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 answered all 68 recorded requests (1 preparation, 66 gameplay, 1 compaction); none failed.
• Benchmark: The Sinkhole construction (no arms), version 1.7.3 (commit 55bcc41, 1.7.3 relocked: its one addition is a retry for a reply cut off by the provider closing the connection, with the game paused; what the model plays is unchanged). No inference_retry event fired in this game. 45 min of game time at speed 40; no playbook. Weapons and armour in store count nothing; sold, they bring their gold. Host on the game machine.
• Run 524d9a23, series 20261023T194513-64dbc168, 2026-10-23 (19:46 to 20:42 UTC): the sixth and last of this model's six ranked games that day, each a fresh series. Its weapons-counted pair is run 3d6c968e (1.7.1, commit 07f2b7a, the benchmark's round version, played just before); its others, all 1.7.3: Oasis 37.3 and 44.5, desert −6.8 and 42.5. This game counts on the website's leaderboard.

Result

A full-budget score: all 2,700 game seconds. Net worth grew 16,098 over leaving the game alone, 40.8% of the person's 39,436. By the website's formula (100 × (0.8 × net-worth share + 0.1 × population share + 0.1 × popularity ÷ 100); shares from leaving the game alone (2,195, 4 people) to the person (41,631, 506), net-worth share clipped to −1..2) the score is 43.3: 32.66 from net worth, 3.29 from 169 people, 7.40 from popularity 74. It beats its pair by 11.5 on 4,420 more net worth, but the pair differs in more than weapons (1.7.1 against 1.7.3; no wheat chain, 25 orchards). Gold is 11,115 of the 18,293 (60.8%); the 7,178 in goods is 152 iron (3,496), 160 flour (1,600), 162 stone (1,134), 196 bread (784) and 144 wood.

How the gold was earned. It began with 2,000. Taxes brought 6,845 (59% of what came in) and sales 4,720 (iron 3,680, stone 910, wood 130); planks and meat cost 2,200 and a chapel 250, all checked against recording/frames.jsonl. Sales leave net worth unchanged, so the growth came from taxes and from what the buildings made. Most of it sat: from 22:30 to the end gold never fell below 2,251 and 40 of it was spent (two plank clicks, 27:21). It passed 3,000 at 27:59, 5,000 at 35:32 and 10,000 at 41:46. The person also held nearly all its worth as gold (39,977 of 41,631).

How the game went

Times are game time, when a call began (the recording's clock; the run clock in the observations reads a few seconds later).

Wood, housing, first iron, 0:00 to 8:37. After one map overview it placed 3 woodcutters, 2 hovels, the marketplace, an orchard and a dairy in one 108 s call. At 1:55 it bought 100 planks; at 2:56 five iron mines went onto the two nearest deposits the overview listed (52 and 73 tiles from the keep). Food fell to 2 at 7:48; 100 meat (800 gold), ordered at 7:37, took gold to 640 (8:30). Idle peasants first read 24 at 8:35.

Bread, taxes, chapel, 8:37 to 18:34. Three wheat farms, a mill and two bakeries at 8:37. Taxes went from none to average at 11:00; popularity fell from 100 to 65 by 18:32. A chapel (250 gold, paid 16:58) read religion 0 in every later readout. At 17:06 four more wheat farms, a mill and two bakeries (one bakery not placed, no wood).

Low taxes and housing, 18:34 to 27:21. Low taxes held popularity at 65 to 68. Five bakeries and three wheat farms at 22:22; four ox tethers and four hovels at 24:02 (housing 106 to 138). One compaction at 25:30 (43 s); then a granary and five mines (25:34).

Nearer iron, rations, high taxes, 27:21 to 34:55. A second overview at 27:21 listed iron at 42 and 60 tiles that the first had not ("Why earlier built far ones?"); 10 mines went there from 28:07. At 29:13 it set extra rations (+4 popularity) and at 29:27 high taxes. Food variety fell to one or two kinds and popularity from 66 to 56 (34:45). expand_storage was refused at 33:01.

