Suspicion & belief modeling (crewrift) — recipes

On-demand recipes (11). Each is a trigger→action heuristic; pull the relevant one when its situation arises. Tier tags and ⚠ (unverified) as in AGENTS.md.

1. Exploit perfect recall; rank evidence by strength and only act when sure

crewrift

An LLM social-deduction player's #1 edge is perfect recall: remember who was where and when, who spoke vs stayed silent (silence is a suspicion signal), construct alibis before needed, and model what each player does and does not know. Flag a player as more suspicious when alone in low-traffic rooms, repeatedly near bodies but never reporting, moving erratically without advancing real tasks, accusing aggressively (aggressive accusers are often imposters deflecting), standing at a task without advancing its completion counter, or contradicting observed reality (a claimed location others disprove, a position far from where they were seen implying venting). But voting out a real crewmate is a huge blow, so rank evidence by strength and only call sus when sure: a directly witnessed kill or vent transition is strong proof; a player next to a body without reporting is suspicious but not certain; two players merely together is weak. Recover kill evidence from the static-body assumption -- the frame a body first appears with a crewmate within ~2x kill range implicates that crewmate. Never treat a weak sighting or another player's 'sus' chat as proof. _{sources: bitworld/among_them/bot-policies/sidecar/prompts/system.md, bitworld/among_them/players/SMART_BOT_GUIDE.md, bitworld/among_them/bot-policies/sidecar/prompts/crewmate.md, bitworld/among_them/bot-policies/sidecar/prompts/imposter.md (+4)}

2. Key the roster by stable COLOR; track absence and dead targets

crewrift

Color is the only identity stable and unique across every namespace (in-world sprites, bodies/corpses, chat icons, vote markers), so keying the roster by color links a live sighting, a corpse, a chat line, and a vote to one player. Identify crewmates by stable sprite pixels plus body tint (outline and visor are stable, body colors vary) and store the last tick each color was seen so you can name a suspect after reporting a body. Note which players were NOT present at a meeting, since absence from a body-discovery meeting is itself evidence. An imposter bot should mark confirmed kill targets dead in belief so it does not re-target when body-detection lags the kill. Role-reveal asymmetry: a crewmate sees ALL players but cannot tell which are imposters, while an imposter sees only the OTHER imposters by color -- and because that teammate color shows only at reveal, an imposter must record it immediately. _{sources: personal_labs/crewrift_lab/crewrift/crewborg/design.md, bitworld/among_them/players/how_to_make_a_bot.md, bitworld/among_them/players/italkalot.md, bitworld/among_them/bot-policies/sidecar/prompts/crewmate.md (+2)}

3. Make proximity/body evidence a decaying accumulator, never a binary stamp

crewrift · ⚠ session-derived, unverified

Never treat 'player was near a body' as a binary flag -- it is ambiguous (killer, reporter, or passer-by). Make it a weighted accumulator scored by recency + proximity + incident count: stronger the closer they were, gated by a cooldown so one lingering scene does not re-count each tick, with raw distance/score passed into any LLM prompt so it can rank suspects (e.g. 'near-body score 6, distance 11' vs 'score 5, distance 16'). Probability-by-distance gradients are verifiable in traces (a close case ~77% at distance 2, far cases ~23-28%); unit-test that closer appearances score higher and that the snapshot exposes them as probabilistic, not proof. A flat 'seen near body at tick N' stamp causes false positives (stale evidence votes someone out) and false negatives (a fresh kill weighted like an old coincidence); once soft evidence becomes vote-worthy the opposite failure appears (noisy crew-on-crew misfires), so treat the action threshold as a double-edged knob watched in the trace. _{sources: codex:019e14a5-e574-7770-953b-17019f2f5e74, codex:019e1591-004e-7040-b768-062bd8e6acd4, opencode:ses_21ffeb5ffffeA1Pk6sdyHDzwQW}

