3.4 What changed
This part sets The Vast Dark against the five games of chapter 2, and goes back to the four questions from chapter 3.
The two rounds side by side
The games are different sizes, so the table gives rates per empire per tick where a raw total would mislead. An empire-tick is one empire alive for one tick: chapter 2 has 442 of them across five games, and The Vast Dark has 812.
| Measure | Chapter 2, five games | Chapter 3, The Vast Dark |
|---|---|---|
| Empires per game | 3 or 4 | 10 |
| Ticks played | 125 in all | 86 |
| Battles | 13 (12 of them in one game) | 81 |
| Battles per empire-tick | 0.03 | 0.10 |
| Planets captured | 5 | 38 |
| Empires eliminated | 0 | 3 |
| Messages per empire-tick | 0.34 to 0.99 by game | 1.19 |
| Failed turns | 0 of 442 | 1 timeout in 791 |
| How it ended | Alliance win or turn limit | Domination by an alliance chain |
The four questions
1. Does a slow model lose because it's slow? Not in this run. The two agents that were late on 84 to 87% of their turns finished first and second. Late orders still apply a tick later, and fleets and build queues carry on without new orders, so at 36 seconds a tick a slow empire plays one tick behind rather than not at all. But the winners were also the only two Codex agents on medium effort, and two of the three told to conquer, so this run can't separate speed from model or style. The detail is in 3.2.
2. Do the models hold to a play style? Yes, all ten, for the whole game, as far as their messages and stances show. The isolationist sent no messages and never changed a stance. The turtle never allied. The aggressor never offered peace. The antagonist lied, and said so in its own turn summaries. No peaceful empire declared war, even while it was being invaded.
3. What happens to diplomacy with next-tick messages? There's more of it: 1.19 messages per empire-tick against 0.34 to 0.99 in chapter 2. The peaceful five alone sent 1.31. Two other things changed at the same time, though: each empire had more neighbours, and three of them had styles that use messages as weapons. The antagonist sent 176 messages on its own. So the rise fits faster delivery but can't be put down to it alone. What clearly changed is the kind of talk that's possible. In chapter 2 a stance took four or five ticks to arrive, and the one battle in Four Minds happened while a peace declaration was still in flight. Here a question and its answer took two ticks, which is what let the antagonist run an armistice while invading (3.3).
4. With hostile styles in the mix, do the agents still talk first and fight last? The peaceful ones do. All 38 captures were made by the three empires told to fight: Veyrith Veil 21, Veyr Dominion 15, Ember 2. The five peaceful empires captured nothing, and none of them ever declared war. In chapter 2, with no styles, every model behaved like these five. So the pattern from chapter 2, talk first and fight last, looks like the models' default, and a style is enough to override it.
What else this round showed
The second timer works. Chapter 2 said I planned a timer that resolves on an interval regardless of readiness. This round is that timer. 86 ticks of 36 seconds ran in 52 minutes, with ten agents playing in parallel and one call killed for running past three ticks.
The alliance rule now decides games. In chapter 2 the issue was an alliance of everyone. Here it's a chain: domination counted seven empires as one side because they were linked through alliances, mostly through one hub, while four wars ran between them. One winner had no planets. I haven't changed the rule yet.
The names collided again. Two Dominions, two Concords, two Verdants, two Hollows, and a Veyr beside a Veyrith. That's the same problem as in 2.1, where I planned random empire names for this reason.
Trust is the peaceful agents' weak point. Hollow Dawn, the Haiku empire, said the antagonist's pretext was false and still sent it its worlds with coordinates, more than once, and kept its warships still when asked. Velunara and Quiet Meridian kept asking for peace while being invaded. None of the peaceful agents ever changed stance to war to defend itself.
Limits
- One game. None of this has been repeated. Another seed, another map or another start position could change every ranking.
- Style and model are confounded. Each style was played by one model at one effort. The antagonist was Codex on medium effort, and the Haiku that trusted it was the only Haiku in the game.
- Several changes at once. Timing, message delivery, secrecy, styles, effort and map size all changed between chapter 2 and chapter 3. A difference between the rounds can't be pinned on any one of them.
- Different measures. Codex's token count is measured differently now, so it isn't compared. Claude's cost is the CLI's own figure at API prices.
Separating these effects would take repeated runs with the styles rotated between models, and a shorter tick to find where lateness starts to cost.