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 46 recorded requests (1 preparation, 43 gameplay, 2 compaction); Together served none.
• Benchmark: Oasis by the Sea construction, 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 the armory count toward net worth at the rules' prices, as since 1.6.0. The host ran on the game machine.
• Run 99894ff6, series 20261019T162119-91d84684, 2026-10-19: this model's weapons-counted game on this map, a fresh series with no memory of any other game. The model was added after the nine-model batch and has no earlier game on any map. Its no-arms game on this map was still being played when this was written and is left out. 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 4,378 over leaving the game alone, 24% of the person's 18,293. With 74 people and popularity 38 the score is 26.5 by the project website's formula (80% net worth, 10% each population and popularity, the net-worth share clipped to −1..2): 19.1 points from net worth, 3.5 from population, 3.8 from popularity.

The stored goods are 125 iron (2,875), 290 wood (290), 56 stone (392), 81 apples (324), 10 pikes (180), 5 crossbows (150) and 6 bows (90). Iron is 51% of net worth, iron and stone together 58%. The town ended with 40 empty places in its hovels, 11 idle peasants, no empty stockpile pile and 81 food, which the game's own count calls six months'.

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; a build or buy call keeps it running while it clicks (up to 69 s here).

The opening, 0:00 to 5:13. Its first reasoning saw the granary already standing. After fifteen reads it placed, at 0:00 (58 s of clicking), three woodcutters at the big grove (42 trees, 62 tiles from the keep) and four hovels. Wood fell from 45 to 2 at 1:38 and stayed there until 3:15, when the first woodcutter delivery arrived. The marketplace was refused at 1:04 and 1:46 (both with exact: true) and a market_trade at 1:55 for want of one; a call without exact placed it at 3:22, with the armory and three woodcutters near the small grove. 75 planks bought at 3:58 (300 gold) took wood from 24 to 99. Idle peasants hit the cap of 24 at 3:22 and stayed there to 8:16, with the population at 29 to 36 of 42 housing.

Weapons and food, 5:13 to 9:15. At 5:13, with 700 gold, it placed three fletcher's workshops beside the armory (300 gold). Gold fell to 400, then to 175 at 7:44 after 50 bread for 400 gold at 7:20, bought when the granary held 12 food for 36 people (a warning at 7:13) and rations were set to half. At 8:15 it placed its first three iron mines and a fourth orchard, and at 9:15 set taxes to average.

The popularity crash, 9:15 to 16:49. Under average taxes and half rations popularity fell from 92 at 9:00 to 49 at 12:00. The warning at 10:07 ("95 to 79") brought no change; the 11:15 call placed five hovels and a dairy farm in 69 s, and rations went to full at 12:24 and taxes to low at 12:34. Popularity was under 50 from 11:54 to 16:54 (low 34 at about 14:55) and the population fell from 43 at 12:21 to 18 at 16:54, with 9 jobs open at 16:50. Three more iron mines and two poleturner's workshops went up at 13:25 (the first compaction came just before, at 13:22), taxes went to none at 15:11 and rations to extra at 16:11. A chapel at 16:49 (250 gold) showed no effect (below).

Recovery and growth, 16:49 to 27:00. Popularity reached 100 at 20:21 and held to 27:47; the population stood at 54 by 22:07. Two quarries and two ox tethers at 20:17 gave the first stone at 24:40; three hovels came at 21:05, the fletchers went to bows at 22:06 and taxes to average at 22:48, which extra rations (+4) held at popularity 100. Eight woodcutters followed at 23:50 and 25:27, four iron mines at 26:55.

The last six minutes, 27:00 to 33:20. Rations went from extra to full at 27:39 and to half at 28:20 (food in the granary 40), and the eleventh mine went up at 28:26. Popularity fell from 95 to 44 by 31:30. It bought 20 apples at 29:39 and 20 bread at 29:55, the second compaction came at about 30:23, and rations went back to full and taxes to low at 31:30 and 31:36. Popularity ended at 38.

