Claude Fable 5.1

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: Claude Fable 5.1 (claude-fable-5-1) on Anthropic's API, reasoning medium, output limit 32,768 tokens
• Benchmark: The Sinkhole construction (no arms), version 1.7.3 (commit e0e0eb4), from the save "The Sinkhole start"; 45 min of game time at speed 40, the length of the person's game (44 min 58 s); no playbook; waits played fast. Weapons and armour in store count nothing toward net worth; sold at a marketplace, they bring their gold.
• Run 31bbbb10, series 20261019T090704-c98f2ddd, 2026-10-19, host on the game machine: Fable's no-arms game on this map, a fresh series with no memory of its others. Its weapons-counted pair is its 1.7.1 game (run 06cc59ca, series 20261016T110959-db76eefd, commit 07f2b7a), which also ran with the host on the game machine, so the host does not differ in this pair; the version does. Its Sinkhole game from an earlier round is run c849942f (1.6.2). 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 16,099 over leaving the game alone, 40.8% of the person's 39,436, against 10,912 (27.7%) in the 1.7.1 game. By the website's formula (80% net worth, 10% each population and popularity; shares from doing nothing to the person; net-worth share clipped to −1..2) the score is 41.8: 32.66 points from net worth, 2.39 from 124 people and 6.80 from popularity 68. It is the highest of Fable's three Sinkhole games and of the four no-arms Sinkhole games played so far (Haiku 23.8, Sonnet 35.3, Opus 39.3).

Gold is 64.5% of the net worth, iron 23.1% (184 units), stone 7.2% (188) and wheat 4.4% (100). By origin, the 464 iron it mined, 280 of it sold, are worth 10,672, 58% of the net worth, and taxes brought about 4,800 gold. The town ended with 24 idle peasants and no open job, 46 empty places in its hovels, three empty piles in the stockpile, average taxes (99 gold a month) with extra rations, and 27 food.

How the game went

Times are the game clock when a call began, read from the recording; the run clock quoted in observations reads a few seconds later.

Mines first, 0 to 10 minutes. The first build call (0:01) placed two woodcutters and two iron mines; marketplace 0:51; 100 planks (400 gold) 1:27; two mines 2:09; three orchards and a dairy farm 3:06; two mines 4:16. Average taxes at 3:55 (−4 popularity a month) took popularity from 100 at 3:06 to 63 at 10:47. Food fell from 55 to 29 by 4:16; at 5:03 it bought 30 cheese, its reasoning having weighed halving rations against buying. Two mines 7:06, three more with two orchards and two hovels 7:35, 40 apples 9:09, low taxes 10:22, none 10:47. The first iron reached the stockpile at 3:37; 11 mines stood at the 9:09 reading.

Quarries, wheat and a first sale, 10 to 20 minutes. Two mines 10:54, planks 11:40, quarries 12:26 (two placed, one not), four tethers, three hovels and an orchard 13:09, three orchards and a wheat farm 14:54. Gold stood at 766 from 12:26 to 16:46. A food warning at 16:08 (27 for 78 people) was answered at 16:09 by four mines and a sale of 50 iron for 1,150 gold, which paid for 50 apples (17:05) and 75 planks (17:28). A mine, three hovels and two wheat farms 18:03. First stone 17:32; 17 mines at 18:04.

Taxes, a compaction and the last mines, 20 to 30 minutes. Average taxes at 19:40 took popularity from 85 to 59 by 22:52; a third quarry 19:54; quarry searches at 20:39 and 21:17 found no site within 25 tiles; five orchards and two hovels 21:27. The context was compacted at 22:51. Low taxes 22:54, no taxes 24:44 (popularity 51 at 24:41, its low), five mines on the (370, 388) patch 26:26, the 24th mine 27:41, five orchards 28:13. Extra rations at 29:14 (popularity factor +4), average taxes at 29:43, and three hovels at 29:59, the last buildings. Population reached 124 at 29:50.

Waiting and selling, 30 to 45 minutes. A mill beside the stockpile (31:17) found no room, nor did expand_storage (32:38), with one pile empty. It sold iron at 32:55 and 50 wheat at 33:13, bought 50 apples at 34:55 after a food warning, sold planks, iron and wheat at 37:52 to 38:34 for room, waited 200 s from 40:24 after another food warning, and at 43:46 sold 100 planks and 100 stone. Population held at 124, popularity at 64 to 68, and food fell from 110 at 29:44 to 27.

