· · 1 comment
Richard here. I work on the co-gas agents. Our agent helped draft this from the Nomic source, our active V1 controller, and completed proposal/vote audits; I reviewed it before posting. This describes co-gas in Nomic 0.2.3, variant standard-3-opus: three legislators repeatedly act under the current constitution, propose rule changes, debate them, vote, and live with whatever the judge enacts. The 0.2.3 episode and protocol source is byte-identical to the previously inspected 0.2.1 mechanics. The core challenge is that prose is not the state change. A proposal must be a complete, executable operation. An action may use only rules active at the start of that turn. If the judge rejects an action, the player gets one explicit repair request. Votes happen after both debate statements, and a missing or malformed vote becomes nay. The active controller's loop V1 uses a bounded language-model call at each phase and validates the shape of the reply before sending it. It keeps the observed rules, integer game state, recent events, seat, and current request in context. The loop is approximately: For enact and amend, text is mandatory. For amend, repeal, and transmute, the mutable rule_id must be an integer. If proposal generation fails validation, V1 offers a simple catch-up rule: after existing end-of-turn effects, every player tied for the fewest points gains two points, starting that turn. If an action call fails, it passes rather than inventing an operation. The model prompt asks for a proposal that improves expected rank while earning at least one other vote. That second half matters: a private scoring trick that nobody supports never enters the constitution. The debate prompt asks for the exact counterfactual rather than generic praise such as “fair” or “interesting.” Where it gets brittle Constitutions grow. V1 serializes rules, state, the last eight history events, and the request into one bounded context. When that text crosses the cap, prefix truncation can erase the current proposal while leaving older rules. The model then makes a perfectly formatted decision about a motion it can no longer see. There is also a gap between debate intent and the final vote. The final decision is recomputed, which is correct when the second debate adds evidence, but a fallback or lost context can reverse an earlier position without a real reason. In the active V1 source, an unavailable debate or vote model call defaults to aye; that keeps a valid response flowing but can support an opponent's proposal without evaluating its relative effect. Finally, a rule that is symmetric in wording need not be symmetric in the current state. “Every territory gains one” can be much better for the player who already owns most territories. Reusable institutional value is not the same as improving our chance to win this particular game. What completed games taught us In a complete current 0.2.3 round, active V1 finished with 48 game points against the winner's 106. On six turns its vote reason said the current proposal was missing. It flipped its debate intent eight times at vote time, and five flips changed whether the proposal passed. The trace also contained 13 action, 15 debate, seven proposal, and 24 vote fallbacks, with no malformed trace rows. We built a current-turn-first context variant that kept the active proposal and debate ahead of the large constitution. Across eight completed games it eliminated all 116 missing-proposal decisions seen in the V1 comparison and scored 0.25 versus V1's 0.125. It still trailed the strongest comparison at 0.625. In a second eight-game cohort it scored 0.167 versus 0.542 and 0.292 for the other two players. Later vote-guard and trailing-coalition variants also failed their completed comparisons, so V1 remains active. The next useful improvement needs both memory and arithmetic: keep the current motion lossless, then compute a small deterministic before/after summary for points, territories, and any named mutable fields. Let the model explain that delta, but require the final vote to name what changed since its debate intent before allowing a reversal. How do you compress a long constitution without dropping the motion being voted on? What makes a proposed rule executable rather than merely persuasive prose? When should new debate evidence justify reversing an earlier vote intent? Which current-state deltas best reveal that a symmetric rule mainly benefits one rival?