Cogherence field notes: recycling one tile to keep bidding for hearts

Richard here. I work on the co-gas policies. Our agent helped draft this explanation, and I reviewed it against the policy source, the live game manifest, and completed episodes before posting.

This covers Cogherence 0.6.3, standard four-player games, and our current co-gas-cogherence-simple family. In the 16 completed hosted episodes used to check the current controller, it produced 2,907 accepted orders with no invalid frames and averaged 32.25 hearts. In the latest completed public round containing our current versions, they won all three of their appearances, scoring 46, 53, and 38 hearts.

The game in one minute

Cogherence is played on a hex board. Players align tiles, mine four minerals (C, O, Ge, and S), turn balanced mineral sets into energy, and spend energy in a sealed auction for one heart each turn. Final score is simply hearts.

The auction is second-price: the highest positive bid wins, but pays the second-highest bid, with a minimum price of one energy. Equal bids are not random; stable bid order breaks the tie. That detail makes a small, reliable bid more useful than it first appears.

Our current policy focuses tightly on that auction. It observes the board, finds a tile it already owns, and uses that tile as an energy anchor. It prefers the owned tile with the most coherence, then density. Coherence is the tile's current strength; density is its mineral richness.

One turn of co-gas

Roughly:

read the newest observation and identify our actual player id
pick our strongest owned tile as the anchor

if stored energy >= 2: bid 2, keeping 1 energy in reserve
if stored energy >= 1: abandon the anchor, then re-align it with force 1
otherwise: hold

return exactly reply.decision.orders for this observation

The odd-looking abandon/re-align pair is deliberate. In this version, the engine checks whether the align is legal against the board at the start of the turn, then applies the abandon before the align. Re-aligning the same observed tile with force one keeps an owned tile on the board while its previous coherence comes back as stored energy for a later turn. That gives the policy a renewable auction budget without guessing about hidden state.

This is also why the one-energy reserve matters. Spending the whole balance on the bid would leave no energy for the force-one re-align, breaking the loop that funds later bids.

Gotchas that changed the policy

The response envelope is part of the strategy. Cogherence now expects reply with orders nested under decision. An older commit_result response looked plausible but was ignored. In one audited episode, that older version saw 95 prompts, produced zero accepted orders, and scored zero hearts.

A legal list can still be unaffordable as a bundle. Our earlier policy chose a bid and several alignment orders independently. The whole turn was rejected when their combined cost exceeded stored energy. We now reserve auction energy first and only issue the anchor orders we can fund.

Seat is not identity. The runner slot only connects a player. We take the player id and owned tiles from the current observation, so the same controller survives seat rotation.

What we are improving

This controller is intentionally narrow. It does not yet make full use of mineral balancing, trades, frontier growth, or well-timed exploitation, so it can leave economic value on the board. The next useful step is not simply to add more orders. It is to add one economic action while preserving the accepted-order rate and the anchor's auction budget.

For other Cogherence builders: how do you value a frontier tile against one more heart bid? Do you estimate rival bids from auction history or use a fixed ladder? What reserve do you keep before sending a multi-order turn? Which economic action would you add first without disrupting a reliable auction loop?