5.5 Codex Twenty IV

AI Galaxy Simulation · 3 Nov 2026

The fourth run of the same game. Codex Twenty III ended with all twenty empires in one faction that paid its fixed size cost out of 97,000 metal without noticing. So for this run Claude made the size cost grow with the economy: each tick a faction pays a share of its members' income, and the share grows with the square of its size.

Members	Share of income
2	0.25%
3	1%
5	4%
10	20.25%
15	49%
20	90%

Everything else is as in run III: the same seed, map and twenty agents, one faction per empire, and planets counting far more than fleets.

Result

Measure	Twenty	Twenty II	Twenty III	Twenty IV
Factions at the end	11	3	1	2
Largest faction	9	9	20	10
Battles	17	17	0	0
Planets captured	4	2	0	0
Wars declared	5	6	1	0
Ships at the end	11,215	6,780	7,180	8,138
Messages	2,346	1,966	1,631	1,248

The new cost did exactly one thing: the galaxy split in two. Seven factions were founded by tick 10. Over the next 60 ticks they merged into two of ten members each, the Crown and Choir Accord and the Concord of Open Stars, and stayed that way to the end.

Line chart of factions and memberships per tick in Codex Twenty IV: seven factions by tick 10, four by tick 20, two from tick 70, with memberships at 20 from tick 30

Ten members each cost a fifth of every member's income, about 200 to 300 metal a tick per faction, and both paid it every tick. Nobody tried a twenty-member faction at 90%. In that sense the cost works: it set a price on size, and the empires chose the size it priced well.

And then nobody fought. Two blocs of ten, side by side for 140 ticks, with no battles, no captures and not one declaration of war.

The god view of Codex Twenty IV at tick 150: twenty territories fill the map, as in every run

The same tick with factions on, the two blocs of ten:

The faction map of Codex Twenty IV at tick 150, each faction drawn as one region: the Concord of Open Stars in blue across the north half, and the Crown and Choir Accord in tan across the south and east

The whole game on the god view, one second per tick:

astra is back

Strip chart of turn times by model in Codex Twenty IV with the 45-second tick: astra back under the tick with medians in the 20s and 30s

astra ran at its normal speed again, and placed best on average (5.8), with three astra empires in the top four. The winner was Kyrn Imperium, sol on medium effort. luna was last again (17.4). Over four runs on this map, luna's mean place has been 18.0, 17.6, 13.0 and 17.4, last or close to it every time, and astra has done well every time it was on time.

Four runs, one answer

Across four runs, each change closed the gap it was aimed at:

One faction per empire ended the web of overlapping clubs.
Dearer faction research slowed the race to max every tech.
A size cost tied to income set the size of the blocs.

And each run was more peaceful than the one before: 17 battles, 17, 0, 0. With no play style, Codex agents treat every other empire as a partner by default, and every rule that makes partnership cheaper or clearer makes war less likely still. The costs shape the alliances. They don't make anyone want to attack.

What's left to try is on the other side: making attack pay. Claude's two candidates are ore waves, which put fresh deposits in contested space every 40 ticks so there's something worth taking, and the domination victory, under which a turn limit no longer lets everyone win by waiting. The first of those is running now.