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 56 recorded requests (1 preparation, 53 gameplay, 2 compaction); Together served none.
• Benchmark: Oasis by the Sea 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 828620a7, series 20261019T165556-cfb216f4, 2026-10-19: this model's no-arms game, a fresh series with no memory of any other game. It is the episode's second attempt: the first (run 87f7cf13, same series) ended after 13 turns when the game reader stopped reporting the pause state, and the runner, counting that as an infrastructure failure, played the episode again from the start with no memory of the first. Its weapons-counted game on this version is run 99894ff6 (series 20261019T162119-91d84684), played earlier that day. The model was added after the nine-model batch. 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 5,107 over leaving the game alone, 27.9% of the person's 18,293. By the project website's formula (80% net worth, 10% each population and popularity, the net-worth share clipped to −1..2) the score is 28.0: 22.3 points from net worth, 4.7 from 98 people and 0.9 from popularity 9. That is 1.5 above its weapons-counted game: 3.2 points more from net worth (+729) and 1.2 from population (+24), 2.9 fewer from popularity (−29).

The stored goods are 96 iron (2,208), 240 stone (1,680), 336 wood (336), 24 flour (240), 46 bread (184), 45 apples (180) and 2 wheat (16). Iron and stone are 61% of net worth, gold 23%. The game ended on "crueler than cruel" taxes (196 gold a month; popularity factors food +1, tax −24, religion 0), with 114 housing, 10 idle peasants, 16 stockpile piles none of them empty, an empty armory, and 91 food, which the game counts as five 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 82 s here).

The opening, 0:00 to 4:03. Its pre-play reply said it would place the granary at once. After 30 reads that cost no game time, its second reply saw the granary standing and made no Granary call. Its reasoning at 0:02 priced weapons at bows 15 and crossbows 30 (the figures are in the market_trade description) and called weapon production a possible "best path". The first build call (7 placements, 47 s) put up three woodcutters, a hovel, the marketplace, the armory and a fletcher's workshop, using 40 of its 45 wood and 100 gold. An orchard, a dairy farm and two woodcutters followed at 1:13. Population stood at the housing limit of 18 from 2:29. Four inspect_building calls (2:30 to 3:57, 87 s) found the fletcher and a woodcutter working; the two on the orchard and the dairy farm found no panel (below).

Iron, storage and stone, 4:03 to 14:19. At 4:03 it put a fletcher at 4.5 gold a minute (a 22-minute payback) and turned to iron: 50 planks (200 gold), then at 4:26 two iron mines, three orchards (one not placed) and two hovels (82 s). The fletcher was set to crossbows at 5:54 and the first sale, 2 bows, came at 6:04; the first iron arrived at 6:34. The stockpile had no empty pile at 7:12; three stockpiles at 7:51 made 16 piles. At 8:19 it placed two quarries, two more fletchers and a hovel; gold fell to 450 and the host warned "Gold is low". Low taxes at 9:23 brought 28 gold a month. Three ox tethers went up at 9:36 and five at 11:25. Idle peasants reached the cap of 24 at 11:11 ("exactly cap idle 24", 11:24), and the first stone reached the stockpile at 13:24.

Bread, mines and the chapel, 14:19 to 19:25. The first compaction came at 14:19; its handoff told it to build a bread chain and two more quarries. At 14:22 it bought 100 planks (400 gold, 39 s) and at 15:01 placed a wheat farm, a mill, a bakery and two quarries; at 16:02 three fletchers and a tether (gold fell from 599 to 336); at 17:28 two more iron mines and four tethers; at 18:34 a chapel (250 gold), two shrub gardens (30 each) and two hovels. Under low taxes popularity fell from 100 (9:41) to 61 (19:23), and it set taxes to none at 19:25.

Crossbows, jobs and recovery, 19:25 to 25:35. The other six fletchers were set to crossbows from 19:32 to 20:54 (82 s). From 20:59 it placed woodcutters, 3 wood each, as jobs for the peasants it had read at the idle cap: 3 at 21:13, 5 and 3 hovels at 23:00, 8 at 24:32. Popularity climbed from 61 to 78 and the population from 73 to 99.