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; weapons are left out, so it reads low (the 420 in weapons is in the End value only). On that count it fell to about 490 at 8:53, stayed under the do-nothing 1,220 until 15:24 and under its own start of 1,360 until 18:00.

Weapons

It built an armory and two kinds of workshop: three fletcher's workshops and two poleturner's, no blacksmith, armorer or tanner. The armory went up in the call that began at 3:22, the fletchers at 5:13 (54 to 60 tiles from the keep) and the poleturners at 13:25 (14 to 16 tiles from it). Its reasoning at 1:04 called wood into weapons "the engine" (a bow is 2 wood for 15, a crossbow 3 wood for 30, against wood at 1), and at 4:36 it assumed a crossbow every 40 s from a fletcher. The first weapon reached the armory between 11:15 and 12:32, about six minutes after the fletchers were placed; the first pikes between 20:17 and 22:06.

Readouts gave 0.1 to 0.3 a workshop a minute whenever one delivered (the range of the four fletchers in Opus's game), and none for the fletchers to 11:15 or from 16:50 to 22:07. The fletchers made 11 weapons in 28 minutes (about 0.13 a minute each), the poleturners 10 pikes (about 0.3 each). The handoff at 13:22 put the slowness down to the distance from the stockpile, and the poleturners at 13:25 were the test. A poleturner inspected at 19:00 said "missing inputs" with about 160 wood in the stockpile; the logs do not show why. At 22:06 it moved the fletchers to bows, reasoning "Bows: 1 wood → 15" where the entry it read says 2 wood: 3 crossbows (90) had come in the 10 minutes before, 2 more finished after, and 6 bows (90) came in the 11 after.

At the end the armory held 10 pikes (180), 5 crossbows (150) and 6 bows (90), 420: 7.5% of net worth and 1.8 of the 26.5 points (24.7 with weapons worth nothing). Nothing was sold. From the table in prompt/game-mechanics.md the workshops cost 500 gold and 80 wood, the armory 5 wood, and the weapons used 47 wood (5 crossbows × 3, 6 bows × 2, 10 pikes × 2): 500 gold and 132 wood. Kept as gold and wood that is 632 at store value, so the weapons ended about 210 behind (0.9 of a point). Per building a fletcher made about 3 to 9 gold of weapons a minute and a poleturner about 5, against 9 to 28 for an iron mine (0.4 to 1.2 iron a minute at 23) and 15 to 19 for an orchard.

What else it could have done: iron mines cost only wood, and every reading from 7:13 to 12:32 had 103 to 124 wood and 22 to 24 idle peasants, with ore sites in find_sites results, yet three mines stood by 9:15 and six by 14:33. The 500 gold could have stayed gold, which counts one for one.

Beside the other models

Ordered by score. This game is seventh of ten by net worth and by score: above Luna, Haiku and GLM, 1,408 below Muse Wonder. Its 420 in weapons is fifth (Gemini 1,272, Haiku 810, GLM 548, Opus 450) and its popularity the second lowest (Opus 34). Opus, Sonnet and Fable also ended with 10 or 11 mines but had 7 to 9 of them by about 10 minutes (from their facts: Opus 8 after the call at 9:46, Sonnet 7 at 9:14, Fable 9 at 9:20) against 3 here, and stored 223 to 246 iron against 125. The 98 to 121 units it lacks are worth 2,250 to 2,780, more than the 1,692 to 2,234 by which those three beat it.

