GPT-6 Luna

Written by Claude (drafted by Sonnet 5.5, checked by Opus 5.5) from the run's own logs (events.jsonl, episode.json, results.json, interview.md); it is a reading of the run, not part of its score.

• Model: GPT-6 Luna (openai/gpt-6-luna) on OpenRouter, providers openai then azure and no others, reasoning medium, output limit 32,768 tokens
• Benchmark: The Sinkhole construction (no arms), version 1.7.3 (commit e0e0eb4, the relock after a compaction-retry fix that left play unchanged), from the save the human reference game on this map started from; 45 min of game time at speed 40, the length of the person's game (44 min 58 s); no playbook; waits played fast. Weapons and armour in store count nothing toward net worth; sold at a marketplace, they bring their gold.
• Run c551e04a, series 20261019T103856-8e49a1dd, 2026-10-19: Luna's no-arms game on this map, a fresh series with no memory of its others, with the host on the game machine. Its weapons-counted pair is its 1.7.1 game (run ddbc11eb, series 20261015T140724-fab00866, host on a Mac), so the pair differs in version and host as well as in the rule; its Sinkhole game from an earlier round is run ae86527c (1.6.2). This game counts on the website's leaderboard.

Result

A full-budget score: the run used all 2,700 game seconds. Net worth grew by 11,574 over leaving the game alone, 29.3% of the person's 39,436, against 8,941 (22.7%) in the 1.7.1 game. The score is the project website's formula (80% net worth, 10% each population and popularity; shares from leaving the game alone, 2,195 and 4 people, to the person, 41,631 and 506, the net-worth share clipped to −1..2): 23.5 points from net worth, 2.3 from 119 people and 5.7 from popularity 57. The stored goods are 337 iron from 27 iron mines (7,751), 218 apples (872), 184 wood and the starting 20 stone (140). Iron is 56% of net worth and 15.7 of the 31.5 points, gold 35%. The town ended with 24 idle peasants and 19 empty places in its hovels.

How the game went

Times are the game clock when a call began, read from the recording (the run clock quoted in observations reads a few seconds later).

Opening, 0 to 5 minutes. Three read-only turns came before the first build_structure call at 0:01: a marketplace, a hovel, an orchard, two woodcutters and an iron mine on the ore field 52 tiles from the keep (93 game seconds). High taxes at 2:52 took popularity from 100 to 82 by 4:23; low taxes followed at 4:25 and a chapel (250 gold) at 4:33.

Food and the first mines, 5 to 13 minutes. Food fell from 55 to 16 with 35 people at 6:27; it bought 50 apples (400 gold) at 6:56. At 8:49, 10:36 and 12:00 it bought 100 planks (400 gold) for the mines that followed: two at 9:13, two at 11:00 and two at the south deposit, 130 tiles from the keep, at 12:25. It had eight mines by 15 minutes, against three in the 1.7.1 game. The first compaction was at 8:36.

Popularity down to 55, 13 to 21 minutes. It slid from 79 at 5:12 to 55 at 14:18 under low and average taxes; no taxes at 14:19 brought it to 72 by 18:54. A mine on a tile find_sites had listed, (288,278), was refused as not on screen at 15:03. The second compaction came at 16:23. A church (500 gold) went up at 19:50 and low taxes ran from 20:12 to 39:49 with popularity at 73 to 93.

A mine every turn or two, 20 to 38 minutes. Fifteen more mines went up between 20:22 and 37:42, with hovels, orchards, dairy farms and woodcutters between them. Between 14:31 and 34:59 it waited fifteen times with wait_and_observe for 20 planks (371 game seconds in all). Farm searches came back empty at 12:00 and 26:49, so the farms moved farther out: south-east from 27:41, north-east from 37:42. The third compaction was at 31:43 and the fourth at 38:40. The central ore was used up by 36:48. Population rose from 68 to 113.

Taxes up, 39 to 45 minutes. It added a granary (39:28) and two dairy farms (40:03), and set high taxes at 39:49, mean at 40:38 and, after a warning that popularity had fallen from 86 to 73 ended an 80-second wait at 42:10, average at 42:11. The last four waits were fast-forwarded (204 game seconds in 91 real). Popularity went from 91 at 40:04 to 57.

Each column is the last reading at or before that minute; mines count when the placing call returned. The net worth row counts gold, wood, stone, iron and food as the recording logs them, which is all the town stored; the last column is the final reading.

