How the AI thinks, bluffs, and challenges — and what changes with difficulty
There are six AI players: Alice, Bob, Charlie, Dave, Eve and Mike. They all share the same decision engine but differ in how aggressively they use each signal. Memory persists across rounds for as long as the room is active — the AI remembers who bluffs and who plays honestly.
| Name | Challenge tendency | Bluff tendency | Claim jump | Style |
|---|---|---|---|---|
| Alice | 1.0× | 1.0× | low | Balanced and professional |
| Bob | 0.7× | 0.6× | low | Cautious and honest |
| Charlie | 0.8× | 0.8× | low | Standard, reliable |
| Dave | 1.1× | 1.5× | high | Aggressive bully |
| Eve | 1.4× | 0.7× | low | Cynical skeptic |
| Mike | 1.2× | 1.3× | medium | Aggressive and risky |
Before rolling, the AI evaluates whether to challenge the current claim. It combines multiple mathematical signals into a single probability.
Two sub-signals are combined: claimLie (how many of the 36 possible rolls physically cannot produce this rank) and forcedBluff (how likely the previous player was physically trapped by the floor they faced). For example, a 6-6 claim has a lie probability of 97%. If the floor was 5-5, almost nobody could honestly claim 6-6 — so the forced bluff signal is also very high.
The AI calculates if the current claim will trap the next player. If a claim is so high that the next player has almost no chance of beating it honestly, the AI is more likely to challenge now. (Balanced & Aggressive only)
The longer a round goes without a challenge, the more statistically certain it is that someone has lied. This bonus is added after the roll discount is applied.
If the AI has a high statistical chance to beat the current claim honestly, it discounts its suspicion. Why risk a challenge when you can just roll and play it safe?
The AI becomes more cautious as its lives dwindle. However it no longer ignores obvious bluffs just because it has 1 life — doing so leads to a death spiral where it keeps rolling through a long chain instead of ever acting.
| Lives remaining | Challenge multiplier | Override |
|---|---|---|
| 3+ | ×1.0 — full aggression | — |
| 2 | ×0.65 — cautious | — |
| 1 | ×0.45 — careful but not paralysed | ×0.80 if chain ≥ 4 |
After rolling, if the honest roll beats the floor, the AI usually claims truthfully. If forced to bluff, it selects from a weighted "window" of the lowest valid ranks above the floor.
The window size widens as the floor gets higher. An Aggressive AI picking a rank above 6-6 will be less predictable than when picking above 3-2. Decay weighting within the window means lower ranks (closer to the floor) are preferred.
| Difficulty | Window (Low Floor) | Window (High Floor) | Style |
|---|---|---|---|
| Calm | 65% | 85% | Sloppy / Random |
| Balanced | 35% | 75% | Calculated |
| Aggressive | 15% | 60% | Surgical / Minimal |
Aggressive AIs avoid weak opening claims. If their honest roll is below 4-1, they have a high chance of bluffing up to the 4-x or 5-x tier immediately — maintaining table pressure from the very first claim.
Aggressive AIs track opponent lives. With a 2+ life lead, they may widen their bluff window to apply maximum pressure on an opponent who can't afford to be wrong.
When bluffing, Aggressive AIs prefer non-pair ranks. Since pairs are rare (only 6 out of 36 outcomes), claiming one when bluffing is statistically "loud" and easier to catch. Pair bluffs are penalized by 40% in the weighting.
Dave and Mike have elevated "jump" parameters, meaning they occasionally take bigger leaps in their bluff window — Dave especially tends to skip ahead to higher claims when bluffing, making him harder to read but easier to catch when he overreaches.
Sometimes the AI inflates an honest roll to apply pressure, even when it doesn't need to. This is most effective in 1v1 situations where the pressure lands directly on the opponent.
| Factor | Calm | Balanced | Aggressive |
|---|---|---|---|
| Base rate | 5% | 10% | 18% |
| 2 active players | ×2.5 — pressure lands directly on opponent | ||
| 3 active players | ×1.5 | ||
| Life advantage | ×1.5 when AI has more lives than next player | ||
| Roll strength penalty | strong rolls don't need inflation — rate reduced for top 50% rolls | ||
If the next player is on their last life, the AI gets a ×1.4 bonus to its voluntary bluff rate. Finishing a weakened opponent with pressure is a key aggressive tactic — it forces them into a hard decision with no safety net.
Maxle (2-1) is the highest-stakes decision: a wrong challenge costs 2 lives, believing costs only 1.
The final result is multiplied by ×0.75 if the AI has exactly 3 lives, to stay conservative until safely ahead.
The forced bluff signal is especially powerful here. If the floor was 6-6, only one outcome (Maxle itself) could honestly beat it — so 35 of 36 players would be lying. This is the strongest possible signal and heavily influences the challenge decision.
Every resolved challenge updates the AI's internal ledger for that player. Suspicion starts at zero for unknown players — the AI gives everyone the benefit of the doubt until it has evidence.
| Event | Recorded as | Effect |
|---|---|---|
| Caught bluffing | +1 bluff (×audacity) | Increases suspicion |
| Proven honest | +2 honest | Rapidly clears suspicion |
| Caught in a big bluff | +audacity bluff | Large rank jumps remembered harder |
Suspicion = bluffs ÷ (bluffs + honest). Proven honesty is weighted double because being caught telling the truth is a very strong signal of a reliable player. Memory ramps up over the first 6 observations.
Audacity scales the bluff weight by the rank distance of the bluff — a player who claimed 6-6 when they had 3-2 gets remembered more harshly than someone who bumped from 4-3 to 4-4.