| name | crewborg-aaln-guide |
| description | The policy-specific optimizer guide for crewborg-aaln — Aaron's Crewrift league policy. Read this first when optimizing crewborg-aaln. Covers the cognitive-stack architecture, the variant/experiment env flags, the Crewrift league + Observatory IDs, build/run/test commands, the exact files to edit when improving behavior (suspicion, mode selector, vote policy, modes, navigation), and the trace/debug format. Pair with games/crewrift skills for game mechanics and skills/ for the optimizer methodology. |
crewborg-aaln — Policy Optimizer Guide
You are optimizing crewborg-aaln, Aaron Landy's policy in the Crewrift
league (an Among Us–style social-deduction game: crewmates do tasks and vote,
imposters kill/vent/blend). This is the particulars of this policy. For game
mechanics and the optimizer loop, see the routing at the bottom.
The policy directory is the parent of this folder:
players/crewborg-aaln/ — a self-contained coplayer (vendored
players/crewrift/crewborg/ + players/player_sdk/, plus Dockerfile,
coplayer_manifest.json, requirements.txt, README.md). Verify any constant
or path against source before relying on it — Crewrift is in active development.
League & Observatory identity (collect live before acting)
| Field | Value |
|---|
| League | Crewrift — league_605ff338-0a2e-4e62-aeda-559df9a9198f |
| Policy | crewborg-aaln (current uploaded tag :v2; vote-deadline fix in :v3) |
| Policy ID | 3154885a-3285-4888-8757-92cb053a078d |
| Policy version ID (v2) | 804c2e83-4daa-4cf0-b538-700b5f542c8f |
| Container image (v2) | img_3f2026ab-199f-4528-b03f-5db0f1fe6561 (docker-img) |
| Player | Aaron — ply_630a768f-d623-44b2-80fa-36968d6fa75a |
| Run argv | python -m players.crewrift.crewborg.coworld.policy_player |
Never trust these from memory. Read live standings before selecting eval
targets: coworld player use ply_630a768f-d623-44b2-80fa-36968d6fa75a, then
coworld leagues league_605ff338-0a2e-4e62-aeda-559df9a9198f --json,
coworld memberships --league <id> --active-only --json,
coworld submissions --league <id> --mine --json. Verify the active player is
Aaron before any upload/submit. (See skills/coworld-operations.)
Architecture — the cognitive stack
crewborg plugs Crewrift-specific code into the Player SDK's two-loop runtime.
Each tick (24 Hz) the runtime runs, under one shared-memory write lock:
perceive(obs) → update_belief(belief, percept) → update_event_log → update_suspicion
→ update_agent_tracking → strategy.decide → mode.decide(belief) → resolve_action → wire Command
This is the perception → world state → integration → cognition → action
stack:
- Perception (
perception/, types.py:perceive) — decode the binary
Sprite-v1 scene tables into labelled entities + world coords. No computer
vision; the protocol is structured. Raw pixels never enter belief.
- World state (
types.py:update_belief, Belief) — the persistent world
model: self, map/nav, roster-by-color, perception tape, bodies, chat,
social history, tasks, phase, voting, suspicion. Belief is the only
interface the strategy and modes see.
- Integration (fast loop, after belief):
strategy/event_log.py
(per-player observation intervals), strategy/suspicion.py (Bayesian
P(imposter)), agent_tracking.py (reachability-disc location estimates +
occupancy grid).
- Cognition — two tiers:
- Strategy (
strategy/rule_based.py) = the mode selector: pure rules
over belief → a ModeDirective (which mode is active), run every tick.
- Mode (
modes/) = a behavioral stance → one symbolic Intent per tick.
- Action (
action.py:resolve_action) — stateful executor: Intent → wire
Command. All pathing (nav.py), momentum, button/cursor timing, and chat
buffering live here. Modes never touch buttons or routes.
Invariants: modes emit symbolic intents only; the inner loop never blocks on the
strategy (default directive + TTLs keep it live). Full spec:
players/crewrift/crewborg/design.md (architecture §1, modes §7, strategy §10,
suspicion §10.1, action §9). Orientation: crewborg/AGENTS.md.
