Improve AI difficulty scaling.

internal/ai/ai.go

@@ -149,6 +149,17 @@ func (b *Bot) pick(cands []candidate) candidate {
 	if len(cands) == 1 {
 		return cands[0]
 	}
+	// Below competent play, the bot sometimes ignores its evaluation entirely
+	// and plays a random legal move. Softmax temperature alone can't make a bot
+	// a genuine pushover: it still samples among sensibly-generated candidates
+	// and, on this game's shop/dice variance, that wins ~1 game in 5 even
+	// against expert play. Real, occasional mistakes — buying the wrong pet,
+	// passing with coins in hand, a scrambled battle order — are what let a
+	// competent human beat "easy" almost every time. The rate is zero at and
+	// above competent level, so medium and hard never blunder.
+	if p := blunderProb(b.level); p > 0 && rand.Float64() < p {
+		return cands[rand.IntN(len(cands))]
+	}
 	temp := 0.02 + 0.30*(1-b.level)
 	best := math.Inf(-1)
 	for _, c := range cands {
@@ -170,6 +181,24 @@ func (b *Bot) pick(cands []candidate) candidate {
 	return cands[len(cands)-1]
 }
 
+// competentLevel is the skill (on the [0,1] level scale) at which the bot is
+// meant to be an even match for a competent human: at or above it the bot
+// never throws a game on purpose. Difficulties are calibrated around it —
+// easy well below, medium at it, hard above.
+const competentLevel = 0.6
+
+// blunderProb is the chance pick discards its evaluation and plays a uniformly
+// random legal move. It is zero at competent level and above, and ramps up
+// steeply below, so only the easy tier makes real mistakes. blunderMax is the
+// rate at level 0 (the weakest possible bot).
+func blunderProb(level float64) float64 {
+	const blunderMax = 0.75
+	if level >= competentLevel {
+		return 0
+	}
+	return blunderMax * (competentLevel - level) / competentLevel
+}
+
 // budget returns the rollout counts for this difficulty: how many opponent
 // deck/order guesses to test against, and how many dice-randomized battle
 // simulations to run per guess. Fewer samples means noisier estimates, which