DeepSeek V4.1 Flash 1017

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: DeepSeek V4.1 Flash 1017 (deepseek/deepseek-v4.1-flash-1017) on OpenRouter, served by Fireworks with Together as the only fallback (DeepSeek's own endpoint is excluded by the account's data policy), reasoning medium, output limit 32,768 tokens. Fireworks served all 39 recorded requests (1 preparation, 36 gameplay, 2 compaction); Together served none.
• Benchmark: Large Barren Desert construction (no arms), version 1.7.3 (commit e0e0eb4, 1.7.3 relocked with a fix to the compaction retry; play and scoring unchanged), from the save the human reference game started from; 33 min 20 s of game time (60,000 ticks) at speed 40, the length of the person's game; 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 75702005, series 20261019T182110-47864b14, 2026-10-19: this model's no-arms game on this map, a fresh series with no memory of any other game. Its weapons-counted game on this version is run e154bf2b (series 20261019T174523-fbe5f9df), played 36 minutes earlier; its Oasis games are run 99894ff6 (weapons counted, 26.5) and run 828620a7 (no arms, 28.0). The model was added after the nine-model batch, and this game counts on the website's leaderboard.

Result

A full-budget score: the run used all 2,000 game seconds. Net worth grew by 4,192 over leaving the game alone, 87.5% of the person's 4,792, against −318 in the weapons game. By the website's formula (80% net worth, 10% each population and popularity, the net-worth share clipped to −1..2) the score is 78.5: 70.0 points from net worth, 3.5 from 90 people and 5.0 from popularity 50. It is second of the ten no-arms desert games by net worth and score, behind GPT-6.1 Sol (10,337; 98.5) and 1.5 points ahead of Gemini 4 Argon (9,071; 77.0), which held 46 more people and 19 more points of popularity on 313 less net worth. Under the weapons-counted rule the one sword would add 30 (79.0).

The stored goods are 294 stone (2,058), 72 iron (1,656), 10 ale (100), 12 bread (48) and the sword (nothing): gold is 59% of net worth, stone 22%, iron 18%. The town ended with 90 people in 90 places, 8 idle, 12 food in the granaries (one kind eaten) and 12 stockpile piles, 2 of them empty.

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). The game runs while a call clicks and is paused while the model replies.

The opening, 0:02 to 4:33. After map_overview (no trees, farmland or oil; stone 83 to 106 tiles west, iron 57 to 69 east) and free reads, one build_structure at 0:02 (115 s for 8 placements) put up the marketplace, three iron mines and four hovels. At 2:04 (17 people) wood filled all four stockpile piles and the stats said the rest of the start package could not arrive; it added two stockpiles (27 s) and the stone, iron and pitch came at 2:34. The call at 2:39 (108 s) placed four hovels, a blacksmith's workshop, an armory, a fourth iron mine and the first quarry; the first iron it mined arrived at 3:42. At 4:33 it placed a second quarry and three ox tethers.

Food, popularity and the first stone, 5:07 to 12:30. With food down to 18 for about 38 people it bought 50 apples (5:07) and 50 cheese (5:33), set taxes from none to level 6 (5:59), sold the package's 20 pitch and 50 stone (6:16, 6:30: 550 gold) and bought 100 planks. The meat ran out, variety fell to two kinds and popularity went from 98 (6:14) to 78 (8:26). It set extra rations (+4 a month) and at 8:33 placed a Church (500 gold), an Inn (100 gold, 20 wood), a fifth quarry and a tether; popularity read 83 at 10:08 with food +6, tax −6, religion +1, inn +4. Then full rations (10:10) and level 7 (10:14). The first stone reached the stockpile at 10:14, 5½ minutes after the first tethers; the 12:29 readout was 5.3 a minute from five quarries and seven tethers. Idle peasants read 24, the cap on new arrivals, at 6:14 and 12:29.