Variant / experiment flags (what the upload bakes in, and your knobs)
crewborg behavior is configured only at the strategy boundary
(rule_based.py reads env once at construction); modes are env-free. The
crewborg-aaln:v2 image bakes the deterministic dumb baseline:
| Env flag | v2 value | Effect |
|---|
CREWBORG_POLICY_VARIANT | dumb | scripted baseline, no external LLM required |
BE_DUMB / CREWBORG_BE_DUMB | 1 | aggressive imposter: skip Pretend/Evade/body-reports, stay in Search/Hunt |
CREWBORG_LLM_MEETINGS | 0 | deterministic vote policy (no LLM latency in meetings) |
CREWBORG_DICK_MODE | 0 | no timed emergency-button taunt |
CREWBORG_HUNTER | (unset) | hunter/stakeout fork; the hunter variant is the separate sussybuster-aaln policy |
Other knobs you can flip per candidate (edit the Dockerfile ENV block, then
rebuild/upload — see Build below):
CREWBORG_LLM_MEETINGS=1 + a provider key, or a Bedrock flag → LLM chat/vote
in meetings. Three backends (selected by CREWBORG_LLM_PROVIDER, else
inferred): OpenRouter (OPENROUTER_API_KEY, OpenAI-compatible gateway,
default model anthropic/claude-haiku-4.5 — the preferred path), direct
Anthropic (ANTHROPIC_API_KEY, default claude-haiku-4-5-20251001), or
Bedrock (USE_BEDROCK=1 / CREWBORG_USE_BEDROCK=1 /
CLAUDE_CODE_USE_BEDROCK=1 + AWS env creds). Override the model with
CREWBORG_LLM_MODEL. On the hosted runner, attach the key as a
Secrets-Manager secret at upload:
coworld upload-policy ... --secret-env OPENROUTER_API_KEY=sk-or-... (or
--use-bedrock). The deterministic vote_policy.fallback_vote is the
always-on safety net. See skills/coworld-operations §Secrets.CREWBORG_DICK_MODE=1 — one-shot crewmate emergency-button call before the
first kill cooldown clears.
Flags are the experiment surface. The cleanest first experiments are flag
flips (dumb vs. LLM meetings; aggressive vs. blended imposter), not code
changes. Each combo is a distinct candidate — pin and record which flags each
uploaded version baked in (Observatory attributes.run / the Dockerfile ENV)
and match the set across arms. For each flag's tradeoff and when to test it,
see CREWBORG_INSIGHTS.md §4.
Build / run / test
docker build -t crewborg-aaln:latest players/crewborg-aaln
coworld upload-policy crewborg-aaln:latest --name crewborg-aaln
coworld submit crewborg-aaln:<tag> --league league_605ff338-0a2e-4e62-aeda-559df9a9198f \
--auto-champion always --no-open-browser
COWORLD_PLAYER_WS_URL='ws://127.0.0.1:8080/player?slot=0&token=' \
python -m players.crewrift.crewborg.coworld.policy_player
cd players/crewborg-aaln && python -m pytest players/crewrift/crewborg/tests/
players/crewrift/crewborg/scripts/fetch_episodes.sh -n 100
Verify the run attribute is non-null on the new version after upload — a
missing run is the most common silent -100 timeout failure.
