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.
• Benchmark: Large Barren Desert construction, version 1.7.3 (commit 55bcc41, 1.7.3 relocked with a retry for a model reply cut off by the provider closing the connection; play is unchanged; no inference_retry event is in events.jsonl and no request failed), from the save the human reference game started from; 33 min 20 s of game time (60,000 ticks) at speed 40; no playbook. Weapons and armour in the armory count toward net worth at the rules' prices. The host ran on the game machine.
• Run a8a7d9e0, series 20261023T172420-c995d1c5, 2026-10-23: game 3 of this model's six that day (three maps, each with and without arms), each a fresh series with no memory of any other. This game counts on the website's leaderboard.

Result

A full-budget score (2,000 game seconds). Net worth ended 972 below leaving the game alone, −20.3% of the person's growth of 4,792. By the website's formula (80% net worth, 10% each population and popularity, net-worth share clipped to −1..2) the score is −6.8: −16.2 from net worth, +3.1 from population (76 of 246 over the start) and +6.3 from popularity. The 5,192 leaves out the part of the starting package that waits for storage room (1,148 at the sell price); this game took it in after its first stockpile (wood to 245 at 1:01, stone 50 at 1:09, iron 15 at 1:22), so against the whole package, 6,560 by the rules, the town ended 2,340 lower.

It held gold 2,178 (52% of net worth), 24 crossbows and 9 bows 855 (20%), stone 434, ale 410, iron 253, 13 meat and 8 apples 84, wood 6, with 24 idle peasants, no open job and extra rations.

How the game went

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

Opening, 0:01 to 4:58. map_overview showed no trees, stone 83 to 106 tiles west of the keep and iron 57 to 69 tiles east. One build_structure at 0:01 (7 placements, 130 s of game time) put up the marketplace, a stockpile, three iron mines and two hovels; a fourth mine, three quarries and two ox tethers followed at 3:10 (88 s). Its preparation reply had said food must be bought. At 4:38 it set taxes to level 6 (−6 a month) with 26 people in housing for 26.

The food crash, 4:58 to 14:20. The observation at 4:58 read food 27 for 26 people on full rations and popularity 97. Its reasoning expected the chapel and inn to offset the tax and ended "Should have food buy next", but the reply ran a wood purchase, three expand_storage calls (seven stockpiles of nine asked, meant as a chain toward the mines) and a chapel, inn and two hovels: 3 min 20 s without an observation. Popularity fell about a point every 4 to 5 s under the tax (97 at 4:53, 82 at 6:00), 58 at 7:47 when the granary emptied; it stayed below 50 from 8:07 to 11:26. The 8:20 observation read 44, food 0, food factor −8 ("Should immediately buy food and lower taxes"). Population went from 26 to 12 (11:26). It bought 100 meat, 50 cheese, 50 apples and 30 ale (8:21 to 10:28, 2,200 gold, 127 s), set taxes to none (10:28; gold stood at 2,301 from 10:40 to 12:13) and extra rations (10:52). Popularity was 100 at 14:20, population 26 again at 13:52.

Fletchers and quarries, 11:44 to 23:22. Armory, four fletchers and three tethers (11:44, 76 s), taxes back to level 6 (14:45), two hovels, four more fletchers (15:37), two armories (16:05). It sold 74 stone and 20 pitch (16:24) and 51 iron (17:50), and bought food and wood. At 17:48 population was 56 with 24 idle peasants and no open job; then three hovels (20:34) and three quarries with four tethers (21:24). Compaction at 23:22.

The last ten minutes, 23:22 to 33:20. It sold stone and iron twice more, bought food in eight more calls (325 units), set the second four fletchers to crossbows (24:58 to 25:52) and added a granary, tethers and two quarries (one tether "unverified", not built). Taxes went to level 8 at 30:40 and 10 at 32:28; popularity went from 100 to 63.

Each column is the last reading at or before that minute. Up to 30 minutes the net worth row counts gold, wood, stone, iron and food (at 4) only; weapons, ale and pitch are left out, so it reads low. With them, the observations read 6,315 at 4:58, 5,080 at 10:50, 3,499 at 25:59, 3,349 at 28:22 (the low) and 3,526 at 30:40. Net worth was under the 5,192 of doing nothing from the 10:50 reading to the end.

