Kimi K3.1 Pro

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: Kimi K3.1 Pro, through Moonshot's pre-release endpoint for it, kimi-1021-preview (released as kimi-k3.1-pro), reasoning effort high, output limit 32,768 tokens; costs at Kimi K3.1 Pro's list prices ($2 per million input tokens, $10 output, $0.20 cached, $2 cache write), as Moonshot reports no cost per request. Moonshot served all 75 recorded requests (1 preparation, 73 gameplay, 1 compaction); none failed.
• Benchmark: Oasis by the Sea construction (no arms), version 1.7.3 (commit 55bcc41, 1.7.3 relocked: it adds a retry for a model reply cut off by the provider closing the connection ("terminated"), with the game paused; what the model plays is unchanged). No inference_retry event appears in this run's events.jsonl or logs.jsonl, so the retry did not fire. From the save the human reference game started from; 2,000 game seconds (33 min 20 s) at speed 40; no playbook. Weapons and armour in store count nothing toward net worth; sold at a marketplace, they bring their gold. The host ran on the game machine.
• Run 2eaf2041, series 20261023T163352-a03aae08, 2026-10-23: the second of this model's six games that day, a fresh series with no memory of any other game. Its weapons-counted game on this map is run 0a155c24 (series 20261023T154701-a8c5da72), played earlier.

Result

A full-budget score: the run used all 2,000 game seconds. Net worth grew by 7,727 over leaving the game alone, 42.2% of the person's 18,293. By the project website's formula (80% net worth, 10% each population and popularity, shares measured from doing nothing to the person, the net-worth share clipped to −1..2) the score is 44.5: 33.8 points from net worth, 5.8 from 118 people and 5.0 from popularity 50. This game counts on the website's leaderboard. It is third of eleven no-arms games on this map by net worth and score, behind Gemini 4 Argon (17,390; 86.2) and GPT-6.1 Sol (12,591; 62.4), and 7.2 points above its own weapons-counted game (37.3).

The stored goods are 169 iron (3,887), 289 stone (2,023), 576 wood (576), 26 flour (260), 10 apples (40) and 2 bread (8): iron is 43% of net worth, gold 24%, stone 23%, wood 6%. The game ended on high taxes (118 gold a month; factors food +1, tax −6, religion +2) with 20 housing places free, 19 idle peasants and 24 stockpile piles, none empty. Open jobs read 0 in 43 of 45 observations; idle peasants read 20 or more in 29 and stood at the 24-idle cap, where new peasants stop arriving, in 15.

How the game went

Times are the game clock when a call began (the recording's clock; the run clock in the observations reads a few seconds later). The game is paused while the model replies and runs while a build or buy call clicks.

Opening, 0:01 to 4:52. The preparation reply planned to place the granary; at 0:01 the stats showed one standing and it went on. Its first build call (7 placements, 101 s) put up the marketplace, 3 woodcutters, 2 hovels and a dairy farm. It bought 100 planks (1:49), added a stockpile (2:48) and at 3:22 placed 4 iron mines, 2 hovels and a dairy farm (82 s). Population hit the 26 housing limit at 2:56. At 4:52 it placed two dairy farms, an orchard and the first chapel; a second orchard spot was refused.

Food scare and stone, 6:05 to 15:36. At 6:05 the granary held 5 food for 41 people. It bought 20 bread and 10 apples, halved rations (6:52 to 8:56) and set low taxes (7:05; first tax gold at 7:26); popularity fell from 98 to 86 inside the 87 s call at 7:21. The first quarry and 3 ox tethers went up at 9:20, a second chapel at 10:05 (gold fell to 51), and after selling 19 iron (10:28) and buying 100 planks (10:50) the 11:40 call placed 2 iron mines, a second quarry, a wheat farm, a mill and a bakery. Its first three woodcutters read 0 from 11:39. High taxes came at 14:43, the first stone at 15:09, and at 15:36 the third chapel, the eighth and last iron mine, a third quarry and a hovel.

Food and storage, 16:50 to 22:18. Food was 11 for 70 people at 16:50; it bought 10 bread, then 40 (18:00). At 18:30 (90 s) it placed 4 woodcutters, a quarry, 2 tethers and a second bakery. The compaction came at 20:06, when no stockpile pile was empty; 2 stockpiles went in at 20:10, and the fifth quarry at 20:56.

Mean taxes, 22:18 to 28:10. It raised taxes to mean at 22:18, bought 60 apples at 23:53 (27 food for 90 people), placed a fourth chapel and 4 hovels at 24:28 and 8 woodcutters at 25:38 (103 s), inside which popularity fell from 59 to 43. It set no taxes at 27:26 (low 42 at 27:27; under 50 from 26:40 to 28:13), added 2 stockpiles and set low taxes at 28:10.

