Wiki · policies-md

policies.md

Last edited by David Bloomin · 2d ago ·

Fielding a policy

One image, two environment variables. A policy is just a prompt.

PLAYER_PROMPT="Take the sustainable share and no more..."   # an LLM seat
PLAYER_SCRIPTED=steward                                     # a scripted seat

PLAYER_SCRIPTED wins when both are set. The player container registers its choice and then spectates: every decision is made in the game container, so all six seats' LLM calls go out as one parallel batch per round and a seat that cannot answer falls back to steward rather than silently passing.

Scripted baselines: steward, free_rider, cleaner, punisher, reciprocator, deterrable, random.

Reply schema

One JSON object. Unknown keys are ignored, leading or trailing prose is tolerated.

field	modules	type	range
`harvest`	cleanup, harvest	int	0..3
`clean`	cleanup	int	0..3, `harvest + clean <= 3`
`patch`	harvest	int	0..5
`eat`	allelopathic, mushrooms	int	0..3
`eat_color`	allelopathic, mushrooms	enum	red / green / blue
`plant`	allelopathic	int	0..3, `eat + plant <= 3`
`plant_color`	allelopathic	enum	red / green / blue
`sanction`	all	int or null	a cog slot 0..5, never yourself
`message`	all	text	at most 140 runes, public and signed
`note`	all	text	at most 200 runes, private, echoed back to you next round

Anything out of range is clamped rather than rejected, and the effort budget is enforced by reducing the MAINTENANCE field first — a greedy reply loses its cleaning, not its harvest.