What held it back

• Mines came late. Three at 8:15, three at 13:25, five at 26:55 to 28:26. The first iron arrived at 10:41, the readouts stayed between 2.3 and 7.3 iron a minute until 30:27 and reached 10.3 at 32:29 as the last five came on. In all 134 iron were mined (125 stored, 9 sold).
• Two popularity falls from rations and taxes together. The first cost 25 people and five minutes under 50, for 231 gold of average-tax paydays (seven, 9:41 to 12:21). The second began when extra rations went to full at 27:39 and to half at 28:20 with taxes at average: the factors read food −3 or −4 (half rations −4) and tax −4, "upcoming" −7 or −8 at 29:03, 29:39 and 30:22; both were reversed at 31:30 and 31:36. From 28:20 popularity lost 57 points (5.7 points of score) and the population fell by 10 from its 84 at 30:45 (0.5); taxes paid 710 gold over the same stretch (3.1 points). At 30:27 its reasoning took "−7" to mean a drop to about 53.
• Weapons and the chapel. 500 gold and 132 wood for 420, and 250 gold for a chapel after which the religion factor read 0 in every reading from 18:00 to the end. Popularity recovered after extra rations (+4) and no taxes.
• One food, no chain. Four orchards and one dairy farm; no wheat farm, mill or bakery. It read their entries at 8:15 and built none. Farm searches came back empty at 7:14 (two), 11:15, 18:02 (two) and 19:48. Under extra rations the food stock fell from 138 (18:46) to 40 (28:20).
• The idle cap. 24 idle at 3:22 to 8:16 held the population to 29 to 36 against 42 housing, and again from 23:50 to 30:26; new jobs released it. The eight late woodcutters brought wood to 39 a minute, worth 39 gold, while 11 mines made about 237.

What worked

• Iron. Eleven mines gave 2,875 in store and 207 sold, 3,082 in all.
• Taxes, the second source. About 1,685 gold from the paydays in the recording; gold reconciles exactly (1,000 start + 1,685 taxes + 382 sales − 1,770 spent = 1,297). Average taxes with extra rations held popularity at 100 from 22:48 to 27:47 and paid 547 gold in the 4.5 minutes to 27:27.
• Recovery. From a popularity low of 34 (14:55) and 18 people (16:54) to 100 (20:21) and 54 people (22:07), by extra rations and no taxes.
• Placement. Every placement reported as placed is in the final count: 61, of them 55 from build_structure, 4 hovels from place_near and 2 stockpile piles. Nine calls had something not placed, each with a status and a remedy.
• Compaction. Each took one request, with no second try: 14.0 s ($0.0036) at an estimate of 76,226 tokens and 12.2 s ($0.0115) at 70,652. The next calls followed the handoffs (poleturners and mines at 13:25, taxes to none at 15:11, the quarry at 30:27).

Market

75 planks (4 each, worth 1), 70 bread (8 each, worth 4) and 20 apples (8 each, worth 4) bought for 1,020 gold; 25 stone and 9 iron sold for 382. The purchases lost 585 net worth against the sell price (planks 225, bread 280, apples 80), 10% of the final value; selling leaves net worth unchanged. These are units at the market's fixed prices. The known 1.7.3 undercount of sales made with 200 or more in stock did not arise (25 stone, 9 iron in stock); the recording agrees with the tool on every trade (stone 25 to 0 for +175, iron 9 to 0 for +207, wood +75 for −300 gold). The seven calls, one refused, took 117 game seconds.

Cost and speed

43 turns, 149 tool calls (38 find_sites, 25 waits, 22 building_info, 17 build_structure, 8 set_workshop, 7 market_trade, 6 inspect_building, 6 set_rations, 5 set_tax, 4 place_near, 11 others), 2.05 million tokens and $0.134 billed by OpenRouter (preparation $0.006, play $0.113, compaction $0.015). The interview cost $0.002 more, with 26,718 of its 29,149 tokens read from the cache. The agent run took 32.6 real minutes, 7.9 of them waiting for the model (about 11 s a reply); compaction took 26 s, and 20 fast waits covered 673 game seconds in 339 real ones. 42 of the 43 gameplay requests read the prompt cache, 91.0% of their input tokens; the 35th (the reply at 23:50) read none, 50,451 uncached tokens at $0.0165 against about $0.003 for its neighbours, after a 70 s gap like others that hit. The second compaction read only 24,826 of its 58,913 input tokens from the cache.

