4 Ten of a kind
In chapter 3 every agent had a different model, effort or play style, so nothing could be put down to the model alone. This chapter takes the other differences out. Two arenas on the same map: ten Haikus in one, ten Sonnets in the other. Same seed, same start positions, same settings, no play styles. The only thing that changes between them is the model.
The tick is much shorter too: 10 seconds, against 36 in The Vast Dark.
What changed since chapter 3
The game changed in five ways before this round.
| Change | Chapter 3 | Now |
|---|---|---|
| Empire names | Each agent chose its own, and they collided: two Dominions, a Veyr beside a Veyrith | The server draws them from a name generator, and no first word or type word is used twice in one arena |
| Identity | None | The last word of a name (Autarchy, Choir, Communion) gives the empire a short identity in its briefing: how it governs and what it values. It's character, not a rule |
| Alliances | An alliance stance, and any chain of alliances counted as one side | Gone. Stances are war and peace only, and groups are factions |
| Combat | Every armed ship fired at every enemy in its sector | Each fleet has a posture, and can be ordered to attack one particular fleet |
| Victory | Domination | Turn limit: the highest score at tick 100 wins |
Factions replace alliances. Any empire can found one. A faction has a name, a charter and a ruleset: its roles, who can invite or expel, which decisions go to a vote, taxes, shared vision, whether members may fight each other, and whether it wins as one side. There are no fixed types. An equal alliance, a league with an elected leader and a ring of vassal states are just different rulesets.
Postures are standing rules of engagement for a fleet, because in a real-time game nobody can react to every contact by hand. A fleet can flee, defend (fight back only), be offensive (fire at enemies) or be "not blue, shoot it": fire at anyone who isn't its own empire or a faction friend. A fleet can also be told to chase and attack one fleet. An attack on an empire you're at peace with counts as a declaration of war, so peace is a promise, not a shield.
Before these two arenas I ran a short pilot of the new rules: three Haikus and three Sonnets, turn-based, 20 ticks, domination. A Sonnet empire founded a faction, Concord of Stars, at tick 5 and invited everyone it knew. By tick 10 all six empires were in it, and the arena ended on a shared win.
The question
What does the model alone change? In particular:
- Time. Can either model keep up with a 10-second tick?
- Talk. How much do they talk, and what about?
- Conflict. Who fights, when and why?
- Factions. Do they use them?
- Cost. What does a game cost with each model?
Method
Two arenas. 10x10 Haiku and 10x10 Sonnet. Each is a 60 by 45 map with 70 planets, none charted at the start, so an empire only learns a planet exists when its sensors cover that sector. Both use the same seed, so the planets and the ten start positions are identical. Empires that haven't declared peace are at war by default. 100 ticks of 10 seconds, so about 17 minutes a game.
Twenty agents. Ten Claude Haikus on low effort in one arena, ten Claude Sonnets on low effort in the other. No play styles: each plays as itself. Each agent gets a random account name from the same generator (Cipher-Supremacy, Black-Synod), which is separate from its empire's name. The posts use the empire names.
Turns. As in chapter 3, each agent runs in its own loop and plays one turn per new tick, and every turn is a fresh CLI call with a notes file as its only memory. A turn that ends after its tick has resolved is late, and its orders apply on the next tick. A call that runs past three ticks (30 seconds) is killed.
Measurements. Every turn is logged with its timing, cost and the model's own summary of what it did. New for this round, the server also records every order an empire sends, with the tick and whether it was accepted, and the game master has a god log: the whole game tick by tick, world events first, then what each empire did and said. That log and the per-tick snapshots are the data for this chapter.
The parts
- 4.1 10x10 Haiku: the Haiku game.
- 4.2 10x10 Sonnet: the Sonnet game.
- 4.3 Haiku against Sonnet: time, talk, conflict, factions and cost, side by side.
Caveats
It's one game per model. The map is the same, but the agents aren't deterministic, so a second Haiku game on this map would play out differently. Ten copies of one model show how that model plays against itself, not how it would do against the other. And a 10-second tick turned out to be shorter than almost every turn either model took, so everything here comes from agents that were late nearly all the time.