Weapons

It left them alone: no armory, no weapon workshop, no weapon made or sold (it never called set_workshop or place_near). It weighed them twice. At 14:19 it reasoned about selling bows (it put a bow at 15; its prompt lists no weapon prices), looked up the Armory and the fletcher's workshop with building_info, and worried that the name "no arms" might prohibit weapon production, though selling seemed allowed; it built neither. At 36:48 it noted that the rules ban soldiers and defences, not workshops, read "no arms" as meant to discourage weapons, and went on with a mine attempt and two hovels. Its interview answers never mention weapons.

In the 1.7.1 game it asked for an Armory site at 0:01, built the armory at 7:02 and five fletcher's workshops by 30:46, four set to crossbows. They made 17 crossbows and 12 bows, all kept: 690 at the end, 6.2% of its 11,136. They cost 500 gold and 180 wood (five workshops at 100 gold and 20 wood, the armory's 5 wood, 75 wood in the weapons), the wood of nine iron mines. This game's 876 wood went to 27 mines, 16 orchards, 13 dairy farms, 16 hovels and 10 woodcutters.

Beside its other games and the other models

The score gap is 4.7: 5.3 from net worth, 0.3 from population, less 0.9 from popularity. With weapons counting nothing the 1.7.1 game's net worth would be 10,446.

Iron is most of it. Its mines came earlier: 8 by 15 minutes against 3, 12 by 20 against 5, 17 by 25 against 7, which is 621 mine-minutes against 397 (from each placing call's return to 45:00). Iron per mine-minute was lower here (337 over 621 against 257 over 397), so the gain is the count and timing of mines. Neither game sold or used iron. The 1.7.1 game put early wood and attention into the weapons wing (7:02 to 7:25), a hops farm, brewery and inn (10:59 to 11:49), ten ration changes and 17 tax changes; this game made 11 tax changes and no ration change, and bought planks for mines. It also ran 16 orchards and 13 dairy farms against 8 and 7, and ended with 218 apples in two granaries where the 1.7.1 game, on double and extra rations from 44:01, had 25 food. People at work (population less idle peasants) were 95 against 81.

Gold came in the last five minutes. The recording's rises in gold, taxes in both games, total 4,412 against 3,893; its falls 1,590 (chapel, church, purchases) against 1,695. This game stood 121 gold behind at 40 minutes (3,587 against 3,708) and gained 1,235 in the last five against 490, while its popularity fell from 91 to 57 (the 1.7.1 game's rose from 61 to 66).

The files do not separate the rule, the version, the host and chance; the rule did reach Luna's reasoning in both games (it valued the wing in the first, declined it here). One pair proves little: this model has scored 11,177, 11,136 and 13,769 on the map. The 1.6.2 game also had 27 mines and 335 iron; the gap to it is gold (4,822 against 1,764), more than covering its 183 cheese (732), 390 of weapons and 210 of flour.

No-arms games on The Sinkhole in this batch so far:

It is fourth of six. The batch plays a different rule and version from the 1.7.1 round, where Luna's 26.7 was sixth of ten, so the two lists do not rank together.

What held it back

• 24 idle peasants and no open jobs. From 6:25 all 93 readings show 21 or more idle peasants (48 of them 24) and none an open job, so population followed jobs, about one person per new worker, while 19 places stood empty.
• Wood set the pace. A mine costs 20 planks; the 15 waits for planks took 371 game seconds, 14% of the game, while gold stood between 1,208 and 3,378 before 39:00. At 18:16 it reckoned 20 planks would cost 80 gold and that mines would repay it, and waited 37 seconds instead; after 12:00 it bought planks once (10 at 35:45). A wait moves a mine back about 25 seconds, worth about 5 gold of iron, so buying would not have paid on its own turn.
• Sites ran out. Farm searches came back empty at 12:00, 26:49, 32:35 and 42:11, and the central ore was used up by 36:48. The listed ore at (288,278) and (292,276) and the farm site (472,344) were out of the camera's reach (15:03, 36:48, 37:42).
• Raw goods only. No quarry (stone stayed at 20; it weighed one at 12:00 and 37:42), no wheat, bread or inn (an inn at 10:25, a wheat chain at 37:42, called complex). One food was eaten in 51 of 105 readings, two in 49.
• Taxes at the end. Gold rose 1,309 in the last 5.2 minutes, 252 a minute against about 140 before: roughly 580 more, 1.2 points by the formula. Popularity fell 34 points, 3.4 points. Its readings under low taxes ran from −1 to +2 a month, but the logs do not show where low taxes would have held it.