The last five minutes, 28:22 to 33:20. A mill and 6 woodcutters (28:22, 87 s), high taxes at 29:56 (popularity 64), then 5 woodcutters and a tether (30:11, 98 s), during which food fell from 28 to 3. The granary read 0 from 31:56 until 30 apples landed at 32:04. Population rose from 93 (28:20) to 118; popularity fell from 65 (30:10) to 50.

Each column is the last reading at or before that minute. Before the end the net worth row counts gold, wood, stone, iron and food (at 4); flour is left out, so it reads a little low, and the last column is the true net worth. On that count it was under the do-nothing 1,220 from 2:15 to 14:54 (low 493 at 6:53) and under its own start of 1,376 until 15:53.

Gold collected is the paydays in the recording, about 27 game seconds apart (the mean row includes one at 27:27, during the call that set no taxes). Taxes brought 4,236 gold, and gold reconciles exactly: 956 at the first frame, +44 still arriving, +4,236 taxes, +437 from iron, −2,520 on purchases, −1,000 on four chapels = 2,153.

Without weapons

It built no armory and no weapon workshop and never searched for one. Its reasoning at 0:01 asked whether "no arms" barred weapon workshops (the rules bar only soldiers and defences) and settled on "Focus stone/iron, food, wood and taxes." Its weapons-counted game built an armory and two workshops at 4:40 and ended holding 379 of weapons, 7% of its stored goods.

Beside its other games and the other models

Against game 1, the same map with weapons counted, this game has 1,512 more net worth: stone +959 (289 against 152), iron +598 (169 against 143), gold +173, food goods +160, wood +1 and weapons −379 (none made). At 80 points over 18,293 that is 6.6 points, plus 0.6 from 12 more people. Game 1 had a quarry at 3:24 (stone from 8:17) and this game at 9:20 (stone from 15:09), yet this game ended with more stone and iron. Game 1 had no wheat chain and bought for 3,500 gold (net worth lost 1,945) against 2,520 (1,460) here; game 1's purchases were checked against its recording in that game's analysis.

This game is third by population, second by stone, seventh by iron, second-lowest by popularity and has the most woodcutters. By map mean (the two games on this map, from unrounded game scores) it is third, 1.9 above Fable; its means on the other maps are 17.8 (desert) and 37.6 (Sinkhole). Its cost is worked out at Kimi K3.1 Pro's list prices; the others are as billed.

What held it back

• Food ran short five times. The warnings read 5 food for 41 people (6:05), 16 for 58 (13:10), 11 for 70 (16:50), 27 for 90 (23:52) and 33 for 99 (29:55). It had seven farms (5 dairy, an orchard, a wheat farm): 28 of its 37 farm-site searches found nothing, all 27 after 11:40, although map_overview at 20:33 still listed 1,617 farmland tiles; the logs give no reason. It bought 70 bread and 145 apples (1,720 gold, 860 of net worth).
• Mean taxes on a miscounted plan. Its reasoning at 22:18 had popularity falling 2 a month to about 56 "if 11 months"; 10 min 53 s remained, about 24 months, and religion was +4. Religion fell to +2 by 24:28 and the monthly change to −4. The 2.5 minutes after the drop, on no and low taxes, brought 208 gold in 5 paydays against about 600 at the mean level (about 390 gold, 1.7 points; an estimate).
• The last call with food at 33. At 29:56 it called food "balanced" by setting production per minute (25.4) against consumption per month (24.75) for a "3 months" buffer with 3 min 24 s left, about 7.6 months, and wrote "no buy needed, saves net worth". The 98 s call that followed ended with food at 3.
• Late stone, few mines. Iron came first and the quarry, 50 tiles from the keep, at 9:20. Stone readings rose from 8 a minute (5 quarries, 5 tethers, 22:17) to 21 (8 tethers, 27:44); the logs do not isolate the cause. Eight iron mines, none after 15:36: searches at 22:18 around three ore patches found no site, though map_overview at 20:33 listed four patches of 101 ore tiles.
• Housing and the idle cap took turns. Population was at or within one of housing at 2:56, 6:05 to 7:19, 13:33 and 20:54. Its answer to idle peasants was woodcutters, 29 after 18:30; the last 19 went up in calls of 103, 87 and 98 s (14% of the budget), and population rose from 93 to 118 while idle stood at 7 to 15.

What worked

• Iron and stone. Four mines at 3:22, eight by 15:36, about 1.0 iron a minute each: 169 iron, 43% of net worth. Five quarries and 9 tethers gave 289 stone, second of eleven, 23%.
• Taxes and religion. 4,236 gold in all; high taxes from 14:43 brought 1,234 gold in 7.6 minutes with popularity 90 to 78 and religion +4 (two chapels from 11:39, a third from 15:36).
• Placement and compaction. 93 of 97 build_structure placements went through, plus 5 stockpiles, and the final count equals the placed count for every building type. The one compaction (20:06, above 65% of the 120,000-token budget) took 34.9 s and $0.041 on a 77,967-token request; the next calls followed the handoff.

