLevel.MEDIUM -> Unit
+                AiLevel.HARD -> { block *= 1.35; reach *= 1.5 }
+            }
+            own *= 1.0 + 0.15 * profile.style.boldness
+            block *= 1.0 - 0.15 * profile.style.boldness + 0.4 * profile.style.spite
+        }
+        return Weights(own, block, reach)
+    }
 
     override fun chooseMove(state: TayuState, seat: Int, legal: List<TayuMove>): TayuMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
+        val profile = profiles.of(seat)
         val setting = level ?: state.options.level
         if (legal.size == 1) return legal.first()
-        return legal.maxByOrNull { rate(state, seat, it, setting) } ?: legal.first()
+        val weights = weightsOf(setting, profile)
+        val rated = legal.map { it to rate(state, seat, it, weights) }
+        if (profile.level != AiLevel.EASY) return rated.maxByOrNull { it.second }?.first ?: legal.first()
+        // Easy: one of its better few placements, not always the best.
+        val ranked = rated.sortedByDescending { it.second }.take(EASY_POOL).map { it.first }
+        return profile.pick(ranked)
     }
 
-    private fun rate(state: TayuState, seat: Int, move: TayuMove, setting: TayuLevel): Int {
+    private fun rate(state: TayuState, seat: Int, move: TayuMove, weights: Weights): Double {
         val after = TayuRules.previewPlacement(state, seat, move)
         val mine = state.axisOf(seat)
         val theirs = if (mine == Axis.NORTH_SOUTH) Axis.EAST_WEST else Axis.NORTH_SOUTH
@@ -43,15 +82,15 @@ class TayuAi(private val level: TayuLevel? = null) : GameAi<TayuState, TayuMove>
         // Skipped rather than multiplied by zero: the gentle setting is the whole
         // reason to avoid walking the open mouths at all, and this runs for every
         // one of a thousand-odd candidate placements.
-        val reach = if (setting.reachWeight == 0) {
+        val reach = if (weights.reach == 0.0) {
             0
         } else {
             (pull(after, mine) - pull(after, theirs)) / REACH_SCALE
         }
 
-        return setting.ownWeight * ours -
-            setting.blockWeight * opposing +
-            setting.reachWeight * reach
+        return weights.own * ours -
+            weights.block * opposing +
+            weights.reach * reach
     }
 
     /**
@@ -82,5 +121,8 @@ class TayuAi(private val level: TayuLevel? = null) : GameAi<TayuState, TayuMove>
     private companion object {
         /** Brings the open-mouth term onto the same scale as the score term. */
         const val REACH_SCALE = 8
+
+        /** How many top-rated placements Easy chooses among. */
+        const val EASY_POOL = 6
     }
 }

app/src/main/java/org/prolibertate/games/game/wizard/WizardAi.kt

@@ -3,8 +3,13 @@ package org.prolibertate.games.game.wizard
 import org.prolibertate.games.game.cards.Card
 import org.prolibertate.games.game.cards.Rank
 import org.prolibertate.games.game.cards.Suit
+import org.prolibertate.games.game.cards.Decks
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
 import kotlin.math.abs
+import kotlin.math.pow
 import kotlin.math.roundToInt
 
 /**
@@ -15,52 +20,99 @@ import kotlin.math.roundToInt
  * high it is and whether it is trump. In play it chases tricks while it is
  * short of its bid and ducks once it has them, which is the whole game.
  */
-class WizardAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
+class WizardAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) :
     GameAi<WizardState, WizardMove> {
 
-    enum class Difficulty(
-        /**
-         * Added to the expected-trick count before rounding. Optimists overbid
-         * and go down; pessimists underbid and get caught taking tricks they
-         * did not want.
-         */
-        val bidBias: Double,
-    ) {
-        EASY(bidBias = 0.35),
-        NORMAL(bidBias = 0.0),
-        CAUTIOUS(bidBias = -0.25),
+    private companion object {
+        /** Optimists overbid and go down; Easy leans this way. */
+        const val EASY_BID_BIAS = 0.35
+        val DECK: List<Card> = Decks.wizard()
+    }
+
+    /** Cards this AI has watched played in the current round, per seat it plays. */
+    private val memory = HashMap<Int, SeenCards>()
+
+    private class SeenCards {
+        var key: Long = Long.MIN_VALUE
+        val cards = HashSet<Card>()
+    }
+
+    /** Public information only: cards that have appeared on the table this round. */
+    private fun observe(state: WizardState, seat: Int): Set<Card> = synchronized(memory) {
+        val seen = memory.getOrPut(seat) { SeenCards() }
+        val key = state.seed * 1_000_003L + state.round
+        if (key != seen.key || state.phase == WizardPhase.BIDDING) {
+            seen.key = key
+            seen.cards.clear()
+        }
+        state.trick.forEach { seen.cards.add(it.card) }
+        state.completedTrick.forEach { seen.cards.add(it.card) }
+        seen.cards.removeAll(state.hands[seat].toSet())
+        seen.cards.toSet()
     }
 
     override fun chooseMove(state: WizardState, seat: Int, legal: List<WizardMove>): WizardMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
-        return when (state.phase) {
-            WizardPhase.BIDDING -> chooseBid(state, seat, legal)
-            WizardPhase.PLAYING -> choosePlay(state, seat, legal)
-            else -> legal.first()
+        val profile = profiles.of(seat)
+        val seen = if (profile.level == AiLevel.HARD) observe(state, seat) else emptySet()
+        val move = try {
+            when (state.phase) {
+                WizardPhase.BIDDING -> chooseBid(state, seat, legal, profile)
+                WizardPhase.PLAYING -> choosePlay(state, seat, legal, profile, seen)
+                else -> legal.first()
+            }
+        } catch (e: RuntimeException) {
+            legal.first()
         }
+        return if (move in legal) move else legal.first()
     }
 
     // -----------------------------------------------------------------------
     // Bidding
     // -----------------------------------------------------------------------
 
-    private fun chooseBid(state: WizardState, seat: Int, legal: List<WizardMove>): WizardMove {
+    private fun chooseBid(
+        state: WizardState,
+        seat: Int,
+        legal: List<WizardMove>,
+        profile: AiProfile,
+    ): WizardMove {
+        val hand = state.hands[seat]
+        val hard = profile.level == AiLevel.HARD
+        val easy = profile.level == AiLevel.EASY
+
         val naming = legal.filterIsInstance<ChooseTrump>()
         if (naming.isNotEmpty()) {
-            val hand = state.hands[seat]
-            val best = naming.maxByOrNull { suitStrength(hand, it.suit) }
-            return best ?: naming.first()
+            return when {
+                hard -> naming.maxByOrNull { hardExpected(state, hand, it.suit) }
+                easy -> {
+                    val noisy = naming.associateWith { suitStrength(hand, it.suit) + profile.random.nextDouble(-0.8, 0.8) }
+                    profile.pick(naming.sortedByDescending { noisy.getValue(it) })
+                }
+                else -> naming.maxByOrNull { suitStrength(hand, it.suit) }
+            } ?: naming.first()
         }
 
         val bids = legal.filterIsInstance<MakeBid>()
         if (bids.isEmpty()) return legal.first()
 
-        val wanted = (expectedTricks(state.hands[seat], state.trump) + difficulty.bidBias)
-            .roundToInt()
-            .coerceIn(0, state.cardsThisRound)
+        var expected = if (hard) hardExpected(state, hand, state.trump) else expectedTricks(hand, state.trump)
+        if (hard) {
+            // The last to bid knows how many tricks the others have left on the table.
+            val others = state.bids.filterNotNull()
+            if (others.size == state.options.playerCount - 1) {
+                expected = expected * 0.75 + (state.cardsThisRound - others.sum()) * 0.25
+            }
+        }
+        if (easy) expected += EASY_BID_BIAS + profile.random.nextDouble(-0.8, 0.8)
+        // A gambler bids up, a cautious seat bids down.
+        expected += profile.bolder(0.0, 0.3)
+
+        val wanted = expected.roundToInt().coerceIn(0, state.cardsThisRound)
         // Screw-the-dealer may have taken the number we wanted away, so settle
         // for the nearest one still on offer.
-        return bids.minBy { abs(it.tricks - wanted) }
+        val ranked = bids.sortedBy { abs(it.tricks - wanted) }
+        return if (easy) profile.pick(ranked) else ranked.first()
     }
 
     /** How much of a hand a given trump suit would be worth. */
@@ -92,22 +144,88 @@ class WizardAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
         return expected
     }
 
+    /**
+     * The chance [card] takes a trick, from the cards that could still beat it.
+     *
+     * [pool] is every card that might be in somebody else's hand and [q] the
+     * chance any one of them is, and each higher card only threatens the trick
+     * if it lands in the same one, which is less likely the longer the round.
+     */
+    private fun winProb(card: Card, pool: List<Card>, trump: Suit?, q: Double, tricks: Int): Double {
+        if (isWizardCard(card)) return 1.0
+        if (isJester(card)) return 0.0
+        val wizards = pool.count { isWizardCard(it) && it != card }
+        val ordinary = { c: Card -> !isWizardCard(c) && !isJester(c) }
+        val trumpCard = trump != null && card.suit == trump
+        val higher = pool.count { ordinary(it) && it.suit == card.suit && it.rank.order > card.rank.order }
+        var danger = (higher + wizards).toDouble()
+        if (!trumpCard && trump != null) {
+            val voidFactor = 1.0 / (1.0 + 0.5 * (tricks - 1))
+            danger += voidFactor * pool.count { ordinary(it) && it.suit == trump }
+        }
+        val sequencing = 1.0 / (1.0 + 0.35 * (tricks - 1))
+        return (1.0 - (q * sequencing).coerceIn(0.0, 1.0)).pow(danger)
+    }
+
+    /** Expected tricks from the cards that could still be out, calibrated so the average hand scores its fair share. */
+    private fun hardExpected(state: WizardState, hand: List<Card>, trump: Suit?): Double {
+        val n = state.cardsThisRound
+        val players = state.options.playerCount
+        val inPlay = DECK.filter { it != state.trumpCard }
+        val mine = inPlay.filter { it !in hand }
+        val q = ((players - 1) * n).toDouble() / mine.size.coerceAtLeast(1)
+
+        val own = hand.sumOf { winProb(it, mine, trump, q, n) }
+        val average = inPlay.sumOf { winProb(it, inPlay, trump, q, n) } / inPlay.size
+        val scale = if (average > 0.0) (1.0 / (players * average)).coerceIn(0.5, 2.0) else 1.0
+        return own * scale
+    }
+
     // -----------------------------------------------------------------------
     // Play
     // -----------------------------------------------------------------------
 
-    private fun choosePlay(state: WizardState, seat: Int, legal: List<WizardMove>): WizardMove {
+    private fun choosePlay(
+        state: WizardState,
+        seat: Int,
+        legal: List<WizardMove>,
+        profile: AiProfile,
+        seen: Set<Card>,
+    ): WizardMove {
         val cards = legal.filterIsInstance<PlayCard>().map { it.card }
         if (cards.size == 1) return PlayCard(cards.first())
 
+        val hard = profile.level == AiLevel.HARD
         val bid = state.bids[seat] ?: 0
         val stillNeeded = bid - state.tricksWon[seat]
         val tricksLeft = state.hands[seat].size
         // Chase while we are short, and only while there are enough tricks left
-        // to be worth chasing.
-        val wantsTrick = stillNeeded > 0 && stillNeeded <= tricksLeft
+        // to be worth chasing. (The careful player keeps chasing even when the
+        // bid is out of reach: every trick still cuts the penalty.)
+        val wantsTrick = stillNeeded > 0 && (hard || stillNeeded <= tricksLeft)
+        // A spoiler who has already blown their own bid takes tricks off anyone
+        // who needs exactly one more.
+        val spoiling = profile.style.spite >= 0.5 && stillNeeded < 0
 