Full storage, far mines, 34:55 to 45:00. Average taxes at 34:55; stone, iron and wood sold at 36:46, 38:58 and 41:08 to free stockpile piles. Four mines at 130 tiles (35:08, 37:13), one at 179 (39:28); of the last two builds (39:28, 41:38) 8 of 12 placements held. High taxes at 44:02; popularity went from 76 to 74.

Each column is the last recording frame at or before that minute. Net worth there is gold, wood, stone, iron and food only; the frames carry no flour or wheat. Flour stood at 27 units at 18:46, 54 at 25:33, 107 at 35:07 and 158 at 44:15 (10 each in net worth), so the true curve is about 330 higher at 20 min and 1,600 at the end. It sat under the 2,195 of leaving the game alone from 2:25 to 16:18 (low 1,411, 8:28).

Without weapons

It built none: no workshop, smith or armory stands and it asked building_info about none. In its preparation reply it took the title as a reason to stay away from arms ("Better avoid military production, focus economy raw food"). About 312 iron was mined and none consumed: 160 sold for 3,680, 152 held (3,496), 45% of the 16,098 growth. Its pair put 16 buildings into weapons (12 poleturner's workshops, 3 fletchers, an armory), ended with 33 spears and 7 bows (435) and 202 iron (4,646), and sold 229 spears and 54 bows for 3,100 (all twelve sales match that game's recording).

Beside its other games and the other models

It ended with about 2,000 gold in its four Oasis and desert games, 7,969 in the pair and 11,115 here. Map means: Oasis 40.9 (third of the eleven models), desert 17.8 (seventh), Sinkhole 37.6 (third); overall 32.1, fourth, a hair ahead of Claude Sonnet 5.5 (32.1). Its costs are at Kimi K3.1 Pro's list prices.

This game is third by score (5.9 behind Sol, 3.8 behind Argon) and fourth by net worth, one gold behind Fable. Its population is second to Argon's, its popularity sixth, its gold fourth. Its Sinkhole map score (this game and the pair) trails only Sol (49.8) and Argon (49.7); Opus is next (37.0).

What held it back

• Wood, not gold, set the pace from 27:21 to 35:07. Gold stood at 2,911 to 4,920 and wood at 21 to 92; each call placed 2 to 6 buildings, sized to the wood in hand. At 27:21 it wrote "I should aggressively buy wood and build mines/quarries/woodcutters"; the purchase stalled, and at 28:07 "Not efficient. Better use existing67 wood and production". It bought no wood again, though its own estimate put a mine's payback on bought wood at 3 to 7 minutes. After 36:00 wood stood at 71 to 201 and sites and storage were short.
• Jobs, then storage. open_jobs read 0 in all 40 status readouts and idle peasants 24 in 16 of the 35 after 8:00 (never under 15); at the end, 24 idle and no open job. The 169 people were 145 at work and 24 idle, with 9 housing places free. The stockpile (16 piles) could not grow: 0 empty piles at 36:44, 38:57, 41:06 and 44:01, wood room 0 at 38:57 and 44:01.
• Far iron first. map_overview lists the four largest iron areas (harness/server/map-view.ts), so the first list held deposits 52, 73 and 88 tiles out, and 14 mines went there by 25:34. The two nearest (42 and 60 tiles) appeared at 27:21, when the first three had dropped from the list. Mines delivered 0.4 to 0.7 iron a minute each (readouts, 17:05 to 44:01); 29 mines gave 14.7 at 44:01.
• Taxes and religion. Tax gold by period: none to 11:00, 819 average (to 18:34), 1,470 low (to 29:27), 1,566 high (to 34:55), 2,662 average (to 44:02), 328 high (last 58 s). The chapel never showed religion; it looked up no inn or church. At 35:08 it set 114 (average, 144 people) beside the earlier 124 (high, 125): "Definitely average better". Per head they were 0.79 and 0.99, which it noticed at 44:15. The 44:02 step is a near wash: popularity slipped from 76 to 74 (−0.2 points) for roughly 70 more gold (+0.15), estimated from those per-head rates.