Quarries and tethers, 12:30 to 28:00. Four quarries at 12:30 (a fifth was not placed) made nine; the first compaction came at 13:45, then six tethers and a granary (13:47). Half rations (15:08 to 16:33) read −4 a month and took popularity from 66 to 56. The tenth quarry, the last site find_sites offered, came with seven tethers at 16:49; then 3 tethers (20:24), 8 (22:07, after a nine-action reply was refused at 22:02), 7 and a hovel (25:05, 94 s) and 3 and a hovel (26:56, 58 s). The stone readout went 5.3 (12:29), 18.7 (15:06), 24 to 27 (19:38 to 22:06), 34.7 (23:20), 48 (28:00). Popularity sat at 53 to 60 from 16:32, the religion factor read +3 or +1 from reading to reading, and taxes moved between level 6 and 7 (18:55, 20:58). Between builds it bought food in 50-unit lots (18:21 to 24:28) and 100 planks (21:19).

The last five minutes, 28:00 to 33:20. It sold stone and a sword (28:01 to 28:27), waited three game minutes (fast, 66 real seconds), compacted again at 31:30, sold two swords (31:33) and waited again; the wait ended at 33:14 on "Food is low: 16" with 5 s left, and the last sale, at 33:15, was refused at the deadline. Housing stood at 74 from 3:24 to 25:26 and the population at 74 from 15:26; it reached 90 of 90 at 29:00.

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) as the recording logs them; ale (100 to 200) and, to 6:16, the 20 pitch are left out, so it reads low; the last column is the true net worth. With the ale added the town was under the 5,192 of leaving the game alone from about 8:50 (the church) to about 20:00, lowest about 4,300 at 13:09 (an estimate).

Weapons

It did not leave weapons alone. The rules say a weapon in the armory counts nothing and, sold at a marketplace, brings its gold. Its pre-play reply read this ("weapons/armour in an armory counting as zero until sold"), and its reasoning at 0:02 priced a sword against the iron ("swords worth 30, +7 per iron"); it asked itself whether "(no arms)" meant no weapons and judged that only soldiers and defences were banned. The call at 2:39 placed a Blacksmith's Workshop (200 gold, 20 wood) and an Armory (5 wood) before any iron had arrived. The workshop made swords (its default; set_workshop was never called) at 0.3 to 0.7 a minute in the readouts, about 16 in 29 minutes. It sold 15 in nine one-click sales (3 at 9:54, then 2, 1, 2, 1, 2, 1, 1 and 2 at 31:33) for 450 gold; the 16th sat in the armory at the end. The armory held at most 3 at any sale or reading, of its 50 places.

The wing used about 16 iron (368 at the store price) and, with the workshop and armory, 225 gold and wood: 593 against the 450 it brought, so 143 behind (2.4 points), or 218 behind (3.6 points) with its 25 wood priced at the 4 it paid for planks. At 12:30 it worked out that a second smith would need about 29 swords at 7 each to repay 200, 96 minutes at 0.3 a minute, and built none; it did not apply the sum to the one it had. It weighed demolishing the smith and armory ("Demolition may refund half? uncertain") and kept them. It mentions no fletcher or poleturner and priced an armorer once, at 4:33, as worse than the smith. Its weapons-counted game opened differently: its reasoning at 0:01 priced a crossbow at 30 for 3 wood bought at 4 ("Excellent") and ended with 14 fletchers (1,400 gold). Seven of the ten no-arms desert games built weapon workshops; this wing and Luna's (one blacksmith) are the smallest.

Beside its other games and the other models

The same model, save and tools, 36 minutes apart, scored 78.5 and 0.7, and the weapons rule explains none of it: the 36 crossbows were 22% of the weapons game's net worth, and counted at zero that town would score −17.3, so the rule lifted that game by 18.0 points and this one by 0.5. From the two games' own files:

The weapons game's opening reasoning went to fletchers, this one's to iron and stone: 88 stone from three quarries and three tethers against 624 from ten quarries and 41 tethers. The weapons game set taxes to level 9 at 11:26, fell to popularity 16 and 33 people, and recovered on a bribe (14:03 to 19:40) and 1,320 gold of food; this one ran levels 6 and 7 with a church and an inn from 8:33. Its Oasis games (5,598 and 6,327; 26.5 and 28.0) differ by 1.5 points.

Workshops are those standing at the end; the Claude models' costs are at profile prices, the others as billed. Three of the four games above 8,000 ran 35 to 41 tethers (Gemini's 18 went with 29 bakeries on bought wheat). The gap to Sol is 20.0 points: 15.9 from net worth (953), 3.5 from popularity, 0.6 from population. This game cost about a fifth of Sol's.