-        if (state.trick.isEmpty()) return PlayCard(chooseLead(cards, state.trump, wantsTrick))
+        val best = if (hard) {
+            hardPlay(state, seat, cards, seen, wantsTrick, spoiling, stillNeeded, tricksLeft)
+        } else {
+            standardPlay(state, cards, wantsTrick, spoiling)
+        }
+        if (profile.level != AiLevel.EASY) return PlayCard(best)
+
+        // Easy sometimes burns a strong card at the wrong moment.
+        val ranked = listOf(best) + (cards - best).sortedByDescending { it.rank.order }
+        return PlayCard(profile.pick(ranked))
+    }
+
+    private fun standardPlay(
+        state: WizardState,
+        cards: List<Card>,
+        wantsTrick: Boolean,
+        spoiling: Boolean,
+    ): Card {
+        if (state.trick.isEmpty()) return chooseLead(cards, state.trump, wantsTrick)
 
         val ledSuit = ledSuitOf(state.trick)
         val order = state.trick.size
@@ -122,15 +240,99 @@ class WizardAi(private val difficulty: Difficulty = Difficulty.NORMAL) :
 
         return when {
             // Take it as cheaply as we can.
-            wantsTrick && winners.isNotEmpty() -> PlayCard(winners.minBy { strengthOf(it) })
+            wantsTrick && winners.isNotEmpty() -> winners.minBy { strengthOf(it) }
+            // Spite: stop somebody from landing the bid they need.
+            spoiling && winners.isNotEmpty() && victimNeedsOne(state) -> winners.minBy { strengthOf(it) }
             // Cannot take it, or do not want it: throw the least useful card.
-            losers.isNotEmpty() -> PlayCard(losers.minBy { strengthOf(it) })
+            losers.isNotEmpty() -> losers.minBy { strengthOf(it) }
             // Everything we hold wins, and we did not want the trick. Take it
             // with the weakest winner so the good cards stay for later.
-            else -> PlayCard(cards.minBy { strengthOf(it) })
+            else -> cards.minBy { strengthOf(it) }
         }
     }
 
+    /** Whether whoever is winning the trick right now needs exactly this trick to make their bid. */
+    private fun victimNeedsOne(state: WizardState): Boolean {
+        val ledSuit = ledSuitOf(state.trick)
+        val leading = state.trick
+            .withIndex()
+            .maxByOrNull { (index, played) -> trickStrength(played.card, index, state.trump, ledSuit) }
+            ?: return false
+        val victim = leading.value.seat
+        return (state.bids[victim] ?: return false) - state.tricksWon[victim] == 1
+    }
+
+    /**
+     * The strong player's choice. It keeps track of which cards are gone to
+     * judge how likely each of its own is to win, spends sure winners when it
+     * needs tricks, and sheds its most dangerous cards onto tricks it will
+     * lose anyway when it does not.
+     */
+    private fun hardPlay(
+        state: WizardState,
+        seat: Int,
+        cards: List<Card>,
+        seen: Set<Card>,
+        wantsTrick: Boolean,
+        spoiling: Boolean,
+        stillNeeded: Int,
+        tricksLeft: Int,
+    ): Card {
+        val hand = state.hands[seat]
+        val unseen = DECK.filter { it != state.trumpCard && it !in seen && it !in hand }
+        val oppCards = state.handCounts.sum() - hand.size
+        val q = if (unseen.isEmpty()) 0.0 else (oppCards.toDouble() / unseen.size).coerceIn(0.0, 1.0)
+        val trump = state.trump
+        fun pWin(card: Card): Double = winProb(card, unseen, trump, q, tricksLeft)
+
+        if (state.trick.isEmpty()) {
+            if (wantsTrick) {
+                val scored = cards.map { it to pWin(it) }
+                val bestOverall = scored.maxBy { it.second }
+                // With tricks to spare, save the wizard for when it is needed.
+                if (tricksLeft > stillNeeded) {
+                    val bestPlain = scored.filter { !isWizardCard(it.first) }.maxByOrNull { it.second }
+                    if (bestPlain != null && bestPlain.second >= 0.6) return bestPlain.first
+                }
+                return bestOverall.first
+            }
+            // Ducking: lead whatever is least likely to win, a jester best of all.
+            return cards.minWith(compareBy<Card>({ pWin(it) }, { it.rank.order }))
+        }
+
+        val ledSuit = ledSuitOf(state.trick)
+        val order = state.trick.size
+        val remaining = state.options.playerCount - order - 1
+        val bestOnTable = state.trick
+            .withIndex()
+            .maxOf { (index, played) -> trickStrength(played.card, index, trump, ledSuit) }
+        val strengthOf = { card: Card -> trickStrength(card, order, trump, ledSuit) }
+        val winners = cards.filter { strengthOf(it) > bestOnTable }
+        val losers = cards.filter { strengthOf(it) <= bestOnTable }
+
+        fun sure(card: Card): Boolean =
+            remaining == 0 || unseen.none { trickStrength(it, order + 1, trump, ledSuit) > strengthOf(card) }
+
+        if (wantsTrick) {
+            if (winners.isNotEmpty()) {
+                val certain = winners.filter { sure(it) }
+                return (certain.ifEmpty { winners }).minBy { strengthOf(it) }
+            }
+            // Cannot take this one: keep the strong cards for a trick that can be won.
+            return losers.minBy { strengthOf(it) }
+        }
+        if (spoiling && winners.isNotEmpty() && victimNeedsOne(state)) {
+            return winners.minBy { strengthOf(it) }
+        }
+        if (losers.isNotEmpty()) {
+            // Unload the card that would be hardest to lose later, but keep the
+            // jester: it is the one card that can never win.
+            val nonJester = losers.filter { !isJester(it) }
+            return (nonJester.ifEmpty { losers }).maxBy { strengthOf(it) }
+        }
+        return cards.minBy { strengthOf(it) }
+    }
+
     private fun chooseLead(cards: List<Card>, trump: Suit?, wantsTrick: Boolean): Card {
         if (wantsTrick) {
             // A led wizard is an unloseable trick.

app/src/main/java/org/prolibertate/games/game/yahtzee/YahtzeeAi.kt

@@ -1,6 +1,11 @@
 package org.prolibertate.games.game.yahtzee
 
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.GameAi
+import kotlin.math.exp
+import kotlin.math.sqrt
 
 /**
  * Plays the dice by eye rather than by arithmetic.
@@ -10,33 +15,53 @@ import org.prolibertate.games.game.engine.GameAi
  * enumerated the thirty-two subsets and applied them would be reading the roll
  * before making it — beating the player by seeing the future rather than by
  * playing well. Everything below decides from the dice already on the table.
+ *
+ * Per seat, from [profiles]: Easy sometimes holds a poor set of dice and
+ * sometimes writes a lesser box than the best on offer. Hard works out, for
+ * every way of holding dice, the *expected* value of what could be thrown
+ * (averaging over all outcomes, never over the roll that will actually come),
+ * valuing each box by what it scores less what that box is usually worth, with
+ * the upper-section bonus tracked. Boldness chases the bonus and big combos
+ * and rolls on with slimmer odds; caution takes a sure score sooner.
  */
-object YahtzeeAi : GameAi<YahtzeeState, YahtzeeMove> {
+class YahtzeeAi(private val profiles: AiProfiles = AiProfiles.DEFAULT) :
+    GameAi<YahtzeeState, YahtzeeMove> {
 
     override fun chooseMove(state: YahtzeeState, seat: Int, legal: List<YahtzeeMove>): YahtzeeMove {
         require(legal.isNotEmpty()) { "No legal move for seat $seat" }
+        val profile = profiles.of(seat)
         val rolls = legal.filterIsInstance<RollDice>()
         val writes = legal.filterIsInstance<ScoreIn>()
 
         // Nothing on the table yet, so there is only one thing to do.
         if (!state.hasRolled) return rolls.firstOrNull() ?: writes.first()
+        if (profile.level == AiLevel.HARD && writes.isNotEmpty()) {
+            return chooseExpected(state, seat, rolls, writes, profile)
+        }
 
-        val best = writes.maxByOrNull { scoreOf(it.category, state.dice) }
+        val ranked = writes.sortedByDescending { scoreOf(it.category, state.dice) }
+        // Easy sometimes writes the second or third best box rather than the best.
+        val best = if (ranked.isEmpty()) null else profile.pick(ranked.take(EASY_BOXES))
         val bestScore = best?.let { scoreOf(it.category, state.dice) } ?: 0
 
         // A hand already worth taking, or no rolls left, ends the turn.
         if (rolls.isEmpty()) return best ?: writes.first()
-        if (bestScore >= WORTH_KEEPING) return best!!
+        if (bestScore >= profile.bolder(WORTH_KEEPING.toDouble(), 4.0)) return best!!
 
-        val keep = whatToKeep(state.dice).toSet()
+        val keep = if (profile.slips()) poorKeep(state.dice) else whatToKeep(state.dice).toSet()
         return rolls.firstOrNull { it.keep == keep } ?: rolls.first()
     }
 
-    /**
-     * Good enough to stop throwing. Set at a full house, which is the point
-     * where a hand is worth more than an average reroll of it.
-     */
-    private const val WORTH_KEEPING = 25
+    /** What a distracted player holds: the high dice, or the pair they notice second. */
+    private fun poorKeep(dice: List<Int>): Set<Int> {
+        val counts = countsOf(dice)
+        val second = counts.entries
+            .sortedWith(compareByDescending<Map.Entry<Int, Int>> { it.value }.thenBy { it.key })
+            .drop(1)
+            .firstOrNull { it.value >= 2 }
+        if (second != null) return dice.indices.filter { dice[it] == second.key }.toSet()
+        return dice.indices.filter { dice[it] >= 4 }.toSet()
+    }
 
