CrewRift Classic: what our player remembers before it moves or votes

I drafted this with help from an agent, then reviewed it against our current source, the pinned simulator, focused tests, and completed hosted episodes. This post is specifically about CrewRift 0.1.63's crewrift-classic variant: eight players, two imposters, eight tasks per crewmate, an 800-tick kill cooldown, and a long 7,200-tick vote window.

The first important detail is that a connection slot is not a role. Unless a fixture explicitly fixes roles, the simulator shuffles the non-fixed players and chooses imposters from that randomized list. Our player learns whether it is crew or imposter from the actual role observation and uses the slot only to keep its own color and protocol identity straight.

At a high level, the loop looks like this:

read role, visible players, bodies, tasks, chat, ballots, and server tick
if crew: route to unfinished tasks; report bodies; preserve witnessed evidence
if imposter: hunt when cooldown is ready; otherwise prowl without fake role assumptions
in a meeting: combine direct evidence and distinct-speaker support
commit only a legal current-screen vote; keep a deterministic fallback running

Movement and meetings are connected. Crew follows A* routes to task rectangles, but watches actual world displacement rather than assuming a changing key mask means progress. One concrete failure was a concave Reactor corner: an 18-pixel lookahead skipped over the first cardinal turn and repeatedly drove diagonally into a wall. The current route follower limits lookahead to the current cardinal segment, preserves a just-consumed pre-turn segment, and forces the final small cardinal step. If position remains unchanged for 240 ticks, it invalidates the route and tries bounded cardinal escape legs, stopping recovery as soon as real movement resumes.

Crew meeting logic tries to distinguish evidence strength. A body, vent, kill, contradiction, visible ballot, or direct accusation can justify acting quickly. A generic “Blue sus” is weaker: the deterministic path requires support from two distinct non-self speakers before following that kind of claim. Body-location memory is committed once, and stale queued body chat is suppressed after the context changes.

An optional meeting helper sees a compact record of visible players, chat, parsed votes, memory, and the deterministic fallback. It may add short chat or choose a legal target, but it runs asynchronously and never stops the simulation-tick loop. Late responses and targets from an expired meeting are discarded. The current player also has an important failure guard: if the model provider fails, a crewmate keeps the controller's already-computed target instead of reinterpreting a lone accusation. Two earlier crew losses came from exactly that mistake, where the failed helper supplied the decisive ballot against an innocent body reporter.

On the imposter side, the structured imposter icon is the role and cooldown signal. When the cooldown is ready but target parsing finds nobody, the player preserves its hunt direction and alternates a fresh action press and release. That pulse is disabled near the meeting button and visible reportable bodies, where an accidental interaction would reveal or derail the plan.

There are two scoring gotchas. Current result scores are pure episode outcomes: an imposter win is 3, a crew win is 1, and a loss is 0. Tasks, kills, stuck events, and vote timeouts explain how an episode went, but do not add directly to that score. Also, skip and uncast ballots count against an ejection threshold, so one extra player ballot can be exactly decisive.

In the completed eight-episode test for the current lower player, its 16 randomized assignments included 11 crew roles and five imposter roles. It won four crew assignments and four imposter assignments, completed 84 crew tasks, converted nine imposter kills, and averaged 1.0. All five of its crew ballots targeted result-confirmed imposters, with no vote, connection, or disconnection timeouts. Every replay hash and task/kill total was checked.

The next useful improvement is better post-meeting danger memory: keep movement distance from a credible accuser or counterclaimer without turning that clue into an automatic vote. I would require a completed episode where the marker actually fires, then compare role-matched survival and wins rather than treating extra tasks alone as success.

• Which observations do you trust enough to vote on after one meeting?
• How long should a crewmate avoid a plausible threat before returning to its task route?
• What is your best signal that a navigation recovery really escaped rather than merely changed inputs?
• How do you keep a meeting model useful without letting wall-clock latency consume simulated time?