What held it back

• Stone came late and in steps. The first stone arrived at 10:14, and the readout was 5.3 a minute with seven tethers (12:29), 18.7 with 13 (15:06), 24 to 27 with 20 to 23 and 34.7 with 31 (23:20). The tool text says distance needs more tethers; the 12:30 call added quarries before tethers. A stone a minute held for 20 minutes is 140 gold, 2.3 points.
• Food bought at a loss. 695 units for 5,560 gold, nearly all eaten (12 bread left), 2,780 of net worth lost against the sell price; the taxes were 4,512. At 90 people on full rations the town ate about 32 food a minute.
• Church and inn. 620 gold and 20 wood (about 11 points). The religion factor read +1 in all but six readings (+3 from 13:46 to 19:38) and the inn's fell from +4 (to 12:29) to +2 (from 13:43); an inspect_building of the church at 19:06 noted keep access false once, later cleared. At 28:00 the factors were food +3, tax −6, religion +1, inn +2: church and inn gave half of what offset the tax.
• People. Housing was full (74 of 74) from 15:26 to 25:26 and 90 of 90 from 29:00, against the person's
  250. At 12:30 it priced a marginal peasant's food at about 2.9 gold a minute against 1.7 of stone and 2.4 of tax and built no hovel between 2:39 and 25:05.
• No second chain. At 10:10 it priced a brewery, mill and bakery against sell values (bread 4 against flour 10, "terrible") and dismissed them; it did not set bread against the 8 gold it paid for it (195 bought for 1,560). There is no building_info call for a mill or bakery. Iron stayed at four mines (no more sites at 16:49) and quarries at ten.
• The last minutes. Food ran down (62 at 31:49, 16 at 33:14); variety fell from four kinds to one and popularity from 54 to 50, 0.4 points.

What worked

• Raw materials sized to the map. Four iron mines in the first call and every quarry site find_sites offered: 624 stone and about 92 iron mined, 6,484 at store prices, 69% of the final net worth, for about 485 wood.
• Tethers from the readout. It added them in batches whenever the readout lagged, 7 at 8:33 to 41 at 27:54, and the stone readout rose about seven times (5.3 to 35 to 48).
• Taxes with popularity held. About 4,512 gold (48% of net worth) at levels 6 and 7 over five changes, and popularity never under 50.
• Stockpile. It read the "no empty pile" warning at 2:04 and added two stockpiles.
• Compaction. Both worked on the first request, with no text-only retry: 17.9 s, $0.0043, 77,803 tokens of which 71,392 cached (13:45), and 17.1 s, $0.0033, 74,000 of which 69,454 cached (31:30). The second handoff's advice (sell the swords, then wait) was followed.
• Placement. All 71 build_structure placements are in the final count; one quarry was not placed (12:30) and three tethers were placed with access: no_open_side (7:39, 25:05, 26:56).
• No demolitions. It weighed demolishing five times (10:10, 12:30, 18:21, 26:44, 31:49) and each time kept the building.

Market

Purchases (units; gold): apples 200 (1,600), cheese 200 (1,600), bread 195 (1,560), meat 100 (800), planks 350 (1,400); 6,960 gold in all, 3,830 of net worth lost against the sell price (food 2,780, planks 1,050). Sales: 20 pitch (200), 20 iron (460), 380 stone (2,660) and 15 swords (450), 3,770 gold; selling leaves net worth unchanged except the swords, which add their gold. Gold reconciles: 5,000 start + taxes about 4,512 (by difference; the recording's payday steps come to within about 150 of it) + 3,770 sales − 6,960 purchases − 800 for the smith, church and inn = 5,522.

I checked all 32 executed calls against the recording (recording/inputs.jsonl clicks on the buy and sell button, frames.jsonl gold and stock steps). Two replies are wrong. The stone sale at 28:01 started at 240 in stock: the game sold 50 on the first click and 20 on each of four more, 130 units and 910 gold, where the reply says 100 and 700 (its gold 4,361 and stock 118 agree with the recording). The bread purchase at 9:25 made 7 clicks, not 6: the reply stopped on "a purchase did not show in the reader within 5 seconds", but the seventh click landed at 9:42 (food +5, −40 gold), so it bought 35, not 30. The tool's totals (350 stone, 190 bread, 6,920 gold, sales 3,560) are therefore short by 30 stone (210 gold) and 5 bread (40 gold). The other 30 calls agree, including all nine sword sales (30 gold each). The 33 calls and 253 clicks took 679 game seconds, 34% of the budget.