     /**
      * What to hold on to: a run if one is nearly there, otherwise the face there
@@ -82,4 +107,213 @@ object YahtzeeAi : GameAi<YahtzeeState, YahtzeeMove> {
             face in wanted && taken.add(face)
         }
     }
+
+    // -- Hard: expected values ------------------------------------------------
+
+    /**
+     * Values every box as what it scores now less what it is normally worth, so
+     * that writing a zero in Ones costs little and writing one in Yahtzee costs a
+     * lot, then looks two throws ahead by averaging over every possible outcome
+     * of every way of holding dice.
+     */
+    private fun chooseExpected(
+        state: YahtzeeState,
+        seat: Int,
+        rolls: List<RollDice>,
+        writes: List<ScoreIn>,
+        profile: AiProfile,
+    ): YahtzeeMove {
+        val valuation = Valuation(state, seat, profile.style.boldness)
+        val dice = state.dice
+        val stop = writes.maxByOrNull { valuation.gain(it.category, dice) } ?: writes.first()
+        if (rolls.isEmpty()) return stop
+        val stopValue = valuation.gain(stop.category, dice)
+
+        // terminal[i]: best write for the i-th multiset of dice with no throws left.
+        val terminal = DoubleArray(DiceTables.multisets.size) { i ->
+            val faces = DiceTables.facesOf(i)
+            writes.maxOf { valuation.gain(it.category, faces) }
+        }
+        // Each throw still to come after the next one, worked backwards.
+        var next = terminal
+        repeat(state.rollsLeft - 1) {
+            val previous = next
+            next = DoubleArray(terminal.size) { i ->
+                maxOf(terminal[i], bestKeep(i, previous).second)
+            }
+        }
+        val here = DiceTables.indexOf(countVector(dice))
+        val (keepCounts, expected) = bestKeep(here, next)
+        val margin = -profile.style.boldness
+        if (expected <= stopValue + margin) return stop
+
+        val remaining = keepCounts.copyOf()
+        val keep = dice.indices.filter { position ->
+            val face = dice[position] - 1
+            if (remaining[face] > 0) {
+                remaining[face]--
+                true
+            } else {
+                false
+            }
+        }.toSet()
+        return rolls.firstOrNull { it.keep == keep } ?: stop
+    }
+
+    /** The best set of dice to hold from multiset [index], and what throwing the rest is worth. */
+    private fun bestKeep(index: Int, after: DoubleArray): Pair<IntArray, Double> {
+        var bestKeep = DiceTables.keepsOf(index).first()
+        var best = Double.NEGATIVE_INFINITY
+        for (keep in DiceTables.keepsOf(index)) {
+            val thrown = DICE_COUNT - keep.sum()
+            var key = 0
+            for (face in 0 until 6) key += keep[face] * DiceTables.POWERS[face]
+            var total = 0.0
+            for (outcome in DiceTables.rerolls[thrown]) {
+                total += outcome.probability * after[DiceTables.indexOfKey[key + outcome.key]]
+            }
+            if (total > best + 1e-9) {
+                best = total
+                bestKeep = keep
+            }
+        }
+        return bestKeep to best
+    }
+
+    private fun countVector(dice: List<Int>): IntArray {
+        val counts = IntArray(6)
+        dice.forEach { counts[it - 1]++ }
+        return counts
+    }
+
+    /** What a box is worth to this card right now, in points relative to its usual yield. */
+    private class Valuation(state: YahtzeeState, seat: Int, private val boldness: Double) {
+        private val card = state.cards[seat]
+        private val yahtzeeBonus = state.options.yahtzeeBonus && card[YahtzeeCategory.YAHTZEE] == YAHTZEE_SCORE
+        private val openUpper = YahtzeeCategory.entries.filter {
+            it.section == YahtzeeSection.UPPER && !card.isFilled(it)
+        }
+        private val secured = card.upperSubtotal >= UPPER_BONUS_AT
+        private val diff = YahtzeeCategory.entries
+            .filter { it.section == YahtzeeSection.UPPER && card.isFilled(it) }
+            .sumOf { (card[it] ?: 0) - 3 * it.face!! }
+        private val bonusValue = UPPER_BONUS * (1.0 + 0.4 * boldness)
+
+        fun gain(category: YahtzeeCategory, dice: List<Int>): Double {
+            val score = scoreOf(category, dice)
+            var value = score - TYPICAL[category.ordinal]
+            if (category.section == YahtzeeSection.UPPER && !secured) {
+                val face = category.face!!
+                val after = if (card.upperSubtotal + score >= UPPER_BONUS_AT) {
+                    1.0
+                } else {
+                    chance(diff + score - 3 * face, openUpper - category)
+                }
+                value += bonusValue * (after - chance(diff, openUpper))
+            }
+            // The big combinations: a bold seat values them a little more.
+            if (score > 0 && (category == YahtzeeCategory.LARGE_STRAIGHT ||
+                    category == YahtzeeCategory.YAHTZEE || category == YahtzeeCategory.FOUR_OF_A_KIND)
+            ) {
+                value += boldness * 2.0
+            }
+            if (yahtzeeBonus && isYahtzee(dice)) value += YAHTZEE_BONUS
+            return value
+        }
+
+        /** Likelihood the upper section still reaches the bonus from [difference] with [open] to fill. */
+        private fun chance(difference: Int, open: List<YahtzeeCategory>): Double {
+            if (open.isEmpty()) return if (difference >= 0) 1.0 else 0.0
+            val expected = open.sumOf { TYPICAL[it.ordinal] - 3 * it.face!! }
+            val spread = 2.2 * sqrt(open.size.toDouble()) + 0.3
+            return 1.0 / (1.0 + exp(-(difference + expected + 1.5) / spread))
+        }
+    }
+
+    /** Every multiset of five dice and every throw of n dice, worked out once. */
+    private object DiceTables {
+        class Outcome(val key: Int, val probability: Double)
+
+        val POWERS = intArrayOf(1, 6, 36, 216, 1296, 7776)
+        val indexOfKey = IntArray(46656) { -1 }
+        val multisets: List<IntArray> = vectors(DICE_COUNT)
+        val rerolls: Array<List<Outcome>> = Array(DICE_COUNT + 1) { n ->
+            val factorial = IntArray(6) { f -> (1..f).fold(1) { a, b -> a * b } }
+            val total = Math.pow(6.0, n.toDouble())
+            vectors(n).map { counts ->
+                val ways = (1..n).fold(1) { a, b -> a * b } / counts.fold(1) { a, c -> a * factorial[c] }
+                Outcome(keyOf(counts), ways / total)
+            }
+        }
+        private val keeps: Array<List<IntArray>?> = arrayOfNulls(multisets.size)
+
+        init {
+            multisets.forEachIndexed { i, counts -> indexOfKey[keyOf(counts)] = i }
+        }
+
+        fun indexOf(counts: IntArray): Int = indexOfKey[keyOf(counts)]
+
+        fun facesOf(index: Int): List<Int> {
+            val faces = ArrayList<Int>(DICE_COUNT)
+            multisets[index].forEachIndexed { f, c -> repeat(c) { faces += f + 1 } }
+            return faces
+        }
+
+        /** Every distinct set of dice that can be held back from multiset [index]. */
+        fun keepsOf(index: Int): List<IntArray> {
+            keeps[index]?.let { return it }
+            val counts = multisets[index]
+            var all = listOf(IntArray(6))
+            for (face in 0 until 6) {
+                all = all.flatMap { partial ->
+                    (0..counts[face]).map { n -> partial.copyOf().also { it[face] = n } }
+                }
+            }
+            keeps[index] = all
+            return all
+        }
+
+        private fun keyOf(counts: IntArray): Int {
+            var key = 0
+            for (face in 0 until 6) key += counts[face] * POWERS[face]
+            return key
+        }
+
+        /** Every way to split [size] dice over six faces. */
+        private fun vectors(size: Int): List<IntArray> {
+            val out = ArrayList<IntArray>()
+            fun build(face: Int, left: Int, current: IntArray) {
+                if (face == 5) {
+                    current[5] = left
+                    out += current.copyOf()
+                    return
+                }
+                for (n in 0..left) {
+                    current[face] = n
+                    build(face + 1, left - n, current)
+                }
+            }
+            build(0, size, IntArray(6))
+            return out
+        }
+    }
+
+    private companion object {
+        /**
+         * Good enough to stop throwing. Set at a full house, which is the point
+         * where a hand is worth more than an average reroll of it.
+         */
+        const val WORTH_KEEPING = 25
+
+        /** How many of the best boxes Easy chooses among when it writes a score. */
+        const val EASY_BOXES = 3
+
+        /**
+         * What each box yields on average under strong play, in the order of
+         * [YahtzeeCategory]: the cost of spending it.
+         */
+        val TYPICAL = doubleArrayOf(
+            2.11, 5.28, 8.57, 12.16, 15.69, 19.19, 21.66, 13.10, 22.59, 29.46, 32.71, 16.87, 22.01,
+        )
+    }
 }

app/src/main/java/org/prolibertate/games/net/LobbyController.kt

@@ -8,6 +8,7 @@ import kotlinx.coroutines.flow.MutableStateFlow
 import kotlinx.coroutines.flow.StateFlow
 import kotlinx.coroutines.flow.asStateFlow
 import kotlinx.coroutines.launch
+import org.prolibertate.games.game.engine.AiDifficulty
 import org.prolibertate.games.game.engine.PlayerKind
 import org.prolibertate.games.game.engine.PlayerSlot
 import org.prolibertate.games.game.engine.TableConfig
@@ -33,6 +34,7 @@ class LobbyController(
         val connected: Boolean = false,
         val gameId: String? = null,
         val optionsJson: String = "{}",
+        val aiDifficulty: AiDifficulty = AiDifficulty.MEDIUM,
         val hostName: String = "",
         val seats: List<PlayerSlot> = emptyList(),
         val discovered: List<DiscoveredHost> = emptyList(),
@@ -88,6 +90,7 @@ class LobbyController(
         hostName: String,
         seatCount: Int,
         teamOf: (Int) -> Int,
+        aiDifficulty: AiDifficulty = AiDifficulty.MEDIUM,
     ) {
         stop()
         val seats = (0 until seatCount).map { seat ->
@@ -102,6 +105,7 @@ class LobbyController(
             hosting = true,
             gameId = gameId,
             optionsJson = optionsJson,
+            aiDifficulty = aiDifficulty,
             hostName = hostName,
             seats = seats,
             message = "Waiting for players…",
@@ -237,6 +241,7 @@ class LobbyController(
             seats = current.seats,
             optionsJson = current.optionsJson,
             seed = seed,
+            aiDifficulty = current.aiDifficulty,
         )
         scope.launch { connections.values.forEach { it.send(StartGame(config)) } }
         _state.value = current.copy(started = config)

app/src/main/java/org/prolibertate/games/ui/AppRoot.kt

@@ -57,6 +57,7 @@ import org.prolibertate.games.update.UpdateChannel
 import org.prolibertate.games.update.Updater
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.withContext
+import org.prolibertate.games.game.engine.AiDifficulty
 import org.prolibertate.games.game.engine.SavedGameStore
 import org.prolibertate.games.game.engine.SavedMatch
 
@@ -92,7 +93,12 @@ sealed interface Route {
     /** Every saved game, to carry on from or delete. */
     data object Resume : Route
     data class Setup(val gameId: String) : Route
-    data class Lobby(val gameId: String, val optionsJson: String, val hosting: Boolean) : Route
+    data class Lobby(
+        val gameId: String,
+        val optionsJson: String,
+        val hosting: Boolean,
+        val aiDifficulty: AiDifficulty = AiDifficulty.MEDIUM,
+    ) : Route
     data class Play(
         val gameId: String,
         val config: TableConfig,
@@ -250,8 +256,8 @@ private fun AppContent(
                 playerName = settings.displayName,
                 onBack = { pop() },
                 onPlayOffline = { config, seat -> push(Route.Play(route.gameId, config, true, seat)) },
-                onHostOnline = { optionsJson ->
-                    push(Route.Lobby(route.gameId, optionsJson, hosting = true))
+                onHostOnline = { optionsJson, aiDifficulty ->
+                    push(Route.Lobby(route.gameId, optionsJson, hosting = true, aiDifficulty = aiDifficulty))
                 },
                 saved = saved,
                 onResume = { match ->

app/src/main/java/org/prolibertate/games/ui/GameSetupScreen.kt

@@ -30,6 +30,7 @@ import kotlinx.serialization.encodeToString
 import kotlinx.serialization.json.Json
 import org.prolibertate.games.game.GameCatalog
 import org.prolibertate.games.game.GameDescriptor
+import org.prolibertate.games.game.engine.AiDifficulty
 import org.prolibertate.games.game.engine.PlayerKind
 import org.prolibertate.games.game.engine.PlayerSlot
 import org.prolibertate.games.game.engine.SavedMatch
@@ -88,7 +89,7 @@ fun GameSetupScreen(
     playerName: String,
     onBack: () -> Unit,
     onPlayOffline: (TableConfig, Int) -> Unit,
-    onHostOnline: (String) -> Unit,
+    onHostOnline: (String, AiDifficulty) -> Unit,
     /** A game of this kind left half-played, if there is one. */
     saved: SavedMatch? = null,
     onResume: (SavedMatch) -> Unit = {},
@@ -124,6 +125,7 @@ fun GameSetupScreen(
     // Pirates and Bulgars is not symmetrical, so which side you take is the
     // first thing about the game rather than a detail of it.
     var playPirates by remember { mutableStateOf(false) }
+    var aiDifficulty by remember { mutableStateOf(AiDifficulty.MEDIUM) }
 
     val optionsJson = when (descriptor.id) {
         GameCatalog.EUCHRE -> setupJson.encodeToString(euchre)
@@ -243,6 +245,22 @@ fun GameSetupScreen(
                 style = MaterialTheme.typography.bodySmall,
             )
 