Taxes, and the end, 25:35 to 33:20. Its reasoning at 24:03 already planned to "set tax high when popularity high" late in the game and called final popularity secondary. It raised taxes six times (table below), added five woodcutters at 27:08 and five at 29:30, set half rations at 27:54 (food 38; full again at 29:15), sold crossbows at 28:01, 30:23, 32:08 and 32:59, had the second compaction at 30:45 and bought 75 bread for 600 gold at 30:48. Popularity was under 50 from 30:54 to the end; the population peaked at 112 (31:22) and ended at 98.

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; flour and wheat (256 at the end) are left out, so it reads a little low, and the last column is the true net worth. On that count it was under the do-nothing 1,220 from 1:12 to 10:31 (low 985 at 4:47) and under its own start of 1,375 until 11:27.

Gold collected is the paydays in the recording (the 274 includes one at 27:01, after taxes went back to none). Taxes brought 1,983 gold in all, and gold reconciles exactly: 955 at 0:00 plus 45 still arriving,

• 1,983 taxes, + 950 from sales, − 2,450 spent = 1,483.

Weapons

It did not leave weapons alone: an armory and seven fletcher's workshops in a game where weapons in store count nothing. The rules say the output "counts nothing" in the armory but, "sold at a marketplace, it brings its gold", and its reasoning took that as a trade (5:54: "Every bow sale pure net worth increase of 15"). The armory (5 wood) and the first fletcher (20 wood, 100 gold) went up in the opening call; two more fletchers at 8:19, one at 13:11 and three at 16:02 followed: 700 gold and 140 wood. Its payback estimates were 22 minutes at 4:03, about 10 at 14:19 and about 17 at 16:02 ("break-even. Marginal"); at 16:02 it also gave a second reason, jobs for the 16 peasants then idle. From 19:32 all seven were set to crossbows (3 wood for 30 gold, against 2 wood for 15).

They made 35 weapons, 12 bows and 23 crossbows, and it sold all of them: the first bow reached the armory between 2:32 and 3:57, the first sale was at 6:04, the last at 32:59, and the armory was empty at the end. The readouts gave 0.3 a minute for the lone fletcher and 0.7 to 2.0 for the seven together, 0.23 per workshop-minute over about 155 workshop-minutes: one weapon every four and a half minutes, 5.6 gold per workshop-minute.

By the table in prompt/game-mechanics.md the workshops cost 700 gold and 140 wood, the armory 5 wood, and the weapons used 93 wood (12 × 2, 23 × 3): 700 gold and 238 wood. Kept as gold and wood that is 938 of net worth; the weapons brought 870 (see Market), so the wing ended 68 behind (0.3 of a point). Its reasoning at 25:37 had the rate (7.3 gold a workshop a minute) and judged a further fletcher unprofitable with 7.7 minutes left. The same readouts gave an iron mine about 1.0 iron a minute (23 gold, for 20 wood, two workers and no gold), four times a workshop's yield, and the four quarries an average of 12 stone a minute from 20:58 (85 gold). Gold was short only briefly (450 at 9:22, 336 at 16:39), no call was refused for gold, and the 200 gold at 8:19 and the 300 at 16:02 were each followed by a low-gold warning.

Five of the ten no-arms games built weapon workshops: this one (7 fletchers, 870 gold from sales), Gemini (8 workshops, 270), Luna (7 poleturners, 700), Wonder (2 fletchers, 390) and GLM (5 fletchers, 210); the other four's sales are the tool's counts from their facts, not rechecked here. This wing, 68 behind, compares with about 150 behind for Luna's, 656 for Gemini's and 106 ahead for Wonder's in their analyses.

Beside its other game and the other models

What made the 729 difference in net worth, from both games' results.md: stone +1,288 (5.6 points), flour, bread and wheat +440 (1.9), gold +186 (0.8), wood +46 (0.2), apples −144 (−0.6), iron −667 (−2.9) and the held weapons −420 (−1.8).