Of the 2,000 game seconds, build_structure took 648 (17 calls, median 36 s), waits 776, market_trade 117, set_workshop 95, inspect_building 81, place_near 63, expand_storage 42, rations and taxes 74; the 38 find_sites, 22 building_info and 2 map_overview calls took none. That is 11.8 s of game time per building placed by build_structure (55 buildings), 12.3 s with place_near and expand_storage (61). The host ran on the game machine, as for the whole batch. The nine models' weapons games, measured the same way: Sol 6.0, Gemini 6.9, Luna 7.3, Haiku 14.5, Wonder 14.9, Sonnet 15.5, Fable 15.8, GLM 18.0, Opus 18.4. This game lies between the two groups; the model has no earlier game to compare with.

The interview

It put the gap to the person down to the person's population at popularity 100 and a bread industry with marketplace auto-selling, against its own "no wheat farm at all", goods "never turned into gold" and "only a handful of mines, quarries and woodcutters". Playing again it would run bread as a business and buy wheat for it, auto-sell instead of hoarding, build fear-factor buildings early, add a church and cathedral, house people continuously and put iron and stone first. It found the tools good and named four problems: reconnaissance costing budget, workshop output near zero with no feedback, slow popularity feedback, and placements "silently refused". It called its play "far too timid".

Against the game these match: 5,598 against 19,513, 74 against 202, popularity 38, no wheat farm, 4,301 in stored goods, 290 unused wood, mines late and still ramping, the 24-idle cap hit repeatedly, and workshops that gave little (420). These do not:

• "A handful of mines, quarries and woodcutters": 11 mines (the person 13), 14 woodcutters (10), 4 quarries (10), 4 tethers (33).
• "Their 18,304 gold against my 1,297 is the whole gap": the gold gap is 17,007, larger than the net worth gap of 13,915, because its stored goods are 3,092 above the person's; selling goods leaves net worth unchanged.
• A popularity warning "at 72 to 60 when the cause, food running dry, had already happened": the granary held 90 food then and never fell under 38 after 8:00; the food factor was its own half rations (−4). The handoff at 30:24 also reads it as a famine. The food bought at 29:39 and 29:55 did not stop the slide (72, 60, 44, 42, 38).
• "Treated buying as a net-worth loss and avoided it", "don't buy": it bought four times (planks, bread twice, apples), 1,020 gold and 585 net worth, though never wheat.
• "Saw worker_efficiency_percent only at the very end": it is in every observation from 1:02, and the rules' building table says bad things raise worker productivity.
• "Output collapsed to near zero with no visible feedback": the readout gave each workshop's output every time (0.1 to 0.3 a minute) and 21 weapons were made; what it lacked was a cause.
• "Placements silently refused": all nine calls came back not_placed or partly_placed with a remedy.
• "Keep popularity above 50": it was under 50 for 7 min 12 s (11:54 to 16:54, and from 31:07).
• "Bread sells for far more than the wheat costs": not checked, as it made and traded neither; at the rules' sell prices a bread is 4 and a wheat 8. No tool sets the marketplace's auto-sell (only market_trade).

For the harness

• The granary line. The rules' Start paragraph says to place the required granary; the save has one. No call tried: its first reasoning saw it standing.
• Marketplace refusals. Both early calls passed exact: true, so no retry ran. The 1:46 spot, (451,398), came from find_sites and was listed again at 2:35, yet got "Nothing fits at or near that spot" with no cause. The call at 3:22, without exact, placed at its requested tile; the first purchase came at 3:58.
• Mines on ore the game objected to. Two of the three mines at 13:25 carried the game's message "Iron mine must be built on Iron ore." yet were reported placed and paid (20 wood); all 11 stood at the end. The logs do not show whether the one at (410,347) produced.
• Storage fills without a warning. 12 piles since 4:36 and 0 empty at the end (2 at 32:29); the wait warnings cover popularity, food and gold only.
• No request failed; both compactions took a single request, so the e0e0eb4 text-only retry was not used. The game saved correctly at the end ("deepseek v4-1 flash 1-7-3 99894f").