+            if (!solo) {
+                ChipRow(
+                    label = "Computer difficulty",
+                    values = AiDifficulty.entries.toList(),
+                    selected = aiDifficulty,
+                    display = { it.label },
+                    onSelect = { aiDifficulty = it },
+                )
+                if (aiDifficulty == AiDifficulty.MIXED) {
+                    Text(
+                        "Each computer player gets its own level.",
+                        style = MaterialTheme.typography.bodySmall,
+                    )
+                }
+            }
+
             // Two games let the local player pick a seat rather than always
             // taking the first one: chess by colour, and Pirates and Bulgars by
             // which side of a very lopsided fight they fancy.
@@ -259,13 +277,13 @@ fun GameSetupScreen(
             }
             PrimaryAction(text = if (solo) "Play" else "Play against the computer") {
                 onPlayOffline(
-                    offlineConfig(descriptor.id, optionsJson, seatCount, playerName, localSeat),
+                    offlineConfig(descriptor.id, optionsJson, seatCount, playerName, localSeat, aiDifficulty),
                     localSeat,
                 )
             }
             if (!solo) {
                 PrimaryAction(text = "Host a game for others to join") {
-                    onHostOnline(optionsJson)
+                    onHostOnline(optionsJson, aiDifficulty)
                 }
             }
         }
@@ -279,6 +297,7 @@ private fun offlineConfig(
     seatCount: Int,
     playerName: String,
     localSeat: Int = 0,
+    aiDifficulty: AiDifficulty = AiDifficulty.MEDIUM,
 ): TableConfig = TableConfig(
     gameId = gameId,
     seats = (0 until seatCount).map { seat ->
@@ -292,6 +311,7 @@ private fun offlineConfig(
     optionsJson = optionsJson,
     // Seeded from the clock so consecutive games are not identical.
     seed = System.currentTimeMillis(),
+    aiDifficulty = aiDifficulty,
 )
 
 /**

app/src/main/java/org/prolibertate/games/ui/LobbyScreen.kt

@@ -60,6 +60,7 @@ fun LobbyScreen(
                 hostName = playerName,
                 seatCount = seatCount,
                 teamOf = { seat -> teamForSeat(route.gameId, seat, seatCount) },
+                aiDifficulty = route.aiDifficulty,
             )
         } else {
             env.lobby.startDiscovery()

app/src/main/java/org/prolibertate/games/ui/PlayScreen.kt

@@ -120,6 +120,7 @@ import org.prolibertate.games.game.wizard.WizardRules
 import org.prolibertate.games.game.wizard.WizardState
 import org.prolibertate.games.net.MatchController
 import org.prolibertate.games.net.MatchSaver
+import org.prolibertate.games.game.engine.AiProfiles
 import org.prolibertate.games.game.engine.PlayerKind
 import org.prolibertate.games.settings.Settings
 import org.prolibertate.games.ui.game.BackgammonScreen
@@ -174,7 +175,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<EuchreState, EuchreMove>(
                         rules = EuchreRules,
-                        ai = EuchreAi(),
+                        ai = EuchreAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -213,7 +214,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<SequenceState, SequenceMove>(
                         rules = SequenceRules,
-                        ai = SequenceAi(),
+                        ai = SequenceAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -230,7 +231,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<PresidentState, PresidentMove>(
                         rules = PresidentRules,
-                        ai = PresidentAi(),
+                        ai = PresidentAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -254,7 +255,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<GolfState, GolfMove>(
                         rules = GolfRules,
-                        ai = GolfAi(),
+                        ai = GolfAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -276,7 +277,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<KaiserState, KaiserMove>(
                         rules = KaiserRules,
-                        ai = KaiserAi(),
+                        ai = KaiserAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -312,7 +313,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<WizardState, WizardMove>(
                         rules = WizardRules,
-                        ai = WizardAi(),
+                        ai = WizardAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -350,7 +351,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<CrazyEightsState, CrazyEightsMove>(
                         rules = CrazyEightsRules,
-                        ai = CrazyEightsAi(),
+                        ai = CrazyEightsAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -377,7 +378,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<CribbageState, CribbageMove>(
                         rules = CribbageRules,
-                        ai = CribbageAi(),
+                        ai = CribbageAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -406,7 +407,7 @@ fun PlayScreen(
                         rules = TayuRules,
                         // As with chess, the strength is part of the table's own
                         // options, so this does not have to decode anything.
-                        ai = TayuAi(),
+                        ai = TayuAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -423,7 +424,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<CheckersState, CheckersMove>(
                         rules = CheckersRules,
-                        ai = CheckersAi(),
+                        ai = CheckersAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -440,7 +441,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<BackgammonState, BackgammonMove>(
                         rules = BackgammonRules,
-                        ai = BackgammonAi(),
+                        ai = BackgammonAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -459,7 +460,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<MastermindState, MastermindMove>(
                         rules = MastermindRules,
-                        ai = MastermindAi(),
+                        ai = MastermindAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -476,7 +477,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<PiratesState, PiratesMove>(
                         rules = PiratesRules,
-                        ai = PiratesAi(),
+                        ai = PiratesAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -493,7 +494,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<HeartsState, HeartsMove>(
                         rules = HeartsRules,
-                        ai = HeartsAi,
+                        ai = HeartsAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -582,7 +583,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<YahtzeeState, YahtzeeMove>(
                         rules = YahtzeeRules,
-                        ai = YahtzeeAi,
+                        ai = YahtzeeAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -684,7 +685,7 @@ fun PlayScreen(
                         rules = MorrisRules,
                         // As with chess, the strength is part of the table's own
                         // options, so this does not have to decode anything.
-                        ai = MorrisAi(),
+                        ai = MorrisAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -705,7 +706,7 @@ fun PlayScreen(
                         rules = ChessRules,
                         // Strength comes from the table's own options, so the setup
                         // screen sets it without this having to decode anything.
-                        ai = ChessAi(),
+                        ai = ChessAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,
@@ -724,7 +725,7 @@ fun PlayScreen(
                 val controller = remember(route) {
                     MatchController<CatanState, CatanMove>(
                         rules = CatanRules,
-                        ai = CatanAi(),
+                        ai = CatanAi(AiProfiles(route.config)),
                         config = route.config,
                         scope = scope,
                         role = role,

app/src/test/java/org/prolibertate/games/game/backgammon/BackgammonAiProfileTest.kt

@@ -0,0 +1,43 @@
+package org.prolibertate.games.game.backgammon
+
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+
+class BackgammonAiProfileTest {
+
+    private fun playout(ai: BackgammonAi, seed: Long, steps: Int): List<BackgammonMove> {
+        var state = BackgammonRules.initialState(BackgammonOptions(), seed)
+        val chosen = mutableListOf<BackgammonMove>()
+        repeat(steps) {
+            if (BackgammonRules.isFinished(state)) return chosen
+            val seat = BackgammonRules.currentSeat(state) ?: return chosen
+            val legal = BackgammonRules.legalMoves(state, seat)
+            if (legal.isEmpty()) return chosen
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move", move in legal)
+            chosen += move
+            state = BackgammonRules.applyMove(state, seat, move)
+        }
+        return chosen
+    }
+
+    @Test
+    fun everyLevelAndStyleReturnsLegalMoves() {
+        for (difficulty in listOf(AiDifficulty.EASY, AiDifficulty.MEDIUM, AiDifficulty.HARD, AiDifficulty.MIXED)) {
+            for (seed in 0L..5L) {
+                val moves = playout(BackgammonAi(AiProfiles(difficulty, seed)), seed, 14)
+                assertTrue(moves.isNotEmpty())
+            }
+        }
+    }
+
+    @Test
+    fun defaultProfilePlaysAsBefore() {
+        val plain = playout(BackgammonAi(), 3L, 20)
+        val medium = playout(BackgammonAi(AiProfiles(AiDifficulty.MEDIUM, 77L)), 3L, 20)
+        assertEquals(plain, medium)
+    }
+}

app/src/test/java/org/prolibertate/games/game/catan/CatanAiProfilesTest.kt

@@ -0,0 +1,105 @@
+package org.prolibertate.games.game.catan
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.GameCatalog
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.AiStyle
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+/** Computer seats at every level and style play legal, forward-moving games in each Katan rule set. */
+class CatanAiProfilesTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private fun config(options: CatanOptions, seed: Long) = TableConfig(
+        gameId = GameCatalog.CATAN,
+        seats = (0 until options.playerCount).map {
+            PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = it)
+        },
+        optionsJson = json.encodeToString(options),
+        seed = seed,
+    )
+
+    /** Plays up to [cap] moves, checking every one; returns the last state. */
+    private fun play(options: CatanOptions, seed: Long, profiles: AiProfiles, cap: Int): CatanState {
+        val ai = CatanAi(profiles)
+        var state = CatanRules.initialState(config(options, seed))
+        var moves = 0
+        while (!CatanRules.isFinished(state) && moves++ < cap) {
+            val seat = CatanRules.currentSeat(state)
+                ?: error("seed $seed: nobody on the clock in ${state.phase}; log tail: ${state.log.takeLast(8)}")
+            val legal = CatanRules.legalMoves(state, seat)
+            assertTrue("seed $seed: seat $seat has no legal move in ${state.phase}", legal.isNotEmpty())
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("seed $seed: chose $move outside the legal list in ${state.phase}", move in legal || CatanRules.isLegalMove(state, seat, move))
+            assertTrue(
+                "seed $seed: seat $seat chose the illegal move $move in ${state.phase}",
+                CatanRules.isLegalMove(state, seat, move),
+            )
+            state = CatanRules.applyMove(state, seat, move)
+        }
+        return state
+    }
+
+    /** A table seed on which some seat among the first three has [style] at [difficulty]. */
+    private fun tableSeedFor(difficulty: AiDifficulty, style: AiStyle): Long =
+        (1L..2_000L).first { s ->
+            val p = AiProfiles(difficulty, s)
+            (0 until 3).any { p.of(it).style == style }
+        }
+
+    private fun forEveryProfile(rules: CatanRuleSet, cap: Int) {
+        val options = CatanOptions(
+            playerCount = 3,
+            pointsToWin = when (rules) {
+                CatanRuleSet.BASE -> 10
+                CatanRuleSet.CITIES_AND_KNIGHTS -> 13
+                else -> 12
+            },
+            ruleSet = rules,
+        )
+        for (difficulty in listOf(AiDifficulty.EASY, AiDifficulty.HARD)) {
+            for (style in AiStyle.entries) {
+                val tableSeed = tableSeedFor(difficulty, style)
+                val state = play(options, 40L + style.ordinal, AiProfiles(difficulty, tableSeed), cap)
+                assertTrue(
+                    "$rules $difficulty $style: the game never got past setup",
+                    state.turnNumber > 3 || CatanRules.isFinished(state),
+                )
+            }
+        }
+    }
+
+    @Test
+    fun `easy and hard seats of every style play the base game`() = forEveryProfile(CatanRuleSet.BASE, 1_500)
+
+    @Test
+    fun `easy and hard seats of every style play Cities and Knights`() =
+        forEveryProfile(CatanRuleSet.CITIES_AND_KNIGHTS, 1_500)
+
+    @Test
+    fun `easy and hard seats of every style play Seafarers`() = forEveryProfile(CatanRuleSet.SEAFARERS, 1_500)
+
+    @Test
+    fun `a mixed table of four finishes a base game`() {
+        val state = play(CatanOptions(playerCount = 4), 77L, AiProfiles(AiDifficulty.MIXED, 5L), 12_000)
+        assertEquals(CatanPhase.GAME_OVER, state.phase)
+    }
+
+    @Test
+    fun `the default profile plays exactly as an explicit medium table does`() {
+        for (rules in listOf(CatanRuleSet.BASE, CatanRuleSet.CITIES_AND_KNIGHTS)) {
+            val options = CatanOptions(playerCount = 3, pointsToWin = 10, ruleSet = rules)
+            val plain = play(options, 9L, AiProfiles.DEFAULT, 600)
+            val medium = play(options, 9L, AiProfiles(AiDifficulty.MEDIUM, 1234L), 600)
+            assertEquals("$rules", plain, medium)
+        }
+    }
+}

