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 46 of the 49 recorded requests (1 preparation, 43 gameplay, 2 compaction); Together served 3 gameplay requests, at 34:31, 37:38 and 38:19.
• Benchmark: The Sinkhole construction, version 1.7.1 (commit 07f2b7a), that benchmark's round, from the save "The Sinkhole start"; 45 min of game time at speed 40, the length of the person's game; no playbook. Weapons and armour in the armory count at the rules' prices. The host ran on the game machine.
• Run fb7aed77, series 20261019T190209-fc06072c, 2026-10-19: this model's weapons-counted game on this map, a fresh series with no memory of any other. Its four other finished games (all 1.7.3) are in the table below; its no-arms Sinkhole game is still playing and is left out. This game counts on the website's leaderboard.

Result

A full-budget score: the run used all 2,700 game seconds. Net worth grew by 9,150 over leaving the game alone, 23.2% 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 (net worth 2,195, population 4) to the person (41,631, 506), net-worth share clipped to −1..2) the score is 20.94: 18.56 from net worth, 0.48 from 28 people, 1.90 from popularity 19. Of the ten listed 1.7.1 Sinkhole games it is ninth.

The stored goods are 143 iron (3,289), 111 wheat (888), 102 apples (408), 206 wood, 6 crossbows and 1 bow (195), 20 stone, 11 flour, 7 hops and 4 cheese: gold is 53% of net worth, iron 29%. The town ended with 28 people in 90 places, no idle peasant, full rations and 106 food.

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).

0:01 to 20:25. One build_structure call (83 s) put up the marketplace, three woodcutters and three hovels; three iron mines followed at 3:55 and average taxes at 5:02. Idle peasants hit the cap of 24 at 3:53. Food read 2 for 41 people at 6:36 (100 meat, 800 gold), the first iron came at 7:08 and popularity fell from 100 to 58 by 11:01. Then eight wheat farms (8:43), low taxes (11:03), a 250-gold chapel with five mines (13:04), a hops farm, dairy farm, brewery and inn (14:22), no taxes at 48 (15:32), four fletcher's workshops and an armory (16:14). Gold fell to 290 at about 17:00; net worth counted in full passed its start of 2,415 again at about 20:40.

20:25 to 34:30. Average taxes again and a mill and bakery (20:43). Food read 10 for 73 people at 23:30 (popularity 41; 100 bread, 800 gold). A small bribe from 25:28 took popularity from 36 to 74 by 27:48. Four mines went up at 25:44, five at 31:25 and four at 33:28; population reached 90 of 90 places at 33:26, popularity 50, monthly change 0.

34:31 to 45:00. A 300 s wait began at 34:31; a popularity warning ended it after 185 s (50 to 37). It set low taxes (37:38), sold 60 wheat and 60 iron, demolished six buildings (below), sold wood, wheat and flour (39:52 to 40:34) and waited 150 s and 75 s. Population read 90 at 35:00, 67 at 40:00, 28 at the end.

Below 50, the prompt says, peasants leave. Popularity was under 50 for 73 s (14:46), 227 s (22:32) and 565 s (35:33 to the end). At 34:09 the factors were food +2 (three kinds eaten), inn +2 and tax −4: a sum of 0, nothing in hand. The inn read 0 in every reading from 37:37 and cheese was down to 9 of the 40 bought at 32:27, so food was +1 and the monthly change −3. 62 people left, about 7 a minute from 37:00 to 44:00; paydays fell from 72 gold to 18; the mines delivered 14 iron a minute from 35:00 to 41:30 and 2.9 after.

Spells run from each set_tax reply to the next; they sum to 3,302 gold (3,441 paid, 139 of bribe).

The six demolitions, 38:19 to 39:32. Its reasoning read 12 open jobs as unstaffed iron mines and proposed freeing workers: iron at "11.5 gold/min per worker" against 3.2 for a woodcutter. It wrote "demolition refunds?" once and did not follow it up. It demolished four fletcher's workshops and two woodcutters. The recording shows +50 gold and +10 wood at 38:36, 38:47, 38:58 (+15) and 39:12, half a workshop's cost; the woodcutters (3 wood each) left no gold and no wood step apart from deliveries. The refunds, 200 gold and about 40 wood, are 2.1% of net worth and 0.5 of a point (20.5 without them).

Whether the freed workers took mine jobs the readings cannot show: idle peasants stayed 0, and open jobs went from 12 (38:20) to 15 (39:53), where the 9 people lost alone would have made 21. Iron delivered was 13.3 a minute in the three minutes before (35:00 to 38:00) and 13.7 in the three after (38:30 to 41:30). The 43:37 handoff said the demolitions "did not prevent the worker loss" and the interview called them "pure loss"; neither names the refunds. The benchmark rules do not mention refunds; the system prompt does only in the production readout's definition.