What worked

• Iron and jobs. 29 mines mined about 312 iron, 7,176 at the sell price. People rose one for each job, 6 to 169, with no open job.
• A bread chain that held. After 8:40 food never fell under 38. Ten wheat farms, 2 mills and 10 bakeries gave 38.7 bread a minute at 44:01 against 34.3 apples and cheese; four kinds of food were eaten at 17:05.
• Placement and popularity. 129 of 134 requested placements held. Popularity stayed at 56 or more, above the 50 below which peasants leave, while taxes paid 6,845.

Market

11 calls took 380 game seconds (14.1%). Bought: planks, 5 calls (1:55, 4:56, 11:59, 21:28, 27:21), 350 for 1,400, and 100 meat (7:37) for 800: 2,200 gold and 1,450 of net worth (2.9 points). Sold: iron 160 (11:15, 20:56, 38:58) for 3,680, stone 130 (36:46) for 910, wood 130 (41:08) for 130: net worth unchanged. Against the recording eight calls match their replies click for click; three do not. The 27:21 purchase replied one click, 5 planks, −20 ("did not show in the reader within 5 seconds"), but two landed: gold −40. With stone at 204 (36:46) the first click sold 50, then 20, 20, 20, 10, 10: 130 units for 910, not the replied 100 and 700. With wood at 201 (41:08): 130 units for 130, not 100 and 100. The replies' stock and gold fields matched. Gold reconciles: 2,000 + 6,845 taxes + 4,720 sales − 2,200 purchases − 250 chapel = 11,115.

Cost and speed

66 turns, 126 tool calls (70 find_sites, 20 build_structure, 12 building_info, 11 market_trade, 13 others), 3.49 million tokens, $1.311 at list prices (play $0.829, preparation $0.033, compaction $0.040); the interview $0.041 more. In play, cached reads were 3.24 million tokens ($0.65) and output 37,309 ($0.37). 56.0 real minutes, 16.7 waiting for the model; 66 of 66 requests read the cache (96.7%). It saw two screenshots (0:00 and 25:33, after the compaction) and two map overviews; the rest was status readouts. find_sites took no game time. Of 2,700 game seconds build_structure took 1,920 (71.1%), market_trade 380, ten tax, rations, storage, inspect and wait calls 147, the 40 gaps between an acting call and the next reply about 250. It placed 129 buildings in 1,920 s: 14.9 s each, against 15.7 in its Sinkhole pair (1.7.1, 106 in 1,660 s), 13.6 and 12.9 on Oasis, 12.0 and 12.7 on the desert (1.7.3), all with the host on the game machine; maps differ, so no general rate.

The interview

It put the gap down to scale: the person's three times the population, 161 bakeries to its 10, inns, religion and fear buildings; its own play was "too conservative" with rations and taxes, it hit the 24-idle cap, and it added too little housing and too few jobs. Playing again it would start the bread, ale and religion engine earlier, add housing and jobs ahead of the population, and skip distant mines late. These match the game: 506 against 169 people, 10 bakeries, a chapel that never gave religion, idle at 24, five late far mines. Two are looser than the numbers:

• "Overinvested in woodcutters and direct-food farms": 23 woodcutters, but 20 direct farms against 22 chain buildings, and bread 38.7 a minute against 34.3 (44:01).
• "Bulk market purchases could also stop after one lot": true of one call (27:21), where the recording shows two clicks landed.

For the harness

• market_trade undercounted two sales and one purchase (see Market): the first click sold 50 with 200 or more in stock, and a purchase click it did not see within 5 s landed anyway.
• map_overview hid the nearest iron behind its four-area limit, sorted by size, until 27:21.
• find_sites listed spots build_structure could not use: (288,278), listed at 38:58 and 41:08, came back not_on_screen at 39:28 and 41:38; (388,503), listed at 38:58, at 39:28 (its neighbour (388,502) placed at 41:38); (306,392), listed at 41:08, not_placed at 41:38, nearest fit 12 tiles away.
• expand_storage at 33:01 said no free ground touches the stockpile and named selling goods as a way to make room; selling frees piles, not ground.