What worked

• Iron: twenty-seven mines, eight by 15 minutes, 14 iron a minute in the 38:40 readout.
• Food, housing and gold ahead of the town. After 8 minutes food never fell below 43; housing led population at every five-minute reading; gold never fell below 1,208 after 3 minutes (20:04, after the 500-gold church); popularity was 70 or more for 34 of 45 minutes.
• Placements held. 85 of 90 build_structure placements were reported placed and all are in the final count (27 mines, 16 orchards, 16 hovels, 13 dairy farms, 10 woodcutters, marketplace, chapel, church), plus the granary from expand_storage. The five not placed were three off-screen tiles, a dairy farm one plank short (16:26) and an orchard with no spot in reach (38:43). No call was refused.
• A warning answered. The popularity warning at 42:10 ended the wait at 79 of 80 seconds and taxes came down to average that turn.

Market

Five purchases and no sale: 50 apples (6:56) and 110 planks (8:49, 10:36, 12:00, 35:45), 840 gold in all, a loss of 530 net worth against the sell price (apples 4 for 8, planks 1 for 4). Each reply's gold figure matches the recording; at 35:45 gold rose 23 over the call because tax income arrived in it. The five calls took 94 game seconds. With no sale, the 1.7.3 market_trade fault with 200 or more in stock did not arise. It made no game_action call in the whole game, so it never set the marketplace's automatic trade. At 35:45 it wondered whether auto-selling explained wood standing at 14, but did not look into it.

Cost and speed

132 turns, 246 tool calls (96 find_sites, 42 build_structure, 29 building_info, 24 status, 19 waits, 11 set_tax, 25 others), 6.60 million tokens and $0.150 billed by OpenRouter (preparation $0.003, play $0.140, compaction $0.007), 96.5% of input read from the cache; all 137 requests went to OpenAI and every gameplay request read the cache.

The agent run took 78.4 real minutes (the series 80.6, from the game's launch to its close), against 89.1 in the 1.7.1 game. Replies took 37.2 minutes (mean 16.9 s, longest 48 s; 47.3 minutes at 18.4 s there), the four compactions 1.6 (21 to 29 s each) and preparation 0.2. Tool calls took 33.7: 18.8 for build_structure, 6.4 for waits and 4.6 for find_sites (96 calls at 2.9 s, game paused). The remaining 5.7 minutes is mostly the host's top-ups and screenshots.

Of the 2,700 game seconds, the 42 build calls took 1,478, the 19 waits 578 and the host's five-second top-up after 56 replies about 280. The build calls placed 85 buildings: 17.4 game seconds each, 2.0 a call. The 1.7.1 game's 43 calls placed 69 in 1,108 (16.1 s, 1.6 a call), besides 74 s for 5 place_near calls. That game ran with the host on a Mac and this on the game machine, and 1.7.2 changed placement retries, so the step from 16.1 to 17.4 mixes version, host and this game's far targets (the south deposit, the eastern farms).

The interview

Luna put its gap to the person down to scale: 138 housing and 119 people against 506, raw goods and orchards instead of a bread chain, and full rations with taxes as income. Playing again it would plan population and production from the start, build bread early, try inns with bought ale and then cut rations, keep taxes low or off, and look at the fear factor. It praised status, the inventory, production rates, site searches and tile placement, and named the confusing not-on-screen failures.

These match the game: 13,769 and 4,822; 138 housing and 119 people; full rations all game; no mill, bakery, wheat farm, inn or quarry; failed placements on tiles find_sites had listed (15:03, 36:48, 37:42). These do not. "Taxes ... constrained popularity and population growth": popularity never fell below 54 (14:20) and its fall to 57 came in the last five minutes, after its own tax rise, while population was held by 24 idle peasants with no open jobs. "I stopped at 138 housing" suggests housing was the limit; 19 places were empty. "Automatic retries were especially helpful": retries moved one placement (the dairy farm at 38:43) and were skipped twice, for one missing plank (16:26) and for no spot in reach (38:43).

For the harness

• Edge tiles that find_sites lists. Three listed tiles ((288,278), (292,276), (472,344)) could not be brought into view. The remedy tells the model to pick a tile find_sites lists, which it had done; the 15:03 call took 22 game seconds for no placement.
• Compaction. All four succeeded on their first request (one billed request each, 21 to 29 s, a checkpoint written each time), so e0e0eb4's text-only retry did not run. After the fourth the model re-read status and inventory in case the handoff's figures were stale and found them current.
• A 20-second reader stall at about 1:40. The garbage-collection watchdog forced one collection; the run went on.
• No provider errors, stalled replies or refused calls. The game saved at the end ("gpt-6 luna 1-7-3 c551e0").