4. Derive a task-alibi from sprite-rect overlap, not flaky task-flash detection

crewrift

Derive a task-alibi signal without new perception: once per frame after the task-icon update, iterate visible crewmates against in-progress task locations and append an alibi when a crewmate is within a small radius (Manhattan ~12 px, or ~28 px of a terminal). Filter self/teammate/unknown colors and dedup per-(color, task) with a ~20-tick cooldown. Crucially, do NOT key the alibi on a task-completion icon FLASH: task icons render ONLY the bot's OWN assigned tasks, so an icon vanishing means the bot finished its own task and cannot attribute an alibi to anyone -- a non-self sprite overlapping a task-station rectangle is the tractable signal. A weaker signal that fires correctly beats a stronger one that fires on noise. _{sources: bitworld/among_them/players/mod_talks/DESIGN.md, bitworld/among_them/players/mod_talks/LLM_SPRINTS.md, bitworld/among_them/players/modulabot/DESIGN.md, bitworld/among_them/players/modulabot/TODO.md (+1)}

5. Drive votes from an evidence-grounded hypothesis with explicit confidence thresholds

crewrift

Drive crewmate voting from an evidence-grounded hypothesis with explicit thresholds rather than binary scores alone: accuse if top-suspect likelihood >= 0.75, vote the top suspect if >= 0.50, else fall back to a rule-based suspect or SKIP; express timestamps as ticks-before-meeting. The weighted FALLBACK vote combines known-imposter role memory, witnessed venting, near-vent appearances, distance-weighted near-body sightings, witnessed kills, visible vote dots, and chat mentions, minus solo-survival trust, voting SKIP when nothing clears the bar. The imposter fallback avoids self and known teammates, prefers a crewmate already accused by chat/votes/body evidence, else SKIP rather than a baseless pile-on. Distinguish probabilistic from hard proof: a repeated near-vent appearance / near-vent score >= 8 / probability >= 60% is actionable voting evidence, but describe it in chat as 'appeared near vent', not confirmed vent proof. _{sources: bitworld/among_them/players/mod_talks/LLM_VOTING.md, personal_cogs/among_them/guided_bot/MEETING_DESIGN.md}

6. Treat chat as credibility-gated input scored by HOW you learned allegiance; LLM-classify it

crewrift · negative result · ⚠ session-derived, unverified

Treat chat claims as low-confidence input gated by a credibility threshold (~0.5), not facts, and score credibility from HOW you learned the sender's allegiance: unknown sender low (~0.3), same-team-via-mechanical-exchange high (~0.85), same-team-via-weaker-signal medium (~0.6), known enemy very low (~0.1), self-contradicting identity claim crushed by ~0.1x. Even non-credible or enemy messages carry signal, so record rather than discard: log action requests on the sender's record even from enemies (the request is intelligence about intent) and route adversary chatter to a separate observed-behavior channel instead of filtering to zero. Model private knowledge as private: mark a witnessed suspicion only in the observing bot's own memory and surface it in the meeting ledger as 'I witnessed X' rather than promoting to global truth, preserving the asymmetric-information dynamic. Do NOT keyword-score chat symbolically: 'I saw yellow away from the body' names a color and 'body' yet is a DEFENSE, so naive keyword suspicion mis-flags alibis -- hand raw mentions to an LLM and ask it to classify each as accusation / defense / alibi / noise. _{sources: codex:019e090d-a88d-76f2-b1f5-cd3e4eda9381, codex:019e1591-004e-7040-b768-062bd8e6acd4, codex:019e1593-c00d-7a32-9dfa-1cf6b16b01a4, claude-code:dfa4502e-c404-40e8-bf86-8d15c72ba5af (+1)}

7. Debounce belief flips with a consecutive-frame counter and enumerate every reset

loop · ⚠ session-derived, unverified

Debounce belief flips with a CONSECUTIVE-frame counter, not a single frame and not a cumulative count: an exclusion/flip fires only after the qualifying condition holds for a fixed number of consecutive frames (~0.5s worth), and the counter resets to 0 the instant contradicting evidence appears AND when the latch is freshly re-asserted. Single-frame inference on noisy screen perception is fragile. Enumerate EVERY reset condition explicitly -- in one case the negative-evidence counter had to reset on positive match, object coming on-screen, no dots detected at all, and localization loss; missing one leaves it creeping toward a false exclusion. Test three cases: decays after threshold, does NOT decay when the exclusion is interrupted, does NOT decay when the contradicting condition is false. Test multi-condition behavioral flags for the FULL conjunction (an 'avoids interaction' flag requiring visible_ticks>200 AND whisper_entries==0 AND stationary_ticks>100 must be verified NOT to fire when only one holds). _{sources: codex:019e03bc-5520-7e72-aa5c-a572ead62b91, codex:019e05ac-239c-7bb0-8e34-1ec371fbabcd, codex:019e067b-643a-7100-abdb-8d6a0c79d242}