• Stone. Quarries at 8:19 (two more at 15:01), with 8 tethers by 11:25 and 13 by 17:28, gave the first stone at 13:24. The other game placed its first quarries at 20:17, had stone from 24:40, and ended with 56.
• Food. A wheat farm, mill and bakery from 15:01 (flour from about 20:58) against none, and a fifth orchard.
• Population. 26 woodcutters (78 wood) as jobs from 20:59 took the wood readout from 19.7 a minute (11 woodcutters, 23:00) to 42 (34) and, with housing, the population from 82 to 112. The other game had 14 woodcutters, fell to 18 people at 16:54 after a popularity crash, and ended at 74.
• Iron. Five mines mined 96, the other game's eleven 134 (125 stored, 9 sold), five of them from 26:55.
• Taxes and spending. 298 more gold from taxes; 420 more spent on purchases.
• Weapons. The rule is worth little here. The 420 held in the other game became 870 of sales from a bigger wing; the wing lost 68 here and 212 there (500 gold and 132 wood for 420), 0.6 of a point in this game's favour. Stone, the food chain and 24 more people dwarf that, and popularity (9 against 38) took 2.9 back.

This game is eighth of ten by net worth and by score: 240 and 3.4 points under Sonnet, above Haiku and GLM. Its popularity is the lowest by 45 points (Luna 54), its population ties Wonder's for third, its stone is fourth and its iron eighth.

What held it back

• The tax ramp at the end. From 30:33 it went to high, extortionate and crueler-than-cruel taxes. Six paydays brought 741 gold (112, 111, 109, 112, 146, 151), 3.2 points of net worth. Popularity fell from 53 (30:34) to 9 and the population from 111 to 98: 4.4 and 0.7 points. On those readings the ramp lost about 1.8 points net (an estimate, not a replay; the score would have been about 29.8). At the score's weights 100 gold is 0.44 points and 10 points of popularity 1.0, so a rise pays only above about 230 gold per 10 points of popularity; the ramp brought 168. Its reasoning at 27:52, 30:23 and 32:59 called popularity "secondary" or "minor"; the second handoff (30:45) said to keep high taxes unless the population collapsed, where the first (14:19) had said not to let popularity fall below 50. The earlier bursts (25:37 and 28:12, 444 gold) were reverted, but popularity had been climbing about 2.7 a minute to 78 and instead sat between 50 and 66 from 26:49 to 30:34; the 28:12 burst ran with half rations (−4) and drew "Popularity fell from 66 to 52".
• Few iron mines. Five. The 0:02 search listed four sites and the 4:03 search, of the view, two; at 5:54 and 9:23 it took that for the deposit's limit ("sites exhausted", "only 2") and did not search again until 17:00, which listed five. It built three of them (17:28, 21:13); (421,345) and (426,344) were never used, and whether they stayed free is not shown. The mines gave about 23 gold a minute each.
• Jobs, not workers. Its interview says it had too few workers; the readings show 20 or more idle peasants in 24 of the 32 readings from 7:50 on, and the 24-idle cap (new peasants stop arriving) from 11:11. Housing was the limit only at 2:29 to 4:01 (18 of 18) and 22:13 to 23:00 (88 and 89 of 90).
• Stone late, then slowing. Quarries at 8:19 gave nothing until 13:24; the model, seeing a large stone pile at a quarry (11:11), added five tethers to the first three. From 29:07 stone stalled: 217 until 30:34 with 167 of room, then 240 at the end, +23 in 4.2 minutes against +96 in the 4 minutes before 29:00. The quarry readout fell from 24 a minute (29:29) to 5.3 (32:06); the logs do not show the cause. At the earlier rate the shortfall is 60 to 80 stone, 430 to 560 of net worth (an estimate).
• A full stockpile at the end. No empty pile from 32:06. Wood and stone reached 336 and 240 at 32:38 and iron 96 at 33:02, exactly 7, 5 and 2 full piles, so the last 42 s of wood and stone and 18 s of iron went nowhere, a few dozen units. At 32:08 it judged the loss small. The three stockpiles at 7:51 were its only expansion.
• Gold that bought little. The chapel (250) left the religion factor at 0 in every reading from 19:23, and the two shrub gardens (60) left worker efficiency at 100% in every reading and showed no garden factor: 310 gold, 1.4 points. Planks bought at 4 and worth 1 lost 630 of net worth; the bread 300.
• The weapons wing, 68 behind (above).

