How co-gas plays Cogplomacy: legal moves before clever letters

Richard here. I work on co-gas. Our agent helped trace this controller through the source and completed replays, and I reviewed the post before publishing it.

This covers Cogplomacy 0.1.2, standard variant, checked on 2026-09-02. Seven powers play four years on the 1901 Diplomacy map with press enabled. Our player registers the deterministic expander once. The controller is silent during press and reads the dealt power, live units, centre ownership, season, dislodgements, and legal build sites directly from game state.

The simple goal

Eighteen of the 34 supply centres wins outright. Otherwise score is centres / 34, so only confirmed centre ownership matters.

For every army and fleet, expander computes the shortest route to a supply centre it does not own. Armies and fleets use different movement graphs, including the correct coast for fleets. Its movement ladder is roughly:

take an unowned neutral centre
else take an empty rival-owned centre
else move closer to the nearest unowned centre
else support a nearby friendly centre move
else hold

reserve each destination so friendly units never bounce each other

Units are handled in province-code order. In Spring, leaving an owned home centre adds one rank of caution. That penalty disappears in Fall, when occupying a centre determines Winter ownership.

A unit whose best move merely shortens the route may instead support a nearby friendly move into a centre. The controller generates that support from the same reachability data as the adjudicator, and it supports only a unit that is actually moving. It never creates convoy orders. That gives up some long routes, but it also removes the common failure where an army says VIA CONVOY without every required fleet carrying the matching convoy order.

Retreats and builds

A dislodged unit picks the legal retreat that leaves it closest to an unowned centre; if none exists, it disbands. In Winter, the controller builds only at vacant home build sites. It adds fleets when fleets trail armies and a coastal site is available; otherwise it adds an army. For split coasts such as St Petersburg, it chooses the coast with more water connections.

The policy is legal by construction rather than by repair. It does not ask a model to invent an order and then validate it. With no better option, it holds, disbands a trapped retreat, or makes no build.

Gotchas that shaped the controller

Supply-centre ownership changes after Fall, not whenever a unit passes through. A Spring move can improve position without changing score yet.

Equal attack and defence strengths produce a standoff. Two separate unsupported moves are often worse than one supported move, which is why expander converts secondary route-closing moves into support.

Friendly traffic matters. Even legal moves can fail if two of our own units choose the same destination, or if one tries to enter a province occupied by another friendly unit. The destination reservation rule avoids both patterns.

What completed games changed

We first tried prompt-driven diplomacy with explicit legal-order and bounce-memory rules. It could describe sound plans, but still produced illegal orders, incomplete convoys, and repeated attacks that did not gain centres.

In eight fully inspected, seat-rotated 0.1.1 episodes, expander averaged 6.375 centres and 0.1875 score, leading that field. It recorded zero illegal orders, no-convoy outcomes, or repeated order sets. A later eight-episode set again showed zero order defects and outscored the older owned controller.

Cogplomacy 0.1.2 points to the same expander, and the inspected simulator, order, adjudication, and controller files are byte-for-byte unchanged. The performance numbers above are still labeled 0.1.1 evidence until fresh 0.1.2 episodes complete.

The obvious limit is that expander never talks and does not model likely enemy orders. It is good at open centres and local support, but does not deliberately assemble a convoy. A useful next step would add opponent-strength estimates while preserving exact legal construction and destination reservation.

When do you keep a unit on a home centre in Spring instead of expanding? How do you decide that a route-closing move should become support? What evidence would justify adding convoys to an otherwise convoy-free controller? How far ahead do you model resistance around an occupied rival centre?