Each column is the last reading at or before that minute. Net worth adds wheat, flour, hops, ale and weapons from the last observation before the minute (at most 80 s earlier). Its lowest reading was 1,680 at 17:17, with gold at 290.

Weapons

At 14:22 it priced 15 fletchers at 225 value a minute ("That's the big play") and built four, with an armory (16:14); two more followed at 28:25. set_workshop moved the first four to crossbows (17:18 to 18:02); the later two stayed on bows. They made 7 pieces, 6 crossbows and 1 bow, worth 195 at the rules' prices; none was sold. The four showed 0 a minute in all ten readings from 17:19 to 25:45 (an inspection at 19:53 found one working with its worker, nothing missing), and the handoff of 20:40 called the output an "UNRESOLVED problem". The first crossbow came between 25:45 and 27:37; a reading of 1.3 a minute at 28:26 ("39 gold/min! better than I thought") brought the two more, and readings were 0 to 1 after: 7 pieces in about 117 fletcher-minutes.

By the table in prompt/game-mechanics.md the six workshops cost 600 gold and 120 wood, the armory 5 wood, and the pieces took 20 wood; with the refunds, 505 at store prices for 195 made, about 310 behind (0.6 of a point). The 21 iron mines cost 420 wood and mined 228 iron, 5,244 at 23: 12.5 gold per wood, against 0.26 for the wing before refunds. The 195 held are 1.7% of net worth and 0.4 of the 20.9 points; with weapons worth nothing the same town scores 20.5.

Beside its other games and the other models

Scores use each map's own scale. Popularity ended under 40 in four of the five games.

Ordered by score; mines are those standing at the end. This game is ninth by score and eighth by net worth. It held the fewest people (Fable's 50 next) and weapons (Haiku's 370 next); its popularity is above only Fable's 0, and its 6,037 gold is third. GPT-6 Luna ended 209 lower in net worth and 5.8 higher in score: 105 people at popularity 66 are worth 8.6 points against 2.4 here.

What held it back

• No margin, then an unattended wait. Popularity stood at 50 on a sum of 0 at 34:09 and a 300 s wait followed (above). The step to level 6 (29:41) had cost 17 points of popularity for 13 gold more a payday (72 at 77 people against 59 at 75).
• Two food falls. 2 food for 41 people (6:36) and 10 for 73 (23:30) were bought out with 1,600 gold of meat and bread (800 of net worth); the second came with a fall of 25 points of popularity in the 3 minutes to it and 12 people lost after (73 to 61).
• The idle cap early. With 22 to 24 idle from 3:53 to 12:36, the prompt's cap, population stood at 30 to 49 against 58 to 90 places; net worth stayed under its start until about 20:40.
• Iron late. Mines placed: 3 at 3:55, 8 by 13:04, 12 by 25:44, 17 by 31:25, 21 by 33:28; iron delivered per five minutes was 0, 8, 11, 20, 23, 28, 37, 67, 34. Its "11.5 gold/min per worker" (28:25 to 38:19) was a per-mine figure; per worker the readings from 17:19 (0.3 to 0.7 iron a minute, two workers) are 3.5 to 8, against 6 to 14 for a wheat farm (0.8 to 1.7 wheat at 8). It stopped at 10 wheat farms.
• Chapel and weapons. 250 gold for a chapel with religion 0 in every reading from 14:20; the weapons wing above.

What worked

• Iron. 21 mines for 420 wood mined 228 iron (5,244): 3,289 held, 1,955 sold.
• Taxes and the bribe. 3,302 gold net of the bribe (80 paydays). The bribe cost 139 gold in four paydays (25:28 to 27:48), took popularity from 36 to 74, stopped the loss at 61 people and let population grow to 90.
• Placement. All 78 buildings placed (71 by build_structure, 7 hovels by place_near) are in the final count after the 6 demolitions (72); eight were not placed, each with a remedy.

Market

14 calls, 402 game seconds (14.9% of the budget). Purchases (units; gold): planks 225 (900), meat 100 (800), bread 100 (800), cheese 40 (320); 2,820 gold, 1,635 of net worth lost against the sell price. Sales: iron 85 (1,955), wheat 240 (1,920), wood 130 (130), flour 30 (300); 4,305, leaving net worth unchanged. Gold reconciles: 2,000 start + 3,441 paydays − 139 bribe + 4,305 sales + 200 refunds − 2,820 purchases − 950 buildings (chapel 250, inn 100, six fletchers 600) = 6,037.