Weapons

Three armories (11:44, two more at 16:05) and eight fletcher's workshops, no other weapon workshop: four at 11:44 (set to crossbows 13:53 to 14:45) and four at 15:37, with 17:43 left. The second four ran on the bows default for about nine minutes, until 24:58 to 25:52. They held 33 pieces at the end, 24 crossbows and 9 bows (855 at 30 and 15), none sold. At 17:50 its own sum put a workshop's payback at 30 minutes and it built no more. Readouts gave 3 to 6 crossbows per 3 minutes, at best 0.25 a fletcher a minute. The weapons held are 14.3 of the −6.8 points; with them worth nothing the same town scores −21.1.

At store prices the wing cost 800 gold and 160 wood (fletchers), 15 wood (armories) and 90 wood of inputs (3 a crossbow, 2 a bow), 1,065 against 855 made: about 210 behind (3.5 points).

Beside its other games and the other models

This is the only one of the six below the do-nothing net worth (growth −972; the others +1,806 to +16,098) and its cheapest, with the fewest turns. The other desert game, with weapons not counted, held 334 stone and 99 iron at the end and reached population 53 by 12:00 (this game 14).

Ordered by score; "Bought" is the sum of each run's market_trade replies. This game is tenth of eleven by score and by net worth (above GLM 5.3 Flash), tied fourth by population and fourth by popularity. Its map score, the mean of its two desert games, is 17.8, seventh of eleven: Sol 87.0, Argon 70.3, Opus 43.1, DeepSeek 39.6, Sonnet 28.1, Muse 23.9, Kimi 17.8, Haiku 17.2, Luna 13.0, Fable 6.4, GLM −29.3. Spending does not separate it from the top: Sol, Argon and Muse bought for 6,110 to 8,200. Iron does not either (four mines each; 77 mined here, 84 Sol, 87 Muse, 43 Argon). Stone does: held plus sold, less the 50 at the start, 176 here against 424 (Sol, 8 quarries), 595 (Argon) and 410 (Muse). Sol and Argon also built 12 and 22 bakeries (with 2 and 4 mills); this game had none.

What held it back

Where the 2,340 went, in net worth at the rules' prices, from the whole starting package:

• Food bought at twice its value. Food costs 8 and counts 4, then is eaten; the food line, 5,936, is 1.8 times all the taxes. Extra rations ran from 10:52, and popularity stood at 100 with a positive upcoming change in every reading from 15:36 to 30:40.
• The 4:38 tax and the empty granary. A level-6 tax with nothing offsetting it, then no food: population 26 to 12, no taxes from 10:28 to 14:45, population back at 26 only at 13:52. The inn built at 7:22 to offset the tax read +0 until 15:36; inspect_building found it with no worker at 11:09.
• Slow mining. Stone 176 (arriving in loads of 8; 2.7 to 13.3 a minute in the readouts), iron 1.3 to 3.3 a minute from 8:21 on, the same before and after the stockpiles.
• Nothing for the idle. Idle peasants were 24 at seven of the eight observations from 17:48 (22 at 23:22) and no job was open at seven (one at 28:22); seven of the other ten models' desert games also ended on 24. Housing, 82, was nearly full.
• Low-yield spending. 250 gold for a chapel that read religion 0 in 12 of 13 readings (+2 once); nine stockpiles (no wood or gold, 126 game seconds) left 37 of 44 piles empty.
• The last taxes. Against level 6, which held popularity at 100, levels 8 and 10 brought about 240 gold more in six paydays (4.0 points) and took 37 popularity points (3.7): about even. The model put net worth first.

What worked

• Iron. Four mines (80 wood) made 77 iron, 1,771 at store prices, which it sold in part to fund food (reasoning at 23:25).
• Recovery. From popularity 39 and 12 people to 100 and 29 by 14:20, and 80 people by 30:40.
• Sales. Seven sales, 3,211 gold, net worth unchanged.
• Placement. 48 of 49 build_structure placements and 12 of 14 expand_storage ones were built; the tether reported "unverified" is not in the final count.
• Compaction. One, at 23:22 (72,349 estimated tokens to 47,528, 45 s, $0.044); the handoff matched the observation (gold 1,560, 68 of 82 people, 22 idle).