Market

Twelve market_trade calls took 289 game seconds, 14% of the budget. It bought 200 planks (4 each, worth 1: 1:49 and 10:50, each funding the next build call), 70 bread (8 each, worth 4) and 145 apples (8, worth 4) for 2,520 gold (800, 560, 1,160), losing 1,460 of net worth (16% of the final value, 6.4 points), and sold 19 iron (437 gold, net worth unchanged).

Checked against the recording (gold and food before and after each call, paydays inside a call set aside), 11 of the 12 replies agree with the game. The exception is the apple purchase at 17:11: the reply was not_traded with 0 clicks and no gold change, but one click landed at 17:19 (gold 415 to 375, food 16 to 20): 5 apples for 40 gold. Its handoff recorded the purchase as not made, and its reasoning took the later apples for orchard output. The known undercount for sales from 200 or more in stock did not arise (the largest sale was 19 iron), and no purchase click was lost to a payday.

Cost and speed

73 turns, 154 tool calls (89 find_sites, 17 building_info, 16 build_structure, 12 market_trade, 7 set_tax, 3 expand_storage, 3 inspect_building, 7 others), 3.77 million tokens and $1.395 at list prices (preparation $0.032, play $1.322, compaction $0.041); the interview cost about $0.045 more. All 73 gameplay requests read the prompt cache, 97.0% of their input tokens; cache reads were $0.703 of the $1.322, output $0.402 (550 tokens a reply, 69% reasoning). The run took 48.0 real minutes, 17.2 of them waiting for the model (14.1 s a reply, the longest 60.6 s). No reply held two acting calls (44 held one, 29 only reads, which cost $0.58); its reasoning at 6:32 doubted that several tools in one reply are possible, and the controls say they run in order.

Of the 2,000 game seconds, build_structure took 1,204 (16 calls, median 80 s, 14 over 60 s), market_trade 289, expand_storage 64, set_tax 57, the final wait 57, inspect_building 38, set_rations 14 and the host's 43 five-second top-ups 215. That is 12.9 s of game time per building placed by build_structure (93 buildings). This model's other games: Oasis with weapons 13.6 (80 in 1,084 s), Desert 12.0 (48 in 577 s) and with no arms 12.7 (67 in 851 s), Sinkhole on 1.7.1 15.7 (106 in 1,660 s) and with no arms 14.9 (129 in 1,920 s).

The interview

It put the gap down to a compounding production-and-trade economy against its accumulation of raw goods: bread exports (57 bakeries against its 2), 202 people at 140% efficiency against 118 at 100%, a church and cathedral with no taxes against its tax-led popularity, and its 39 woodcutters against 10. Playing again it would lead with bread, add a church, cathedral and fear-factor buildings, tax nothing and expand housing continuously. It named five obstacles: farmland with no sites and no reasons, not_on_screen near the edge, a batch it doubted, calls over a game minute, and no auto-trade tool.

Against the game these match: the net worth and population figures, 2 bakeries, 39 woodcutters, high and mean taxes with popularity down to 42, apples bought several times, farmland with no sites, two not_on_screen placements, and calls over a game minute (14 of 16). These do not:

• "One early construction batch appeared to report placements that were not reflected": the batch it doubted was at 18:30. The status at 20:10 listed 10 woodcutters, 4 quarries, 5 tethers and 2 bakeries, the new ones included, and the final counts equal every placement reported. Its doubt rested on open jobs (1, not the 9 or so it expected) and wood (62, which is 78 less the batch's 52 plus about 36 delivered).
• "Place the granary": the save already had one, as its 0:01 reasoning and handoff say.
• "Repeatedly damaged popularity": it fell below 50 once (26:40 to 28:13) and ended at 50.
• "Expanding to substantial quarry, ox-tether and iron-mining districts": find_sites offered no more iron or quarry sites after 22:18 (earlier for farms); whether any existed is not shown.

For the harness

• Warnings arrive after a call, not inside it. Popularity fell 98 to 86 in the 87 s call at 7:21 and 59 to 43 in the 103 s call at 25:38; food fell from 28 to 3 in the 98 s call at 30:11, after a food warning (29:55) just before it.
• market_trade reported nothing for a click that landed (17:11; see Market).
• Sites listed and then refused. (439,379) and (439,378) were listed as fitting and refused with no_free_spot_nearby; a wheat farm at (439,380) was placed at 11:40.
• Searches without reasons. The 28 empty farm searches and the 22:18 iron and quarry searches gave no cause, while map_overview still showed farmland and ore.
• Edge placements. Woodcutters at (474,421) and (470,425) returned not_on_screen ("map edge?").
• The granary line. The rules' Start paragraph says to place the required granary; the save has one.
• No request failed and no retry fired. The reader's memory watchdog forced one collection (a 20.0 s stall, logged at 30:11); the game saved at the end ("kimi k3-1 preview 1-7-3 2eaf20").