Each column is the last recording frame at or before that minute. Before the end the net worth row counts gold, wood, stone, iron and food (at 4) and leaves out wheat (up to 100 units, 800), so it reads low; the last column is the game's own figure. Iron mines standing is from the last production readout at or before the minute. Iron dips where it was sold (50 at 16:46, 130 at 32:55, 100 at 38:16); with the sales added back, the iron delivered in each five minutes was 3, 17, 29, 42, 60, 65, 85, 80 and 83 (464).

Weapons

It left them alone. No armory or weapon workshop was placed (its 91 buildings are mines, orchards, hovels, wheat farms, tethers, woodcutters, dairy farms, quarries and the marketplace), it never called set_workshop or looked up a weapon building, and it sold no weapons. The preparation reply says to make sure the goods at the end are stored "(not weapons)"; none of its 38 logged reasoning blocks mentions weapons, and the interview mentions them only as the person's.

Its 1.7.1 game planned a wing in the preparation reply ("weapons workshops with an armory to convert wood/iron to higher-value goods"), built an armory and three fletcher's workshops at 12:25 and a fourth workshop at 15:20, and made 30 bows, worth 450 in store at the end (results.json): 3.4% of that game's 13,107 and 0.9 of its 23.1 points. It sold none. The wing cost 400 gold, 85 wood in buildings and 60 in bows (2 each), and four jobs, filled while idle peasants stood at 24 with no open job in 16 of that game's 19 readings from 15:03 to 39:28. Under its own rule it returned 450 against 400 gold and 145 wood: −95 of net worth, or about −530 with the planks at the 4 gold they cost to buy. Bows take wood only, so no iron went into weapons in either game. Under this game's rule the 1.7.1 game would have ended at 12,657 and 22.1 points.

In this game the wood went into its 91 buildings (the 24 mines alone took 480) and the 400 gold stayed in the treasury; it mined 464 iron against 336, and the jobs were mines, orchards, wheat farms and quarries. The 24 idle peasants stood as before.

Beside its other games and the other models

Mined, kept and sold together: iron 464 against 336, stone 288 against 220, wheat 200 against none.

The pair differs in more than weapons: the version (1.7.2 retries a blocked build_structure target only within its footprint's side plus a tile, where 1.7.1 searched 12, and keeps room beside the stockpile; 1.7.3 added this benchmark and fixed a compaction retry that did not come up here) and the model's own choices. Here 5 of 90 build_structure placements were not placed and none was moved; the 1.7.1 game had 74 of 75 placed (a hovel moved by 4 and 3 tiles, an orchard by 11 and 11) and one not placed.

The score is 18.8 points higher: net worth +10.5, popularity +6.8, population +1.5. Scored by this game's rule the 1.7.1 game ends at 12,657, so the gap is 5,637 and 19.7 points; the weapons are 450 of it. What the two games' files show:

• The end. The 1.7.1 game had popularity 91 and 104 people at 32:42, set average taxes at 32:36 on full rations, and ended at 0 and 50 with its food gone (2 at 39:28, the granary at 0 from 39:11); holding its 32:42 position would have scored about 33. This game ended at 68 and 124: 24 orchards and 3 dairy farms where that one had 12 and 2 (apples 39 to 48 a minute in its readouts from 29:44, against 21 to 26 there), and extra rations (food +4) set against average taxes (−4). Food never fell under 22 (15:32).
• Iron, earlier. 11 mines at 9:09 and 17 at 18:04, against 5 from 6:52 to 18:26 there (10 at 19:38, 17 at 29:53, 25 at 33:49). The first 20 minutes delivered 91 iron against 49, the game 464 against 336 (+2,944). It bought 300 planks by 12:26 against 175 by 9:30 there, where 400 gold went into workshops from 12:25; here the sale of 50 iron at 16:46 paid for apples and planks.
• Taxes, +2,270. About 3,400 of the 4,803 came after 29:43, on average taxes with extra rations.
• Wheat and stone. Six wheat farms (90 wood) gave 200 wheat, 1,600 of net worth; the 1.7.1 game built none. Three quarries delivered 288 stone (first at 17:32) against 220 (18:30).
• Buying cost more: 2,015 of net worth against 765.