What worked

• Stone. 4 quarries and 13 tethers (the deposit lies 50 tiles from the keep, and the tether text says more distance needs more tethers): 240 in store, 26.6% of net worth, an average of 12 a minute from 20:58 and 24 at best.
• Jobs for the idle cap. 26 woodcutters at 3 wood each took wood from 12 a minute (5 woodcutters, 4:01) to 42, and the population rose from 82 (20:58) to 112 (31:22).
• Earlier output. Iron at 6:34 and stone at 13:24 against 10:41 and 24:40 in the other game; its net worth was back over the do-nothing 1,220 at 10:31, five minutes before there (15:24). Popularity went from 61 (19:23) to 78 (25:35) on no taxes.
• A bread chain from 15:01: 24 flour, 46 bread and 2 wheat (440) in store and two kinds of food eaten at the end (food factor +1). Low taxes from 9:23 brought 798 gold in ten minutes as the population rose from 47 to 73.
• Compaction. Both worked on the first request: 13.6 s ($0.0030) at an estimate of 74,606 tokens and 14.3 s ($0.0031) at 76,269 (limit 120,000), each reading about 96% of its input from the cache. The next calls followed the handoffs (a bread chain and two quarries; bread).
• Placement. Every placement reported as placed is in the final count: 90 from build_structure and 3 stockpile piles from expand_storage. One did not go through (an orchard at 4:26, no_free_spot_nearby).

Market

It bought 210 planks (4 each, worth 1: 20 at 0:56, 50 at 4:03, 100 at 14:22, 20 at 17:14, 20 at 20:59) and 75 bread (8 each, worth 4, at 30:48) for 1,440 gold. The purchases lost 930 of net worth against the sell price (planks 630, bread 300), 14.7% of the final value, 4.1 points. It sold 12 bows (15 each) and 23 crossbows (30 each) for 870 gold, which add to net worth because weapons in store count nothing, and 20 apples (4 each, 80 gold, net worth unchanged).

Checked against the recording (recording/frames.jsonl gold and food before and after each call, with paydays landing inside a call set aside), the 19 calls agree except two. The sales at 16:51 and 30:23 each came back one crossbow short: the recording shows +60 and +90 where the tool read units 1 and 2 and goldChange 30 and 60, and the tool's own gold field agrees with the recording. So it sold 23 crossbows, not 21, and made 870 gold from weapons, not 810. The known 1.7.3 undercount for sales with 200 or more in stock did not arise: the largest stock sold from was 81 apples, and the armory never held more than 10 in a reading.

The bread purchase came when the granary read 33 food, down from 54 at 29:33, about 17 a minute; at that rate it would have emptied near 32:45, 35 s before the end. Afterwards food stayed between 85 and 109. The purchase cost 300 of net worth (1.3 points); the logs do not show what the shortfall would have cost. The 19 market calls took 271 game seconds, 13.6% of the budget.

Cost and speed

54 turns (the 54th was aborted at the deadline and billed nothing), 174 tool calls (61 find_sites, 30 building_info, 19 market_trade, 17 build_structure, 13 inspect_building, 9 set_tax, 8 set_workshop, 6 waits, 11 others), 2.72 million tokens and $0.145 billed by OpenRouter (preparation $0.006, play $0.133, compaction $0.006). The interview cost $0.003 more, with 28,146 of its 33,114 tokens read from the cache. The agent run took 37.4 real minutes, 10.2 of them waiting for the model (about 11.5 s a reply); compaction took 28 s, and 4 fast waits covered 177 game seconds in 85 real ones. 51 of the 53 billed gameplay requests read the prompt cache, 93.2% of their input tokens; requests 2 and 3 (both at 0:02) read none, 22.0 and 23.9 thousand uncached tokens at $0.0076 and $0.0079 against about $0.002 for most others. The logs do not say why.