app/src/test/java/org/prolibertate/games/game/checkers/CheckersAiProfileTest.kt

@@ -0,0 +1,40 @@
+package org.prolibertate.games.game.checkers
+
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+
+class CheckersAiProfileTest {
+
+    @Test
+    fun `every difficulty and seeded style plays only legal moves`() {
+        val seen = mutableSetOf<Pair<Any, Any>>()
+        for (difficulty in AiDifficulty.entries) for (seed in 0L until 12L) {
+            val profiles = AiProfiles(difficulty, seed)
+            for (seat in 0..1) profiles.of(seat).let { seen += it.level to it.style }
+            val ai = CheckersAi(profiles, CheckersLevel.CASUAL)
+            var state = CheckersRules.initialState(CheckersOptions(level = CheckersLevel.CASUAL))
+            var plies = 0
+            while (!CheckersRules.isFinished(state) && plies < 10) {
+                val legal = CheckersRules.legalMoves(state)
+                val move = ai.chooseMove(state, state.turn, legal)
+                assertTrue("$difficulty/$seed illegal move $move", move in legal)
+                state = CheckersRules.applyMove(state, state.turn, move)
+                plies++
+            }
+        }
+        assertTrue("expected a spread of level/style pairs, saw ${seen.size}", seen.size >= 8)
+    }
+
+    @Test
+    fun `default profile plays the same as an explicit medium table`() {
+        val state = CheckersRules.initialState(CheckersOptions(level = CheckersLevel.CASUAL))
+        val legal = CheckersRules.legalMoves(state)
+        val a = CheckersAi(level = CheckersLevel.CASUAL).chooseMove(state, state.turn, legal)
+        val b = CheckersAi(AiProfiles(AiDifficulty.MEDIUM, 5L), CheckersLevel.CASUAL)
+            .chooseMove(state, state.turn, legal)
+        assertEquals(a, b)
+    }
+}

app/src/test/java/org/prolibertate/games/game/chess/ChessAiTest.kt

@@ -3,6 +3,11 @@ package org.prolibertate.games.game.chess
 import org.junit.Assert.assertEquals
 import org.junit.Assert.assertTrue
 import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiLevel
+import org.prolibertate.games.game.engine.AiProfile
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.AiStyle
 
 class ChessAiTest {
 
@@ -128,4 +133,52 @@ class ChessAiTest {
             assertTrue("$level took ${millis}ms", millis < 15_000)
         }
     }
+
+    @Test
+    fun `every level and style returns a legal move over a few playouts`() {
+        val opening = "r1bqkb1r/pppp1ppp/2n2n2/4p3/2B1P3/5N2/PPPP1PPP/RNBQK2R w KQkq - 4 4"
+        for (diff in listOf(AiDifficulty.EASY, AiDifficulty.MEDIUM, AiDifficulty.HARD, AiDifficulty.MIXED)) {
+            for (seed in 1L..4L) {
+                val ai = ChessAi(AiProfiles(diff, seed))
+                var state = ChessRules.fromPosition(ChessOptions(level = ChessLevel.CASUAL), opening)
+                repeat(6) {
+                    if (ChessRules.isFinished(state)) return@repeat
+                    val legal = ChessRules.legalMoves(state)
+                    val move = ai.moveIn(state)
+                    assertTrue("$diff/$seed: $move illegal", legal.contains(move))
+                    state = ChessRules.applyMove(state, state.turnSeat, move)
+                }
+            }
+        }
+    }
+
+    @Test
+    fun `easy and every style still take a hanging queen`() {
+        val state = position("3r3k/8/8/3q4/8/8/8/3R2K1 w - - 0 1", ChessLevel.CLUB)
+        for (style in AiStyle.entries) {
+            val ai = ChessAi(AiProfilesFixed(AiLevel.HARD, style))
+            assertEquals("d1d5", ai.moveIn(state).toString())
+        }
+    }
+
+    @Test
+    fun `the default profile matches the no-argument ai`() {
+        val state = position("r1bqkb1r/pppp1ppp/2n2n2/4p3/2B1P3/5N2/PPPP1PPP/RNBQK2R w KQkq - 4 4")
+        assertEquals(ChessAi().moveIn(state), ChessAi(AiProfiles.DEFAULT).moveIn(state))
+    }
+
+    /** Profiles with one fixed level and style for every seat. */
+    private fun AiProfilesFixed(level: AiLevel, style: AiStyle): AiProfiles {
+        val diff = when (level) {
+            AiLevel.EASY -> AiDifficulty.EASY
+            AiLevel.MEDIUM -> AiDifficulty.MEDIUM
+            AiLevel.HARD -> AiDifficulty.HARD
+        }
+        // Search seeds until the derived style matches; the style space is tiny.
+        for (seed in 0L..500L) {
+            val profiles = AiProfiles(diff, seed)
+            if (profiles.of(0).style == style && profiles.of(1).style == style) return profiles
+        }
+        return AiProfiles(diff, 0L)
+    }
 }

app/src/test/java/org/prolibertate/games/game/crazyeights/CrazyEightsAiProfileTest.kt

@@ -0,0 +1,71 @@
+package org.prolibertate.games.game.crazyeights
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+class CrazyEightsAiProfileTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private fun config(options: CrazyEightsOptions, seed: Long) = TableConfig(
+        gameId = "crazy8s",
+        seats = (0 until options.playerCount).map {
+            PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = it)
+        },
+        optionsJson = json.encodeToString(options),
+        seed = seed,
+    )
+
+    private fun play(ai: CrazyEightsAi, options: CrazyEightsOptions, seed: Long): List<CrazyEightsMove> {
+        var state = CrazyEightsRules.initialState(config(options, seed))
+        val moves = mutableListOf<CrazyEightsMove>()
+        var guard = 0
+        while (!CrazyEightsRules.isFinished(state) && guard++ < 6000) {
+            val seat = CrazyEightsRules.currentSeat(state)
+            if (seat == null) {
+                state = CrazyEightsRules.nextRound(state)
+                continue
+            }
+            val legal = CrazyEightsRules.legalMoves(state, seat)
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move", legal.contains(move))
+            moves += move
+            state = CrazyEightsRules.applyMove(state, seat, move)
+        }
+        assertTrue("game did not finish", CrazyEightsRules.isFinished(state))
+        return moves
+    }
+
+    @Test
+    fun `every difficulty and style plays complete games legally`() {
+        val variants = listOf(
+            CrazyEightsOptions(playerCount = 4, roundsToPlay = 2),
+            CrazyEightsOptions(playerCount = 2, roundsToPlay = 2, drawUntilPlayable = false),
+            CrazyEightsOptions(playerCount = 6, roundsToPlay = 2),
+        )
+        for (difficulty in listOf(AiDifficulty.EASY, AiDifficulty.HARD, AiDifficulty.MIXED)) {
+            for (seed in 0L until 9L) {
+                val options = variants[(seed % variants.size).toInt()]
+                play(CrazyEightsAi(AiProfiles(difficulty, seed)), options, 300L + seed)
+            }
+        }
+    }
+
+    @Test
+    fun `the default profile plays like a medium balanced seat`() {
+        val options = CrazyEightsOptions(playerCount = 4, roundsToPlay = 2)
+        for (seed in 1L..3L) {
+            val default = play(CrazyEightsAi(), options, seed)
+            val medium = play(CrazyEightsAi(AiProfiles(AiDifficulty.MEDIUM, 31L + seed)), options, seed)
+            assertEquals(default, medium)
+        }
+    }
+}

app/src/test/java/org/prolibertate/games/game/cribbage/CribbageAiLevelsTest.kt

@@ -0,0 +1,101 @@
+package org.prolibertate.games.game.cribbage
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.AiStyle
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+class CribbageAiLevelsTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private fun config(options: CribbageOptions, seed: Long) = TableConfig(
+        gameId = "cribbage",
+        seats = (0 until options.playerCount).map {
+            PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = options.teamOf(it))
+        },
+        optionsJson = json.encodeToString(options),
+        seed = seed,
+    )
+
+    /** Seeds that between them give some seat every style at this difficulty. */
+    private fun seedsCoveringStyles(difficulty: AiDifficulty, seats: Int): List<Long> {
+        // The standard opponent is balanced by definition, so there is no style to cover.
+        if (difficulty == AiDifficulty.MEDIUM) return listOf(0L, 1L, 2L)
+        val missing = AiStyle.entries.toMutableSet()
+        val seeds = mutableListOf<Long>()
+        var seed = 0L
+        while (missing.isNotEmpty() && seed < 500) {
+            val styles = (0 until seats).map { AiProfiles(difficulty, seed).of(it).style }
+            if (styles.any { it in missing }) {
+                seeds += seed
+                missing.removeAll(styles.toSet())
+            }
+            seed++
+        }
+        assertTrue("styles never covered: $missing", missing.isEmpty())
+        return seeds
+    }
+
+    /** Plays a whole game with [ai] in every seat; returns the moves chosen. */
+    private fun play(ai: CribbageAi, options: CribbageOptions, seed: Long): List<CribbageMove> {
+        var state = CribbageRules.initialState(config(options, seed))
+        val moves = mutableListOf<CribbageMove>()
+        var guard = 0
+        while (!CribbageRules.isFinished(state) && guard++ < 20000) {
+            val seat = CribbageRules.currentSeat(state)
+            if (seat == null) {
+                state = CribbageRules.nextHand(state)
+                continue
+            }
+            val legal = CribbageRules.legalMoves(state, seat)
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move", legal.contains(move))
+            moves += move
+            state = CribbageRules.applyMove(state, seat, move)
+        }
+        assertTrue("game ran away", guard < 20000)
+        return moves
+    }
+
+    @Test
+    fun `every level and style plays a whole game legally`() {
+        val options = CribbageOptions(pointsToWin = 61)
+        listOf(AiDifficulty.EASY, AiDifficulty.MEDIUM, AiDifficulty.HARD).forEach { difficulty ->
+            seedsCoveringStyles(difficulty, options.playerCount).forEach { seed ->
+                play(CribbageAi(AiProfiles(difficulty, seed)), options, 100L + seed)
+            }
+        }
+    }
+
+    @Test
+    fun `a four handed table plays legally at every level`() {
+        val options = CribbageOptions(playerCount = 4, pointsToWin = 61)
+        listOf(AiDifficulty.EASY, AiDifficulty.HARD, AiDifficulty.MIXED).forEach { difficulty ->
+            play(CribbageAi(AiProfiles(difficulty, 3L)), options, 7L)
+        }
+    }
+
+    @Test
+    fun `the default profile plays as the standard opponent`() {
+        val options = CribbageOptions(pointsToWin = 61)
+        val standard = play(CribbageAi(), options, 55L)
+        val medium = play(CribbageAi(AiProfiles(AiDifficulty.MEDIUM, 9L)), options, 55L)
+        assertEquals(standard, medium)
+    }
+
+    @Test
+    fun `easy does not always play the best card`() {
+        val options = CribbageOptions(pointsToWin = 61)
+        val standard = play(CribbageAi(), options, 55L)
+        val easy = play(CribbageAi(AiProfiles(AiDifficulty.EASY, 9L)), options, 55L)
+        assertTrue("easy played exactly like medium", standard != easy)
+    }
+}