Where to edit when improving behavior
All paths under players/crewborg-aaln/players/crewrift/crewborg/. Map the
hypothesis to the smallest file:
| Want to change… | Edit | Notes |
|---|
| Who we suspect / flee / vote | strategy/suspicion.py | Bayesian P(imposter): prior, likelihood-ratio table, social cues. FLEE_PROBABILITY gates Flee. Full model: docs/designs/suspicion.md. |
| Which mode is active (role/phase rules) | strategy/rule_based.py | the mode selector + all env-flag gating (BE_DUMB, DICK_MODE, LLM meetings). Crewmate & imposter priority orders live here. |
| The meeting vote decision | strategy/meeting/vote_policy.py | state-dependent vote bar, must-eject endgame, anti-split swap, per-role fallback vote. Often the highest-leverage crewmate fix (missed/bad votes are big penalties). |
| Meeting chat / LLM prompts | strategy/meeting/prompts.py, context.py, llm.py, schema.py | role-specialized prompts (the ROLE_STRATEGY tier is the tunable knob); valid_vote_targets enforces never-vote-teammate. |
| A specific behavioral stance | modes/<mode>.py | crewmate: normal, attend_meeting, report_body, flee, dick_mode, idle. imposter: pretend, search, hunt, evade. The intent each emits + its "done" detection. |
| Movement / pathing / stuck recovery | nav.py (graph + routing) and action.py (follower, momentum, replan) | momentum bang-bang + predictive stop; route re-roots every REPLAN_INTERVAL. See skills/map-navigation. |
| Imposter target selection / interception | strategy/trajectory.py, strategy/opportunity.py, modes/hunt.py, modes/search.py | victim pick, trajectory lead, witness/urgency gate. |
| Occupancy / location estimates | agent_tracking.py, strategy/occupancy.py | reachability-disc beliefs + Pretend room scoring. |
| What gets logged (for a hypothesis) | events.py, trace.py, artifact.py | add domain events / artifact fields first if the eval can't distinguish your hypothesis. See skills/data-collection-design, skills/player-artifacts. |
| Tuning constants (no structural change) | the relevant module's constants; design.md §12 indexes them | e.g. EVADE_TICKS, SEARCH_LEAD_TICKS, isolation/witness/urgency, anti-split timing, Pretend room penalty. |
Keep the diff one hypothesis per candidate so the eval result is
attributable (playbooks/optimize-policy.md).
Debug / trace format
Stdout = protocol channel; stderr = logs/traces. Logging defaults into the
per-episode artifact (artifact.py → in-memory SQLite trace.db +
summary.json, zipped and PUT to COWORLD_PLAYER_ARTIFACT_UPLOAD_URL). The
summary.json block is always echoed to stderr (greppable fallback).
The trace.db schema (traces/metrics/positions), the server_tick replay
join key, and the CREWBORG_TRACE verbosity levels are documented canonically in
games/crewrift/skills/crewrift-optimization. Set the level before the episode
and match it across arms. The single most useful forensic record is the
per-meeting domain.suspicion_snapshot (ranked posteriors + each suspect's event
log + the would-be vote and the bar) — it explains a vote after the fact.
Crewrift facts that shape every decision
Scoring (task +1, kill +10, win +100, vote-timeout −10, stuck −1), the
time-based-seed role assignment (no color→role correlation), the −100 lobby
taint, and the high-variance / 40–80-game consequence for evals are the
objective — the canonical statement is in
games/crewrift/skills/crewrift-optimization (game facts) and
crewrift-eval-design (eval sizing). The one rule to internalize: +10/kill and
+100/win dominate task points, so league score weights imposter performance and
wins.
Routing — read in this order
- This guide — the crewborg-aaln particulars (you are here).
../AGENTS.md — the optimizer entry point + workflow.../playbooks/optimize-policy.md — the one-loop procedure to follow.../games/crewrift/skills/crewrift-optimization/SKILL.md and
crewrift-eval-design/SKILL.md — Crewrift commands, schema, eval sizing
(the most relevant skills).../skills/base-optimizer-framework — the methodology, when you need depth.- Then pull focused
../skills/* per the loop step:
hosted-xp-evals + eval-variance-design (run/size evals),
replay-artifact-analysis + opponent-strategy-mining +
replay-variance-miner (analyze), policy-hypothesis-loop +
data-collection-design (hypothesize/instrument), eval-aggregation +
promotion-gate (verdict), spatial-temporal-analysis + map-navigation
(Crewrift's spatial traits), coworld-operations (CLI), continuous-optimizer
../playbooks/defend-leaderboard.md (keep #1 automatically).