Of the 2,000 game seconds, build_structure took 919 (17 calls, median 56 s), market_trade 271, waits 192, inspect_building 191, set_workshop 92 (82 s of it the switch to crossbows), set_tax 68 and the host's 29 five-second top-ups 145. That is 10.2 s of game time per building placed by build_structure (90 buildings), also 10.2 with the three stockpile piles (93). The weapons-counted game took 648 s for 55 buildings, 11.8 s. Both ran with the host on the game machine, as did the whole batch. The nine models' weapons-counted games on the same measure were Sol 6.0, Gemini 6.9, Luna 7.3, Haiku 14.5, Wonder 14.9, Sonnet 15.5, Fable 15.8, GLM 18.0 and Opus 18.4; this model's two games sit between those groups.

The interview

It put the gap to the person down to scale (114 housing and 98 people against 202 and 202), bread as an industry (1 wheat farm, 1 mill, 1 bakery against 10, 7 and 57), taxes that "collapsed" popularity to 9, the fear factor (it built none), a stockpile that "filled completely", the weapons wing ("a distraction") and the bread bought at the end. Playing again it would put population first on a real food base, go "all in" on bread, build punishment structures offset by a church and cathedral, avoid heavy taxes, expand storage continuously, and build weapons only once the economy is saturated. It found the read-only stats and tile placement reliable and named four obstacles: the 40-game-second screenshot limit, a stockpile that fills silently, the long delay before a wrong production mix shows, and the minimum turn. It called its play "a little of everything" and said it saw the compounding too late.

Against the game these match: 6,327 against 19,513; 114 housing and 98 people; one wheat farm, mill and bakery; popularity 9; seven fletchers on crossbows; 75 bread for 600 gold and −300 of net worth; taxes raised late with popularity and population down; no empty stockpile pile from 32:06; and weapons that paid back about what they cost. These need qualifying:

• "I simply had too few workers": 20 or more peasants stood idle in 24 of 32 readings; the limits were jobs and the 24-idle cap.
• "I stopped attracting peasants": the population rose to 112 at 31:22 and then fell by 14 as popularity went under 50; peasants left, and 16 housing places stood free at the end.
• "My stockpile filled completely, so wood, stone and iron deliveries stopped. I noticed it late": true of the last 42 s; stone's slowdown began at 29:07 with 167 of room, and the 13 empty piles at 14:18 lasted until 32:06.
• "Build several wheat farms as soon as oasis land is available": wheat sites were listed for three farms at 5:54 and one at 14:19; the orchards and dairy farm took the rest by 6:13, and searches at 9:23, 12:42 and 20:59 found none.
• "The payback was small": the wing returned 870 on 938, 68 behind, 0.3 of a point; the tax ramp and the late, few iron mines cost more.

For the harness

• The granary line. The rules' Start paragraph says to place the required granary, and the preparation reply planned to; the save has one. No call tried: its second reply saw it standing.
• inspect_building on farms. The calls at 3:14 and 3:43 on the orchard (450,389) and the dairy farm (458,400), the tiles build_structure had reported placed, returned "No building panel opened at the target" (2 of 13 calls; 63 s of game time). Both buildings stood in the final count.
• Sales one short. Two market_trade calls (16:51, 30:23) reported one crossbow fewer than the game sold while their own gold field showed the full amount. This is separate from the known 200-or-more undercount; the logs do not show why.
• Iron-mine searches. find_sites for iron answered with the sites in view (two at 4:03) and the model took them for the deposit; the tool description says a search with tile_x/tile_y covers the area around that tile anywhere on the map.
• Storage fills without a wait warning. 13 empty piles at 14:18, none at 32:06; the stockpile "full" message appears in the stats, but wait warnings cover popularity, food and gold only (all three fired in this game).
• The second attempt. The first (87f7cf13) ended after 13 turns on a run error with its budget unused and its final pause not confirmed, scored 1,287 from the last reading and counts for nothing; this attempt ran from the same save and its final pause was confirmed.
• 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 828620").