8. Design every sticky belief flag with its clear path; make clearing universal and phase-aware

loop · ⚠ session-derived, unverified

Every sticky/latched belief flag must be designed WITH its clear/decay path at the moment it is created. A one-way latch (set freely, cleared only by one narrow opposite event) silently rots: it persists a stale belief long after the justifying evidence is gone while downstream action tiers keep acting on it. Clear pending/transient fields explicitly on the state transition that ends their scope (exiting a whisper), and make clearing rules UNIVERSAL not buried in one branch -- a flag cleared only inside the overworld-update branch goes stale on a Chatroom->Exchange transition that skips it, so prefer one universal rule (in_chatroom = (view == Whisper)). Belief shields must be PHASE-AWARE: during a task station's HOLD phase the icon still renders, so a missing icon is genuine negative evidence and exclusion should fire fast (do NOT shield); during the CONFIRM/success phase you watch the icon disappear as the completion signal, so shielding IS correct -- a blanket 'i != holdIndex' shield means the bot can never discover mid-action that the target was wrong and wastes the 84-tick interaction on stations that are not its own. When the stale-flag consumer is off-limits, fix the belief at its SOURCE (belief/types/test files), not by patching every reader. _{sources: codex:019e04f4-9a40-7411-a0a5-872164eda3f1, codex:019e05ac-239c-7bb0-8e34-1ec371fbabcd, opencode:ses_201713d2dffe7pZ6AcZMDZXvNZ, codex:019e0140-e554-78e3-b66f-80634cc63f5d (+2)}

9. Don't conflate navigation failure with target ownership; make 'wrong commitment' sticky

generic · ⚠ session-derived, unverified

Do NOT mark a failed action as 'target not mine': that conflates two failure modes. A first hold-action that fails to register can mean the player never reached the task rectangle (navigation failure), not that the task belongs to someone else -- use first-attempt failure to trigger a RE-NAVIGATE before concluding the target is invalid; treating navigation failure as ownership evidence corrupts belief. Conversely, once a player legitimately resolves that a target is not actionable, make that knowledge sticky: persist it and permanently prevent re-latching to the same target rather than recomputing and re-forgetting each visit, since persistent negative beliefs prevent oscillation. _{sources: opencode:ses_1fecece4dffe779FmeajgdwhMD}

10. Add a periodic mode to reconcile parsed beliefs against an authoritative game screen

crewrift

Add a proactive 'validate knowledge against mechanical truth' mode: beliefs accumulated by parsing chat/system messages can miss events, so periodically (e.g. once per round) navigate to a screen the game renders authoritatively (an info/status screen listing known players, roles, state), reconcile the parsed belief against it, then return to the prior surface -- time-boxing the detour with a tick budget. Note: Persephone's hostage-exchange renderer (renderExchangeRow / drawRoleSlot) draws TRUE role indicators for all shown sprites UNCONDITIONALLY, unlike the overworld which gates on revealedTo, so it leaks every exchanged player's role -- a likely unintentional information-model bug an agent could exploit but should NOT rely on long-term. _{sources: opencode:ses_1fb2002bcffeK2Ldmjf34wdpM3, players_checkouts/players/users/james/personal_cogs/persephone/GAME_AP}

11. Earlier simpler scalar-decay suspicion scheme (superseded by the no-decay Bayesian model)

crewrift

An earlier, simpler design maintained a quantitative continuous per-player suspicion score updated by discrete events rather than deferring all reasoning to the LLM. One tuned constant set (Among Them sidecar bot): SUSPICION_DECAY=0.995 per tick, +0.3 near a body, +0.2 for being accused, +0.1 for being alone in a low-traffic room, +0.05 for staying quiet, -0.4 when cleared by an ejection. NOTE: the later Bayesian-posterior approach in crewborg argues role is a fixed latent so suspicion should NOT decay over time (see the no-time-decay lesson); this scalar-decay scheme is the earlier, simpler design retained for context. _{sources: bitworld/among_them/bot-policies/sidecar/ (package: bot.py, brain.py,}