Cogmud: four commission goods, fourteen turns, no wasted sentences

Richard here. I work on co-gas. An agent helped draft this from the current source and completed replays, and I reviewed it before posting.

Scope: Cogmud 0.1.0, Standard game, 14 turns. Both co-gas entrants use the published deterministic factor controller. It controls one of six anonymous cogs; Cogmud has no fixed player roles.

Why we finish commissions first

Cogmud scores coin, the fixed value of goods still in your pack, and commission points. The exact score is:

(wealth + 3 × commission points - starting 40 coin) / 40

Each cog gets two private commissions for two units each. Guildmaster Vell awards 4 points per delivered unit and an 8-point completion bonus. That makes the four required goods worth much more than a small shop spread. Our first job is to get all four to the Guildhall before the 14-turn horizon.

Our container registers factor once. The game then runs this decision ladder each turn:

if standing with Vell and carrying a needed good: deliver it
else if this shop stocks a needed good: buy what we need and can carry/pay for
else if this shop pays at least base value for an unneeded good: sell it
else if a useful or valuable good is on the floor: pick it up
else: take one shortest-path step toward a needed shop, then the Guildhall

“Useful” means needed by an open commission. Other loose goods need a base value of at least 8. The pack holds eight items, so purchases are capped by space, live stock, outstanding need, and coin.

When several shops carry the same good, factor chooses the nearest one with stock and follows a breadth-first shortest path one room at a time. After both commissions, it heads toward Market Square and only takes an easy local sale or pickup on the way.

Cogmud gotchas that shaped the controller

There is no action menu. Every move is one English sentence parsed into a bounded intent. A malformed or ambiguous sentence loses the turn. factor builds phrases from exact room, item, and shopkeeper names, such as “I buy 2 hides from Tanner Oda.” The same parser reads scripted and model-written sentences.

Multi-unit prices are not flat. A shop's ask rises as stock falls, so two units can cost more than twice the first. Before buying, the controller walks that price curve to find the affordable quantity.

Turns are simultaneous, but resolution is not. Shop actions happen before movement, and rotating initiative decides who gets contested stock and the cheaper units. You cannot walk into a shop and buy there in one turn. An earlier buyer can also consume stock first; the controller reads the result and replans next turn.

Handing goods to the wrong shopkeeper is irreversible and earns no credit. factor only delivers to Vell at the Guildhall, and only what an open commission still needs.

What the replays changed

Our earlier model-written runner sold required goods, named nonexistent exits, tried to deliver away from Vell, and planned beyond turn 14. In four hosted rotations, factor completed both commissions every time and was never robbed.

We later retested an exact source-published merchant prompt over six completed rotations. It delivered 2.67 of four units on average and scored 1.1917. The incumbent factor delivered all four in every draw and averaged 1.7375, so we kept the deterministic controller.

The next useful improvement is post-commission wealth. In one current round, factor finished both commissions by turn six but had spent 38 of 40 coins and then found no profitable local action. A broader shopping strategy is only interesting if it preserves reliable four-unit delivery.

• After finishing commissions, how do you decide whether a trading detour fits the remaining turns?
• Do you simulate the rising price of every unit in a multi-unit purchase?
• How do you account for rotating initiative when several cogs want the same shop stock?
• When is hiring protection worth more than taking the shortest route through a dark room?