app/src/test/java/org/prolibertate/games/game/engine/AiProfilesTest.kt

@@ -0,0 +1,77 @@
+package org.prolibertate.games.game.engine
+
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertFalse
+import org.junit.Assert.assertTrue
+import org.junit.Test
+
+class AiProfilesTest {
+
+    @Test
+    fun `fixed difficulties give every seat that level`() {
+        for ((difficulty, level) in listOf(
+            AiDifficulty.EASY to AiLevel.EASY,
+            AiDifficulty.MEDIUM to AiLevel.MEDIUM,
+            AiDifficulty.HARD to AiLevel.HARD,
+        )) {
+            val profiles = AiProfiles(difficulty, seed = 42)
+            for (seat in 0..5) assertEquals(level, profiles.of(seat).level)
+        }
+    }
+
+    @Test
+    fun `medium is the balanced standard opponent and never slips`() {
+        val profiles = AiProfiles(AiDifficulty.MEDIUM, seed = 7)
+        for (seat in 0..5) {
+            val profile = profiles.of(seat)
+            assertEquals(AiStyle.BALANCED, profile.style)
+            repeat(200) { assertFalse(profile.slips()) }
+        }
+    }
+
+    @Test
+    fun `the same table always gives a seat the same personality`() {
+        val a = AiProfiles(AiDifficulty.MIXED, seed = 1234)
+        val b = AiProfiles(AiDifficulty.MIXED, seed = 1234)
+        for (seat in 0..6) {
+            assertEquals(a.of(seat).level, b.of(seat).level)
+            assertEquals(a.of(seat).style, b.of(seat).style)
+        }
+    }
+
+    @Test
+    fun `mixed tables vary by seat and across seeds`() {
+        val seen = (0L..200L).flatMap { seed ->
+            val profiles = AiProfiles(AiDifficulty.MIXED, seed)
+            (0..3).map { profiles.of(it).level to profiles.of(it).style }
+        }.toSet()
+        assertTrue(seen.map { it.first }.toSet() == AiLevel.entries.toSet())
+        assertTrue(seen.map { it.second }.toSet() == AiStyle.entries.toSet())
+    }
+
+    @Test
+    fun `only easy slips, and not every time`() {
+        val easy = AiProfiles(AiDifficulty.EASY, 1).of(0)
+        val slips = (1..1000).count { easy.slips() }
+        assertTrue("slipped $slips of 1000", slips in 200..400)
+        val hard = AiProfiles(AiDifficulty.HARD, 1).of(0)
+        repeat(200) { assertFalse(hard.slips()) }
+    }
+
+    @Test
+    fun `pick takes the best unless easy slips and then stays in the better half`() {
+        val ranked = (1..10).toList()
+        assertEquals(1, AiProfiles(AiDifficulty.HARD, 3).of(0).pick(ranked))
+        val easy = AiProfiles(AiDifficulty.EASY, 3).of(0)
+        val picks = (1..500).map { easy.pick(ranked) }
+        assertTrue(picks.all { it <= 5 })
+        assertTrue(picks.toSet().size > 1)
+    }
+
+    @Test
+    fun `a config without a difficulty reads as medium`() {
+        val config = TableConfig("chess", emptyList(), "{}", 1L)
+        assertEquals(AiDifficulty.MEDIUM, config.aiDifficulty)
+        assertEquals(AiLevel.MEDIUM, AiProfiles(config).of(0).level)
+    }
+}

app/src/test/java/org/prolibertate/games/game/euchre/EuchreAiProfilesTest.kt

@@ -0,0 +1,65 @@
+package org.prolibertate.games.game.euchre
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+class EuchreAiProfilesTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private fun config(seed: Long) = TableConfig(
+        gameId = "euchre",
+        seats = (0 until 4).map {
+            PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = it % 2)
+        },
+        optionsJson = json.encodeToString(EuchreOptions(pointsToWin = 4)),
+        seed = seed,
+    )
+
+    /** Plays a whole game with every seat on [ai], returning every move chosen. */
+    private fun playGame(ai: EuchreAi, seed: Long): List<EuchreMove> {
+        val moves = mutableListOf<EuchreMove>()
+        var state = EuchreRules.initialState(config(seed))
+        var guard = 0
+        while (!EuchreRules.isFinished(state) && guard++ < 3000) {
+            val seat = EuchreRules.currentSeat(state)
+            if (seat == null) {
+                state = EuchreRules.nextHand(state)
+                continue
+            }
+            val legal = EuchreRules.legalMoves(state, seat)
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move in ${state.phase}", legal.contains(move))
+            moves += move
+            state = EuchreRules.applyMove(state, seat, move)
+        }
+        assertTrue("game with seed $seed did not finish", EuchreRules.isFinished(state))
+        return moves
+    }
+
+    @Test
+    fun `every difficulty and style plays whole games legally`() {
+        for (difficulty in AiDifficulty.entries) {
+            for (seed in 1L..8L) {
+                playGame(EuchreAi(AiProfiles(difficulty, seed)), seed)
+            }
+        }
+    }
+
+    @Test
+    fun `the default profile plays as a medium balanced seat`() {
+        for (seed in 1L..3L) {
+            val standard = playGame(EuchreAi(), seed)
+            val explicit = playGame(EuchreAi(AiProfiles(AiDifficulty.MEDIUM, 99L)), seed)
+            assertEquals(standard, explicit)
+        }
+    }
+}

app/src/test/java/org/prolibertate/games/game/golf/GolfAiProfileTest.kt

@@ -0,0 +1,72 @@
+package org.prolibertate.games.game.golf
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+class GolfAiProfileTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private fun config(options: GolfOptions, seed: Long) = TableConfig(
+        gameId = "golf",
+        seats = (0 until options.playerCount).map {
+            PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = it)
+        },
+        optionsJson = json.encodeToString(options),
+        seed = seed,
+    )
+
+    private fun play(ai: GolfAi, options: GolfOptions, seed: Long): List<GolfMove> {
+        var state = GolfRules.initialState(config(options, seed))
+        val moves = mutableListOf<GolfMove>()
+        var guard = 0
+        while (!GolfRules.isFinished(state) && guard++ < 8000) {
+            val seat = GolfRules.currentSeat(state)
+            if (seat == null) {
+                state = GolfRules.nextHole(state)
+                continue
+            }
+            val legal = GolfRules.legalMoves(state, seat)
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move", legal.contains(move))
+            moves += move
+            state = GolfRules.applyMove(state, seat, move)
+        }
+        assertTrue("game did not finish", GolfRules.isFinished(state))
+        return moves
+    }
+
+    @Test
+    fun `every difficulty and style plays complete games legally`() {
+        val variants = listOf(
+            GolfOptions(playerCount = 2, gridSize = 6, holes = 2),
+            GolfOptions(playerCount = 4, gridSize = 9, holes = 2),
+            GolfOptions(playerCount = 3, gridSize = 4, holes = 2, lineUpFinalPutt = true),
+            GolfOptions(playerCount = 6, gridSize = 8, holes = 1),
+        )
+        for (difficulty in listOf(AiDifficulty.EASY, AiDifficulty.HARD, AiDifficulty.MIXED)) {
+            for (seed in 0L until 8L) {
+                val options = variants[(seed % variants.size).toInt()]
+                play(GolfAi(AiProfiles(difficulty, seed)), options, 400L + seed)
+            }
+        }
+    }
+
+    @Test
+    fun `the default profile plays like a medium balanced seat`() {
+        val options = GolfOptions(playerCount = 3, gridSize = 6, holes = 2)
+        for (seed in 1L..3L) {
+            val default = play(GolfAi(), options, seed)
+            val medium = play(GolfAi(AiProfiles(AiDifficulty.MEDIUM, 13L + seed)), options, seed)
+            assertEquals(default, medium)
+        }
+    }
+}

app/src/test/java/org/prolibertate/games/game/hearts/HeartsAiProfileTest.kt

@@ -0,0 +1,63 @@
+package org.prolibertate.games.game.hearts
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+class HeartsAiProfileTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private fun config(seed: Long) = TableConfig(
+        gameId = "hearts",
+        seats = (0 until 4).map { PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = it) },
+        optionsJson = json.encodeToString(HeartsOptions(targetScore = 50)),
+        seed = seed,
+    )
+
+    /** Plays a whole game and returns every move made, asserting each was legal. */
+    private fun play(ai: HeartsAi, seed: Long): List<HeartsMove> {
+        var state = HeartsRules.initialState(config(seed))
+        val moves = mutableListOf<HeartsMove>()
+        var guard = 0
+        while (!HeartsRules.isFinished(state) && guard++ < 5000) {
+            val seat = HeartsRules.currentSeat(state)
+            if (seat == null) {
+                state = HeartsRules.startNextRound(state)
+                continue
+            }
+            val legal = HeartsRules.legalMoves(state, seat)
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move", legal.contains(move))
+            moves += move
+            state = HeartsRules.applyMove(state, seat, move)
+        }
+        assertTrue("game did not finish", HeartsRules.isFinished(state))
+        return moves
+    }
+
+    @Test
+    fun `every difficulty and style plays complete games legally`() {
+        for (difficulty in listOf(AiDifficulty.EASY, AiDifficulty.HARD, AiDifficulty.MIXED)) {
+            for (seed in 0L until 8L) {
+                play(HeartsAi(AiProfiles(difficulty, seed)), 100L + seed)
+            }
+        }
+    }
+
+    @Test
+    fun `the default profile plays like a medium balanced seat`() {
+        for (seed in 1L..3L) {
+            val default = play(HeartsAi(), seed)
+            val medium = play(HeartsAi(AiProfiles(AiDifficulty.MEDIUM, 77L + seed)), seed)
+            assertEquals(default, medium)
+        }
+    }
+}

app/src/test/java/org/prolibertate/games/game/kaiser/KaiserAiProfilesTest.kt

@@ -0,0 +1,68 @@
+package org.prolibertate.games.game.kaiser
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+class KaiserAiProfilesTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private fun config(seed: Long) = TableConfig(
+        gameId = "kaiser",
+        seats = (0 until 4).map {
+            PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = it % 2)
+        },
+        optionsJson = json.encodeToString(KaiserOptions()),
+        seed = seed,
+    )
+
+    /** Plays up to [hands] hands with every seat on [ai], returning every move chosen. */
+    private fun playHands(ai: KaiserAi, seed: Long, hands: Int): List<KaiserMove> {
+        val moves = mutableListOf<KaiserMove>()
+        var state = KaiserRules.initialState(config(seed))
+        var dealt = 1
+        var guard = 0
+        while (!KaiserRules.isFinished(state) && guard++ < 4000) {
+            val seat = KaiserRules.currentSeat(state)
+            if (seat == null) {
+                if (dealt >= hands) break
+                dealt++
+                state = KaiserRules.nextHand(state)
+                continue
+            }
+            val legal = KaiserRules.legalMoves(state, seat)
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move in ${state.phase}", legal.contains(move))
+            moves += move
+            state = KaiserRules.applyMove(state, seat, move)
+        }
+        return moves
+    }
+
+    @Test
+    fun `every difficulty and style plays hands legally`() {
+        for (difficulty in AiDifficulty.entries) {
+            for (seed in 1L..8L) {
+                val moves = playHands(KaiserAi(AiProfiles(difficulty, seed)), seed, hands = 4)
+                assertTrue(moves.isNotEmpty())
+            }
+        }
+    }
+
+    @Test
+    fun `the default profile plays as a medium balanced seat`() {
+        for (seed in 1L..3L) {
+            val standard = playHands(KaiserAi(), seed, hands = 3)
+            val explicit = playHands(KaiserAi(AiProfiles(AiDifficulty.MEDIUM, 99L)), seed, hands = 3)
+            assertEquals(standard, explicit)
+        }
+    }
+}