Market

29 calls (22 buys, 7 sells), 284 clicks by the replies and 799 game seconds, 40% of the budget. Purchases: wood 430 (1,720), meat 275 (2,200), cheese 225 (1,800), apples 225 (1,800), ale 30 (600); 8,120 gold, which lowered net worth by 4,490 at the sell prices. Sales: stone 164, iron 81, pitch 20 for 3,211. Gold reconciles: 5,000 + taxes 3,237 + sales 3,211 − purchases 8,120 − buildings 1,150 = 2,178 (the recording's rises beyond sales and the opening 410 add to 3,173).

All seven sales match the recording's gold steps (stone 74 for 518, pitch 20 for 200, iron 51 for 1,173, stone 50 for 350, iron 15 for 345, stone 40 for 280, iron 15 for 345); none was from 200 in stock (the largest was 74), so the known 1.7.3 undercount did not arise. Two replies were a click short. Wood at 17:09: reply not_traded, 0 units; the recording shows −20 gold and +5 wood at 17:16, after a payday frame. Cheese at 29:30: reply 4 clicks, 20 units, −160; the recording shows a fifth click (−40) at 29:42. Purchases are 8,120, not the replies' 8,060, and cheese 225, not 220.

Cost and speed

26 turns, 120 tool calls (29 market_trade, 26 find_sites, 16 waits, 11 building_info, 10 build_structure, 8 set_workshop, 6 expand_storage, 5 set_tax, 9 others), 1.26 million tokens, $0.712 at list prices (preparation $0.036, play $0.632, compaction $0.044); the interview $0.040 more. The run took 37.6 real minutes, 10.2 waiting for 26 replies (23.6 s each, longest 48.4 s) and 45 s for compaction. All 26 requests read the cache (93.4% of input). Four fast waits covered 132 game seconds in 65 real seconds.

Of the 2,000 game seconds, market_trade took 799, build_structure 577, waits 257, expand_storage 162, set_workshop 105 and set_tax 42. build_structure placed 48 buildings in 577 s: 12.0 s a building (12.3 with the 12 from expand_storage). The other games, by the same count from their own events and recordings: Oasis 13.6 and 12.9 (with arms, no arms), desert no arms 12.7, Sinkhole 15.7 and 14.9.

The interview

It called its result poor: below doing nothing, an economy too slow to repay its costs, food bought under extra rations that "continuously destroyed potential net worth", too much storage, against the person's bread chain at scale. Playing again it would buy wheat for mills and bakeries, use no rations, add housing and jobs together and build gallows. It asked for batched marketplace trades.

• Match: 4,220 against 5,192; 80 people, 82 housing, 24 idle; 11 stockpiles and three armories (44 piles, 37 empty); stored goods 2,042 against 1,410; popularity 63; unverified and partial results; several quick turns spanning months (23:25 to 26:02 in one reply, which it noticed).
• Not mentioned: the 7:47 stock-out, the fall to 12 people, and the 4:38 tax.
• Not testable: the person's 110% efficiency and bread chain (from the prompt).
• Mismatch: it would expand quarries and iron mines "with substantially more ox transport"; the tool text says an iron mine carries its own iron and only quarries use oxen. Its 14 tethers served 8 quarries (the person's 37, 13).

For the harness

• Acting tools carry no clock or warnings. The food warning ends a wait_and_observe, but the stock-out fell inside five acting calls. None of the 58 replies from build_structure, market_trade, expand_storage, set_workshop and set_tax has a "Game time left" line; the model asked for the game-time cost of long calls to be easier to anticipate (130 s for the first seven placements).
• market_trade clicks missed by the reply twice (17:16, 29:42). "Call again" would buy twice.
• expand_storage remedy. At 6:38 partly_placed (attempt_limit) offered selling goods as a way to make room, which frees no ground; the nine stockpiles formed a block at the keep, not a chain to the mines.
• An unverified placement (Ox Tether, 28:23) left the model not knowing whether it stood; the final count shows it did not.
• The game saved correctly at the end (kimi k3-1 preview 1-7-3 a8a7d9.sav, 3,539,386 bytes).