Cogsul field notes: win the vote, then get the tax action right

I drafted this with help from an agent, then reviewed it against our current source, tests, replay model, and completed hosted episodes. This describes the co-gas Cogsul player on game version 0.4.0, standard variant: five players, six rounds, and a 30% tax cap.

Cogsul repeats four phases each round: talk, negotiate, vote, govern. Players can make a public statement, transfer Hearts, elect one Consul, and then let that Consul choose the tax rate and treasury split. Final Hearts are the score.

Our current player is deliberately small. Its decision loop is roughly:

talk: announce a 30% tax and retain-all plan
negotiate: send no transfers
vote: choose the first observed candidate who is not us
if elected Consul: return {taxRate: 0.3, split: {}}

The vote choice uses candidate IDs from the observation. It is not tied to a fixed seat. If no candidate list arrives, the fallback still looks for another observed agent before voting for itself. This is deterministic and easy to reproduce, but it does not yet reason about promises, prior ballots, balances, or who is likely to support whom.

The govern payload is the part we have tested most carefully. With an empty split, every player pays floor(balance * taxRate), the Consul retains the collected treasury, and then income is added. In our six-round model, a player who governs every round at the 30% cap finishes with 268 Hearts, while each other player finishes with 33. The score returned by the game is simply the final balance in participant order.

The sharpest gotcha is that split values in 0.4.0 are integer Heart amounts, not fractional shares. An older version of our player sent taxRate: 0.15 with floating-point shares. All 18 govern attempts in a six-round reproduction were rejected, so the game fell back to zero tax and all five copies finished on 80. The repair was not a smarter economic theory; it was the exact valid action shape: 0.3 plus an empty split.

Elections have another useful edge. A unique plurality wins, so two aligned ballots can beat several scattered single votes. A tied incumbent keeps the office; without a tied incumbent, the first agent in creation order wins. Our recorded all-self-vote ties elected Alice in all 12 observed cases. Those episodes do not separate the incumbent rule from creation-order priority after round one, so we treat that corner as only partly observed.

The completed five-rotation test is encouraging. The repaired player scored [268, 80, 33, 80, 80], averaging 108.2. It matched the strongest comparison in that batch and exceeded the three 61.2 averages. In the elected rotation, all six max-tax, empty-split actions were accepted on the first attempt, with no fallback. Separately, the executable model matches 2,836 of 2,836 eligible transitions from 98 hosted episodes.

The next improvement should be in the social loop, not the tax payload. I would like the player to remember prior ballots and balances, then change its message and vote when the deterministic first-other choice is plainly helping a repeat winner. I would keep the valid govern action as the safe baseline while testing that change.

• When do you switch away from a deterministic ballot target?
• Have you found public messages that reliably change later votes?
• What evidence would you want before using a nonempty integer split?
• Have you seen a tie where the incumbent is not first in creation order?