app/src/test/java/org/prolibertate/games/game/morris/MorrisAiProfileTest.kt

@@ -0,0 +1,40 @@
+package org.prolibertate.games.game.morris
+
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+
+class MorrisAiProfileTest {
+
+    @Test
+    fun `every difficulty and seeded style plays only legal moves`() {
+        val seen = mutableSetOf<Pair<Any, Any>>()
+        for (difficulty in AiDifficulty.entries) for (seed in 0L until 12L) {
+            val profiles = AiProfiles(difficulty, seed)
+            for (seat in 0..1) profiles.of(seat).let { seen += it.level to it.style }
+            val ai = MorrisAi(profiles, MorrisLevel.CASUAL)
+            var state = MorrisRules.initialState(MorrisOptions(level = MorrisLevel.CASUAL))
+            var plies = 0
+            while (!MorrisRules.isFinished(state) && plies < 8) {
+                val legal = MorrisRules.legalMoves(state)
+                val move = ai.chooseMove(state, state.turn, legal)
+                assertTrue("$difficulty/$seed illegal move $move", move in legal)
+                state = MorrisRules.applyMove(state, state.turn, move)
+                plies++
+            }
+        }
+        assertTrue("expected a spread of level/style pairs, saw ${seen.size}", seen.size >= 8)
+    }
+
+    @Test
+    fun `default profile plays the same as an explicit medium table`() {
+        val state = MorrisRules.initialState(MorrisOptions(level = MorrisLevel.CASUAL))
+        val legal = MorrisRules.legalMoves(state)
+        val a = MorrisAi(level = MorrisLevel.CASUAL).chooseMove(state, state.turn, legal)
+        val b = MorrisAi(AiProfiles(AiDifficulty.MEDIUM, 5L), MorrisLevel.CASUAL)
+            .chooseMove(state, state.turn, legal)
+        assertEquals(a, b)
+    }
+}

app/src/test/java/org/prolibertate/games/game/morris/MorrisAiTest.kt

@@ -29,7 +29,7 @@ class MorrisAiTest {
     )
 
     private fun choose(state: MorrisState, level: MorrisLevel = MorrisLevel.CLUB): MorrisMove =
-        MorrisAi(level).chooseMove(state, state.turn, MorrisRules.legalMoves(state))
+        MorrisAi(level = level).chooseMove(state, state.turn, MorrisRules.legalMoves(state))
 
     @Test
     fun `it always returns a move it is allowed to play`() {

app/src/test/java/org/prolibertate/games/game/pirates/PiratesAiProfileTest.kt

@@ -0,0 +1,45 @@
+package org.prolibertate.games.game.pirates
+
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+
+class PiratesAiProfileTest {
+
+    private fun playout(ai: PiratesAi, steps: Int): List<PiratesMove> {
+        var state = PiratesRules.initialState(PiratesOptions(level = PiratesLevel.CASUAL))
+        val chosen = mutableListOf<PiratesMove>()
+        repeat(steps) {
+            if (PiratesRules.isFinished(state)) return chosen
+            val seat = PiratesRules.currentSeat(state) ?: return chosen
+            val legal = PiratesRules.legalMoves(state, seat)
+            if (legal.isEmpty()) return chosen
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move", move in legal)
+            chosen += move
+            state = PiratesRules.applyMove(state, seat, move)
+        }
+        return chosen
+    }
+
+    @Test
+    fun everyLevelAndStyleReturnsLegalMoves() {
+        for (difficulty in listOf(AiDifficulty.EASY, AiDifficulty.MEDIUM, AiDifficulty.MIXED)) {
+            for (seed in 0L..5L) {
+                assertTrue(playout(PiratesAi(AiProfiles(difficulty, seed)), 12).isNotEmpty())
+            }
+        }
+        for (seed in 0L..3L) {
+            assertTrue(playout(PiratesAi(AiProfiles(AiDifficulty.HARD, seed)), 4).isNotEmpty())
+        }
+    }
+
+    @Test
+    fun defaultProfilePlaysAsBefore() {
+        val plain = playout(PiratesAi(), 10)
+        val medium = playout(PiratesAi(AiProfiles(AiDifficulty.MEDIUM, 77L)), 10)
+        assertEquals(plain, medium)
+    }
+}

app/src/test/java/org/prolibertate/games/game/president/PresidentAiProfileTest.kt

@@ -0,0 +1,71 @@
+package org.prolibertate.games.game.president
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+class PresidentAiProfileTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private fun config(options: PresidentOptions, seed: Long) = TableConfig(
+        gameId = "president",
+        seats = (0 until options.playerCount).map {
+            PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = it)
+        },
+        optionsJson = json.encodeToString(options),
+        seed = seed,
+    )
+
+    private fun play(ai: PresidentAi, options: PresidentOptions, seed: Long): List<PresidentMove> {
+        var state = PresidentRules.initialState(config(options, seed))
+        val moves = mutableListOf<PresidentMove>()
+        var guard = 0
+        while (!PresidentRules.isFinished(state) && guard++ < 6000) {
+            val seat = PresidentRules.currentSeat(state)
+            if (seat == null) {
+                state = PresidentRules.nextRound(state)
+                continue
+            }
+            val legal = PresidentRules.legalMoves(state, seat)
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move", legal.contains(move))
+            moves += move
+            state = PresidentRules.applyMove(state, seat, move)
+        }
+        assertTrue("game did not finish", PresidentRules.isFinished(state))
+        return moves
+    }
+
+    @Test
+    fun `every difficulty and style plays complete games legally`() {
+        val variants = listOf(
+            PresidentOptions(playerCount = 4, roundsToPlay = 3),
+            PresidentOptions(playerCount = 5, roundsToPlay = 2, fourOfAKindBomb = true),
+            PresidentOptions(playerCount = 3, roundsToPlay = 2, twosClear = false),
+        )
+        for (difficulty in listOf(AiDifficulty.EASY, AiDifficulty.HARD, AiDifficulty.MIXED)) {
+            for (seed in 0L until 6L) {
+                val options = variants[(seed % variants.size).toInt()]
+                play(PresidentAi(AiProfiles(difficulty, seed)), options, 200L + seed)
+            }
+        }
+    }
+
+    @Test
+    fun `the default profile plays like a medium balanced seat`() {
+        val options = PresidentOptions(playerCount = 4, roundsToPlay = 2)
+        for (seed in 1L..3L) {
+            val default = play(PresidentAi(), options, seed)
+            val medium = play(PresidentAi(AiProfiles(AiDifficulty.MEDIUM, 55L + seed)), options, seed)
+            assertEquals(default, medium)
+        }
+    }
+}

app/src/test/java/org/prolibertate/games/game/sequence/SequenceAiLevelsTest.kt

@@ -0,0 +1,89 @@
+package org.prolibertate.games.game.sequence
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.AiStyle
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+class SequenceAiLevelsTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private fun config(options: SequenceOptions, seed: Long) = TableConfig(
+        gameId = "sequence",
+        seats = (0 until options.playerCount).map {
+            PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = it % options.teamCount)
+        },
+        optionsJson = json.encodeToString(options),
+        seed = seed,
+    )
+
+    /** Seeds that between them give some seat every style at this difficulty. */
+    private fun seedsCoveringStyles(difficulty: AiDifficulty, seats: Int): List<Long> {
+        // The standard opponent is balanced by definition, so there is no style to cover.
+        if (difficulty == AiDifficulty.MEDIUM) return listOf(0L, 1L, 2L)
+        val missing = AiStyle.entries.toMutableSet()
+        val seeds = mutableListOf<Long>()
+        var seed = 0L
+        while (missing.isNotEmpty() && seed < 500) {
+            val styles = (0 until seats).map { AiProfiles(difficulty, seed).of(it).style }
+            if (styles.any { it in missing }) {
+                seeds += seed
+                missing.removeAll(styles.toSet())
+            }
+            seed++
+        }
+        assertTrue("styles never covered: $missing", missing.isEmpty())
+        return seeds
+    }
+
+    private fun play(ai: SequenceAi, options: SequenceOptions, seed: Long): List<SequenceMove> {
+        var state = SequenceRules.initialState(config(options, seed))
+        val moves = mutableListOf<SequenceMove>()
+        var guard = 0
+        while (!SequenceRules.isFinished(state) && guard++ < 4000) {
+            val seat = SequenceRules.currentSeat(state) ?: break
+            val legal = SequenceRules.legalMoves(state, seat)
+            if (legal.isEmpty()) break
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move", legal.contains(move))
+            moves += move
+            state = SequenceRules.applyMove(state, seat, move)
+        }
+        assertTrue("game ran away", guard < 4000)
+        return moves
+    }
+
+    @Test
+    fun `every level and style plays a whole game legally`() {
+        val options = SequenceOptions(teamCount = 2, playersPerTeam = 1)
+        listOf(AiDifficulty.EASY, AiDifficulty.MEDIUM, AiDifficulty.HARD).forEach { difficulty ->
+            seedsCoveringStyles(difficulty, options.playerCount).forEach { seed ->
+                play(SequenceAi(AiProfiles(difficulty, seed)), options, 300L + seed)
+            }
+        }
+    }
+
+    @Test
+    fun `three teams play legally at every level`() {
+        val options = SequenceOptions(teamCount = 3, playersPerTeam = 1, sequencesToWin = 1)
+        listOf(AiDifficulty.EASY, AiDifficulty.HARD, AiDifficulty.MIXED).forEach { difficulty ->
+            play(SequenceAi(AiProfiles(difficulty, 4L)), options, 11L)
+        }
+    }
+
+    @Test
+    fun `the default profile plays as the standard opponent`() {
+        val options = SequenceOptions()
+        val standard = play(SequenceAi(), options, 21L)
+        val medium = play(SequenceAi(AiProfiles(AiDifficulty.MEDIUM, 6L)), options, 21L)
+        assertEquals(standard, medium)
+    }
+}