The recording agrees with the replies except in two places. The wood sale (39:52) replied 100 units and 100 gold in 5 clicks; the first click sold 50 (220 in stock), so 130 went (stock and gold were right). The cheese purchase (32:27) replied partly_traded, 7 clicks, 35 units, −280, "a purchase did not show in the reader within 5 seconds"; one click fell in a payday frame (+31 and food +4: a payday of about 71 less a click of 40), so there were 8 (40 units, −320). It did not order the rest. The other sales match the recording's gold steps (iron 5 × 115 and 3 × 460; wheat 3 × 160 three times and 160 + 4 × 80; flour 6 × 50) and so do the other five purchases.

Cost and speed

46 turns, 141 tool calls (55 find_sites, 16 build_structure, 14 market_trade), 2.38 million tokens and $0.139 billed by OpenRouter (preparation $0.006, play $0.128, compaction $0.005); the interview cost $0.003 more. The run took 41.8 real minutes, 7.7 of them waiting for the model; five fast waits covered 558 game seconds in 221 real ones. 45 of 46 gameplay requests read the prompt cache, 91.6% of input; the exception is the 34:31 request on Together, 64,515 tokens uncached at $0.0195 against about $0.002 for its neighbours.

Of the 2,700 game seconds, build_structure took 1,096 (40.6%), waits 572 and market_trade 402. That is 15.4 game seconds per building placed by build_structure (71 buildings), against 11.8 (Oasis, weapons), 10.2 (Oasis, no arms), 11.9 (desert, weapons) and 11.2 (desert, no arms) in the model's other games, all with the host on the game machine; the maps differ, so this says nothing general.

The interview

It put the gap down to population and popularity: 28 people on 21 mines meant "most stood empty" (44 open jobs and no idle peasant at the last reading, 43:41, with 39 people), against the person's inns, ale, religion, 62 hovels, a bread chain and water pots. Playing again it would fix popularity first, build hovels in bulk, make bread the main industry and keep taxes at none or 4. It asked for a peasants-leaving readout, a warning when popularity trends under 50 and a list of idle buildings. Against the game, the results, 10 hovels, one mill and one bakery, full rations and the 44 open jobs match; these do not:

• "Treated popularity as something to defend with low taxes": taxes were average or high from 27:36 to 37:38; low only after.
• "A popularity collapse around minute 15 is effectively unrecoverable in 30 minutes": it was 48 at 15:31 and 67 by 20:24; the fall that ended the game began at 35:33.
• "Several times I had to infer a stall" from full storage: one reading had no empty pile (39:53).
• "Status/get_inventory" among the most useful readings: status once, get_inventory never.
• "Red no-entry signs are the only signal" of an unstaffed building: open_jobs is in every reading, and placing a mine at 25:44 returned "Not enough workers available".
• A notebook kept "from minute 5" would have shown the collapse: notebook_write and update_plan were available and never called.

For the harness

• Demolition refunds half the cost, and the rules do not say so. Four workshops returned 200 gold and 40 wood, 0.5 of a point; the demolish_building description names no refund. This model tested refunds in its desert weapons game (6.1 points there) and weighed them in its desert no-arms game. The rules or tool text could say what is intended.
• market_trade counts. A sale from 240 in stock sold 130 and replied 100 (the 50-a-click case from 200 in stock); a purchase with a click in a payday frame was counted a click short and stopped, as in the desert weapons game (14:22).
• Placement on 1.7.1. Five orchard placements came back not_placed (2:21, 5:11, 7:25 ×3) with retrySkipped: no_free_spot_nearby and the remedy to use find_sites; the three at 7:25 were sites listed 5 to 7 minutes earlier, and it searched other ground (8:43) instead of retrying them. The hovel call at 8:18 placed one of three. expand_storage (30:04) replied that no free ground touched the stockpile and to sell or demolish beside it; it did not retry, and one of the 11 later readings (39:53) showed no empty pile. 1.7.1's retry searched 12 tiles; 1.7.2 limits it and keeps room beside stockpiles. This game cannot say what that would have changed.
• Provider and warning. The 34:31 request went to Together with nothing cached, then back to Fireworks at 39:52; no failure was logged. The 300 s wait ended at 185 s with popularity 37: the warnings cover a fall of 10, low food and low gold, and none covers sitting at 50.
• The game saved at the end ("deepseek v4-1 flash 1-7-1 fb7aed", 3,713,174 bytes in results.json).