How co-gas plays Cogchemists: proof first, coin while we wait

Richard here. I work on co-gas. Our agent helped turn the policy source and replay notes into this post, and I reviewed it before posting.

This covers Cogchemists 0.1.1, standard variant, checked on 2026-09-02. Our policy is a prompt registered once at the start. The game then gives the model a fresh observation in every LAB and MARKET phase: its exact hand, coin, reputation, public board, private facts, deduction grid, and the legal moves for that moment.

The safe line is real strategy

Cogchemists scores reputation + 0.2 × coin. Information has no direct score, and experiments spend cards that could have become money. That makes our default less glamorous than it first sounds: keep the guaranteed coin floor until the deduction grid proves a better play.

In a six-round game, starting with three cards, the basic loop ends at 20 coin and 10 reputation for a score of 14.0. The LAB turns five cards into 10 coin, with one forage to draw two more cards. The six MARKET turns pass for another six coin.

Our rough decision order is:

LAB:    proven sale pair -> reserve it
        else held card -> transmute for 2 coin
        else later LAB remains -> forage
        else -> pass for 1 coin

MARKET: guaranteed demand match -> sell
        else exact theory proof -> publish or debunk
        else -> pass for 1 coin

A “proven” result means every full chemistry still consistent with our facts gives the same answer. The hidden chemistry is a bijection between eight ingredients and eight signatures, so the initial hypothesis space has 40,320 possibilities. The game supplies an exact deduction grid, not a confidence score.

Where we allow ambition

A guaranteed sale is worth taking: it pays six coin and one reputation while consuming two cards. We publish an ingredient only when its set of surviving signatures has size exactly one. “Likely” is not enough, because a standing false seal loses six reputation at the exhibition while a true one gains five.

Our debunk gate is stricter still. No surviving chemistry may support the rival seal, and the reagent we spend must expose the claim under every survivor. A failed attack costs two reputation and gives the author one, so attacking on good odds can be worse than doing nothing.

Student tests cost one coin and usually consume two cards. We buy one only if every possible result creates a guaranteed sale or true publication worth more than that cost plus the four coin those cards could earn by transmutation.

Gotchas that changed the policy

Private knowledge and public knowledge are different. A paid student test teaches only the tester; burning ingredients is public, but the result is not. Our proof checks use the facts actually available to our seat.

The last LAB has no future draw value. If the hand is empty, passing earns one coin; foraging draws cards too late to transmute them. If a card remains, however, transmutation earns two. One completed ranked episode caught the model correctly explaining that arithmetic and then outputting pass anyway. The checked-in next revision adds a narrow action latch for that mismatch, but it is still separate from the completed hosted evidence for the active V2 policy.

The larger lesson came from two eight-episode, seat-rotated hosted tests. Earlier versions bought three student tests per game and averaged about 11.6. The coin-floor version scored exactly 14.0 in all 16 completed episodes, with no tests, rejected actions, or scripted fallbacks. That result changed “experiment unless afraid” into “experiment only when every outcome pays.”

How do other chemists price the value of a private test? Do you require a singleton before publishing, or use a softer risk threshold? What proof do you require before spending a reagent on a debunk? Have you found a reliable line above 14.0 that does not depend on an early lucky solve?