app/src/test/java/org/prolibertate/games/game/tayu/TayuAiLevelsTest.kt

@@ -0,0 +1,88 @@
+package org.prolibertate.games.game.tayu
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.AiStyle
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+class TayuAiLevelsTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private val options = TayuOptions(tileCopies = 2)
+
+    private fun config(seed: Long) = TableConfig(
+        gameId = "tayu",
+        seats = (0 until options.playerCount).map {
+            PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = it % 2)
+        },
+        optionsJson = json.encodeToString(options),
+        seed = seed,
+    )
+
+    /** Seeds that between them give some seat every style at this difficulty. */
+    private fun seedsCoveringStyles(difficulty: AiDifficulty, seats: Int): List<Long> {
+        // The standard opponent is balanced by definition, so there is no style to cover.
+        if (difficulty == AiDifficulty.MEDIUM) return listOf(0L, 1L, 2L)
+        val missing = AiStyle.entries.toMutableSet()
+        val seeds = mutableListOf<Long>()
+        var seed = 0L
+        while (missing.isNotEmpty() && seed < 500) {
+            val styles = (0 until seats).map { AiProfiles(difficulty, seed).of(it).style }
+            if (styles.any { it in missing }) {
+                seeds += seed
+                missing.removeAll(styles.toSet())
+            }
+            seed++
+        }
+        assertTrue("styles never covered: $missing", missing.isEmpty())
+        return seeds
+    }
+
+    private fun play(ai: TayuAi, seed: Long): List<TayuMove> {
+        var state = TayuRules.initialState(config(seed))
+        val moves = mutableListOf<TayuMove>()
+        var guard = 0
+        while (!TayuRules.isFinished(state) && guard++ < 500) {
+            val seat = TayuRules.currentSeat(state) ?: break
+            val legal = TayuRules.legalMoves(state, seat)
+            if (legal.isEmpty()) break
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move", legal.contains(move))
+            moves += move
+            state = TayuRules.applyMove(state, seat, move)
+        }
+        assertTrue("game ran away", guard < 500)
+        return moves
+    }
+
+    @Test
+    fun `every level and style plays a whole game legally`() {
+        listOf(AiDifficulty.EASY, AiDifficulty.MEDIUM, AiDifficulty.HARD).forEach { difficulty ->
+            seedsCoveringStyles(difficulty, options.playerCount).forEach { seed ->
+                play(TayuAi(AiProfiles(difficulty, seed)), 500L + seed)
+            }
+        }
+    }
+
+    @Test
+    fun `the gentle table setting is respected at every level`() {
+        listOf(AiDifficulty.EASY, AiDifficulty.HARD).forEach { difficulty ->
+            play(TayuAi(AiProfiles(difficulty, 2L), TayuLevel.GENTLE), 9L)
+        }
+    }
+
+    @Test
+    fun `the default profile plays as the standard opponent`() {
+        val standard = play(TayuAi(), 31L)
+        val medium = play(TayuAi(AiProfiles(AiDifficulty.MEDIUM, 8L)), 31L)
+        assertEquals(standard, medium)
+    }
+}

app/src/test/java/org/prolibertate/games/game/wizard/WizardAiProfilesTest.kt

@@ -0,0 +1,71 @@
+package org.prolibertate.games.game.wizard
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+class WizardAiProfilesTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private fun config(seed: Long, options: WizardOptions) = TableConfig(
+        gameId = "wizard",
+        seats = (0 until options.playerCount).map {
+            PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = it)
+        },
+        optionsJson = json.encodeToString(options),
+        seed = seed,
+    )
+
+    /** Plays a whole game with every seat on [ai], returning every move chosen. */
+    private fun playGame(ai: WizardAi, seed: Long, options: WizardOptions): List<WizardMove> {
+        val moves = mutableListOf<WizardMove>()
+        var state = WizardRules.initialState(config(seed, options))
+        var guard = 0
+        while (!WizardRules.isFinished(state) && guard++ < 5000) {
+            val seat = WizardRules.currentSeat(state)
+            if (seat == null) {
+                state = WizardRules.nextRound(state)
+                continue
+            }
+            val legal = WizardRules.legalMoves(state, seat)
+            val move = ai.chooseMove(state, seat, legal)
+            assertTrue("illegal move $move in ${state.phase}", legal.contains(move))
+            moves += move
+            state = WizardRules.applyMove(state, seat, move)
+        }
+        assertTrue("game with seed $seed did not finish", WizardRules.isFinished(state))
+        return moves
+    }
+
+    @Test
+    fun `every difficulty and style plays whole games legally`() {
+        for (difficulty in AiDifficulty.entries) {
+            for (seed in 1L..6L) {
+                val options = WizardOptions(
+                    playerCount = 3 + (seed % 4).toInt(),
+                    rounds = 6,
+                    screwTheDealer = seed % 2L == 0L,
+                )
+                playGame(WizardAi(AiProfiles(difficulty, seed)), seed, options)
+            }
+        }
+    }
+
+    @Test
+    fun `the default profile plays as a medium balanced seat`() {
+        val options = WizardOptions(playerCount = 4, rounds = 6)
+        for (seed in 1L..3L) {
+            val standard = playGame(WizardAi(), seed, options)
+            val explicit = playGame(WizardAi(AiProfiles(AiDifficulty.MEDIUM, 99L)), seed, options)
+            assertEquals(standard, explicit)
+        }
+    }
+}

app/src/test/java/org/prolibertate/games/game/yahtzee/YahtzeeAiLevelsTest.kt

@@ -0,0 +1,95 @@
+package org.prolibertate.games.game.yahtzee
+
+import kotlinx.serialization.encodeToString
+import kotlinx.serialization.json.Json
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import org.prolibertate.games.game.engine.AiDifficulty
+import org.prolibertate.games.game.engine.AiProfiles
+import org.prolibertate.games.game.engine.AiStyle
+import org.prolibertate.games.game.engine.PlayerKind
+import org.prolibertate.games.game.engine.PlayerSlot
+import org.prolibertate.games.game.engine.TableConfig
+
+class YahtzeeAiLevelsTest {
+
+    private val json = Json { encodeDefaults = true }
+
+    private fun config(options: YahtzeeOptions, seed: Long) = TableConfig(
+        gameId = "yahtzee",
+        seats = (0 until options.playerCount).map {
+            PlayerSlot(seat = it, name = "P$it", kind = PlayerKind.AI, team = it)
+        },
+        optionsJson = json.encodeToString(options),
+        seed = seed,
+    )
+
+    /** Seeds that between them give some seat every style at this difficulty. */
+    private fun seedsCoveringStyles(difficulty: AiDifficulty, seats: Int): List<Long> {
+        // The standard opponent is balanced by definition, so there is no style to cover.
+        if (difficulty == AiDifficulty.MEDIUM) return listOf(0L, 1L, 2L)
+        val missing = AiStyle.entries.toMutableSet()
+        val seeds = mutableListOf<Long>()
+        var seed = 0L
+        while (missing.isNotEmpty() && seed < 500) {
+            val styles = (0 until seats).map { AiProfiles(difficulty, seed).of(it).style }
+            if (styles.any { it in missing }) {
+                seeds += seed
+                missing.removeAll(styles.toSet())
+            }
+            seed++
+        }
+        assertTrue("styles never covered: $missing", missing.isEmpty())
+        return seeds
+    }
+
+    /** Plays a card out, returning the final state. */
+    private fun play(ai: YahtzeeAi, options: YahtzeeOptions, seed: Long): YahtzeeState {
+        var state = YahtzeeRules.initialState(config(options, seed))
+        var guard = 0
+        while (!YahtzeeRules.isFinished(state) && guard++ < 1000) {
+            val legal = YahtzeeRules.legalMoves(state, state.turn)
+            val move = ai.chooseMove(state, state.turn, legal)
+            assertTrue("illegal move $move", legal.contains(move))
+            state = YahtzeeRules.applyMove(state, state.turn, move)
+        }
+        assertTrue("game ran away", guard < 1000)
+        assertTrue(state.cards.all { it.isComplete })
+        return state
+    }
+
+    @Test
+    fun `every level and style fills a card legally`() {
+        val options = YahtzeeOptions(playerCount = 2)
+        listOf(AiDifficulty.EASY, AiDifficulty.MEDIUM, AiDifficulty.HARD).forEach { difficulty ->
+            seedsCoveringStyles(difficulty, options.playerCount).forEach { seed ->
+                play(YahtzeeAi(AiProfiles(difficulty, seed)), options, 40L + seed)
+            }
+        }
+    }
+
+    @Test
+    fun `the default profile plays as the standard opponent`() {
+        val options = YahtzeeOptions(playerCount = 2)
+        repeat(3) { i ->
+            val standard = play(YahtzeeAi(), options, 70L + i)
+            val medium = play(YahtzeeAi(AiProfiles(AiDifficulty.MEDIUM, 3L + i)), options, 70L + i)
+            assertEquals(standard.cards, medium.cards)
+        }
+    }
+
+    @Test
+    fun `hard scores clearly more than easy over a run of solo games`() {
+        val options = YahtzeeOptions(playerCount = 1)
+        fun average(difficulty: AiDifficulty): Double =
+            (1..12).map { i ->
+                play(YahtzeeAi(AiProfiles(difficulty, i.toLong())), options, 900L + i).totalFor(0)
+            }.average()
+
+        val easy = average(AiDifficulty.EASY)
+        val hard = average(AiDifficulty.HARD)
+        println("yahtzee solo average: easy $easy, hard $hard")
+        assertTrue("hard $hard did not beat easy $easy", hard > easy + 20)
+    }
+}

app/src/test/java/org/prolibertate/games/game/yahtzee/YahtzeeRulesTest.kt

@@ -272,7 +272,7 @@ class YahtzeeRulesTest {
         var guard = 200
         while (!YahtzeeRules.isFinished(state) && guard-- > 0) {
             val legal = YahtzeeRules.legalMoves(state, state.turn)
-            state = YahtzeeRules.applyMove(state, state.turn, YahtzeeAi.chooseMove(state, state.turn, legal))
+            state = YahtzeeRules.applyMove(state, state.turn, YahtzeeAi().chooseMove(state, state.turn, legal))
         }
         assertTrue("the game should end on its own", guard > 0)
         assertTrue(YahtzeeRules.isFinished(state))
@@ -288,7 +288,7 @@ class YahtzeeRulesTest {
         var guard = 400
         while (!YahtzeeRules.isFinished(state) && guard-- > 0) {
             val legal = YahtzeeRules.legalMoves(state, state.turn)
-            state = YahtzeeRules.applyMove(state, state.turn, YahtzeeAi.chooseMove(state, state.turn, legal))
+            state = YahtzeeRules.applyMove(state, state.turn, YahtzeeAi().chooseMove(state, state.turn, legal))
         }
         assertTrue(guard > 0)
         assertTrue(state.cards.all { it.isComplete })

app/src/test/java/org/prolibertate/games/net/SaveRoundTripAllGamesTest.kt

@@ -155,9 +155,9 @@ class SaveRoundTripAllGamesTest {
     @Test fun backgammon() = exercise(GameCatalog.BACKGAMMON, BackgammonRules, BackgammonAi(), opts(BackgammonOptions()))
     @Test fun mastermind() = exercise(GameCatalog.MASTERMIND, MastermindRules, MastermindAi(), opts(MastermindOptions()))
     @Test fun pirates() = exercise(GameCatalog.PIRATES, PiratesRules, PiratesAi(), opts(PiratesOptions()))
-    @Test fun hearts() = exercise(GameCatalog.HEARTS, HeartsRules, HeartsAi, opts(HeartsOptions()))
+    @Test fun hearts() = exercise(GameCatalog.HEARTS, HeartsRules, HeartsAi(), opts(HeartsOptions()))
     @Test fun pegSolitaire() = exercise(GameCatalog.PEG_SOLITAIRE, PegSolitaireRules, PegSolitaireAi, opts(PegSolitaireOptions()))
-    @Test fun yahtzee() = exercise(GameCatalog.YAHTZEE, YahtzeeRules, YahtzeeAi, opts(YahtzeeOptions()))
+    @Test fun yahtzee() = exercise(GameCatalog.YAHTZEE, YahtzeeRules, YahtzeeAi(), opts(YahtzeeOptions()))
     @Test fun klondike() = exercise(GameCatalog.KLONDIKE, KlondikeRules, KlondikeAi, opts(KlondikeOptions()))
     @Test fun freecell() = exercise(GameCatalog.FREECELL, FreeCellRules, FirstLegalAi(), opts(FreeCellOptions()))
     @Test fun spider() = exercise(GameCatalog.SPIDER, SpiderRules, FirstLegalAi(), opts(SpiderOptions()))