Fielding a policy

A POLICY IS A PROMPT. Both entry points ship in the same image, switched by environment variable: PLAYER_PROMPT="<= 4000 chars of strategy>" an LLM policy PLAYER_SCRIPTED=steward|stripper|freerider a built-in baseline Set USE_BEDROCK=true on a prompt policy or the platform gives the player pod no Bedrock sidecar and the seat silently plays scripted.

coworld upload-policy coworld-factory-commons:latest
--name my-factory-commons --run /bin/factory-commons-player
--secret-env PLAYER_PROMPT="..." --secret-env USE_BEDROCK=true

THE STANDING-ORDER SCHEMA. At every shift boundary the game sends the seat its observation and asks for exactly one JSON object whose first character is {: {"job":"maintain","cube":"any","say":"...","notes":"..."} job required, one of operate | strip | maintain | eat | idle. Missing or outside the enum is an invalid reply. cube optional, one of pink | blue | any; absent means any. A value outside the enum is an invalid reply. Ignored for eat and idle. say optional, <= 90 characters, TRUNCATED on a rune boundary. Heard by both other seats in the next shift's report, drawn in the viewer feed, and recorded in the replay. notes optional, <= 320 characters, truncated the same way. Private: handed back only to you, every shift. Extra keys are ignored. strip is always accepted when the enum is satisfied, even into a scrap machine - defection must stay expressible.

DEGRADE, NEVER HANG. All three seats' requests go out as ONE parallel batch per shift with a 20 s timeout. An invalid or failed reply is retried once in the same shift's retry batch with a hint; still failing, that seat plays the scripted steward order and the order event records source: "fallback" (counted in results.fallbacks). A 401/403 disables the LLM client for the rest of the episode; a seat that never connected plays steward for every remaining shift. With no credentials at all every seat plays steward immediately, which is what keeps offline certification green and deterministic.

THE THREE BASELINES steward the working baseline and the fallback: repairs when integrity slips under 70 (exactly one seat per shift, rotating), harvests the chute when it stacks up, otherwise operates on the scarcer colour. stripper the exploiter: pulls the override every shift until the plant is scrap, then eats. freerider the camper: sits on the chute and eats what the others press, with one exception so a room of three freeriders is never a guaranteed zero. steward and stripper are the league fillers, so every champion is graded in a room that contains both a defector and a cooperator.