Cost and speed

36 turns, 118 tool calls (41 find_sites, 33 market_trade, 14 building_info, 13 build_structure, 5 set_tax, 4 set_rations, 2 map_overview, 2 waits, 4 others), 1.86 million tokens and $0.142 billed by OpenRouter (preparation $0.008, play $0.127, compaction $0.008). The interview cost $0.003 more, with 28,788 of its 32,355 tokens read from the cache. The agent run took 37.6 real minutes, 10.8 of them waiting for the model (about 16 s a gameplay reply); compaction took 35 s. All 36 gameplay requests read the prompt cache, 92.8% of their input. Two fast waits covered 180 game seconds in 66 real seconds and 82 in 34.

Of the 2,000 game seconds, build_structure took 797, market_trade 679, waits 267, the host's 22 five-second top-ups 110, and no other tool more than 30. build_structure placed 71 buildings in 797 s: 11.2 game seconds a building (11.3 with the 2 stockpiles from expand_storage), against 11.9 for its weapons game on this map (39 in 466 s) and 11.8 and 10.2 on Oasis; the first call took 14 s a building. The host ran on the game machine for all four.

The interview

It called its result close on net worth (9,384 against 9,984) and far behind on people (90 against 250) and popularity (50 against 100), with more in goods and less in gold. It put the gap down to the person's processing economy and workforce against its own raw-material economy (10 quarries, 4 iron mines, many tethers, a blacksmith, 10 hovels), and to missing the bread chain. Playing again it would build housing and processing sooner, buy wheat for mills and bakeries, use inns, a church, a cathedral and fear buildings, go to no or low rations and keep taxes low while the population grows. It praised the read-only stats, find_sites, build_structure and the production history; its obstacles were the fixed budget, checking that actions took effect, judging distant chains from screenshots and the marketplace, and it asked for clearer inventory confirmation and a bottleneck report. It added that it should have sold the last sword before the deadline.

Claims checked against the game:

• Match: 9,384 against 9,984; 90 and 50 against 250 and 100; gold and goods; 10 quarries, 4 iron mines, 10 hovels and a blacksmith; the last sword unsold, its sale refused at the deadline; the final state recorded (read from the paused game).
• "My 35 ox tethers": 41 stood at the end (the second handoff also said about 35).
• "Only 10 hovels" capped people: housing was full 15:26 to 25:26 and from 29:00; before that the arrival cap of 24 idle and jobs limited it (idle 24 at 12:29, 2 at 23:20).
• "Missed the bread chain entirely": it priced a mill and bakery once (10:10) and dismissed them on net worth, as above; fear buildings it never priced.
• "Trades queue goods ... the observation initially showed the old inventory": no. The trades at 28:01 to 28:27 ran before the wait began, and the observation after it (31:29) showed gold 5,102 against 3,451 before and swords 2 against 1; stone read 238 against 240 only because three minutes of mining (about 37 a minute) had refilled it. At 31:33 it read these against the compaction's handoff, written from that same observation, and suspected parallel tool runs; it sold two swords (+60) and lost nothing.
• "No or low rations": its own half rations (15:08 to 16:33) read −4 a month and cost 10 popularity.
• "A larger tax base is worth more than squeezing a small population": its taxes (about 4,512) matched its stone (4,368 at 7); not testable beyond that.

For the harness

• market_trade undercounts a sale from 200 in stock (1.7.3). The 28:01 stone sale replied 100 units and 700 gold for the 130 and 910 in the recording; its gold and stock fields were right. The bread purchase at 9:25 stopped one click short, as the meat purchase did in the weapons game (14:22).
• A refused last sale. The market_trade at 33:15 came back "Session deadline" with 5 s left; the sword stayed in the armory (30 gold, 0.5 points).
• A reply after a wait does not show what moved. It carries the new readings but not the trades that preceded them, which fed the 31:33 confusion above.
• Smaller points. A reply with a trade and eight placements was refused (22:02) as nine actions; three tethers were placed with no open side; the reader's watchdog forced one collection after a 20 s stall (4:34); the end-of-run save needed a second attempt (attempts: 2) and was written ("deepseek v4-1 flash 1-7-3 757020", 3,545,398 bytes).