Both games stood at 24 idle peasants with no open job, and both stockpiles filled.

In the 1.7.1 round (weapons counted) Gemini 4 Argon scored 52.4 and GPT-6.1 Sol 50.4, ahead of Opus 34.7, Sonnet 33.1 and Muse Wonder 33.0; this game's 41.8 would stand third of eleven. The no-arms Sinkhole games so far, by net worth and score: Claude Haiku 5.5 8,243 and 23.8, Claude Sonnet 5.5 15,246 and 35.3, Claude Opus 5.5 16,889 and 39.3, and this game. On the batch's other maps Fable's no-arms game also beat its weapons-counted game (Oasis by the Sea 40.5 against 37.5, Large Barren Desert 11.6 against 1.3; those pairs share a version).

What held it back

• Jobs, not housing. Idle peasants stood at 24 with no open job in 19 of the 24 readings from 18:04 (23, 24 with one job open, 17, 18 and 22 in the others). The five mines of 26:26 took idle from 24 to 17 at 27:42, and population rose from 105 to 124 as it refilled. After the last building (29:59) population stayed at 124 for 15 minutes with housing at 170; its reasoning at 29:59 said more hovels would not help, and it placed three. Searches for job sites came back empty: quarries (20:39, 21:17), orchards (22:51, 25:00, 25:54; at 29:59 one site it judged too far), wheat (29:59), dairy (28:08), iron (27:41, one site). The ore at (341, 575) and (289, 277), 130 and 179 tiles from the keep, was never tried.
• Nothing built for the last 15 minutes while gold rose from 1,847 to 11,803. The mill and expand_storage failed for room (31:17, 32:38); at 33:31 its reasoning judged farms 110 or more tiles out not worth the time left.
• Buying. 3,140 gold lost 2,015 of net worth (4.1 points): planks 1,335, food 680. Four woodcutters delivered 16 wood a minute.
• Taxes against popularity. Two average spells without extra rations took popularity from 100 to 63 (3:55 to 10:47, about 420 gold) and from 85 to 57 (19:40 to 23:15, about 510 gold).
• Storage. One pile was empty at 32:39 and 37:52, none at 43:46; stone sat at 288 with room 0 from 42:36 to 44:22 (about 19 stone not mined, about 130 of net worth).
• A doubt after the compaction. The handoff counted 16 orchards and about 14 hovels, which already held the seven buildings of 21:27. At 22:54 and 23:14 its reasoning took the 16 orchards and the unchanged wood (49) to mean that the seven had not registered; it called status (16 orchards, 15 hovels) and get_inventory and placed another set at new sites (one orchard and two hovels placed, two orchards not; 86 s). The recording shows wood falling in five steps of 5 (21:48 to 22:22) and two of 6 (22:39, 22:45), the batch's 37, and every building type's final count equals what it placed. Nothing was lost, but the interview repeats the claim.
• Plan and notebook unused. update_plan and notebook_write were never called, so only its own handoff crossed the compaction.

What worked

• Iron. Four mines placed by 2:33, 13 standing by 11:40, 24 by 28:14; 464 iron, 58% of the net worth, at about 0.7 a minute a mine; 191 of it came in the last 11.8 minutes.
• Selling to keep the mines running. With iron room 77 and one empty pile at 32:39, at 17 iron a minute, the stockpile had under five minutes left; the sales freed piles at no cost in net worth.
• Popularity. Never under 51; extra rations at 29:14 paid for average taxes (about 3,400 gold); it backed out of two tax spells at 10:22 and 22:54 on the factors, and each food warning was answered or weighed.
• Placements held: all 85 placed are in the final count.

Market

Bought: 445 planks in five trades (1:27, 6:05, 11:40, 17:28, 25:54; 1,780 gold), 30 cheese (240) and 140 apples (9:09, 17:05, 34:55; 1,120): 3,140 gold, 2,015 of net worth lost (planks 1,335, cheese 120, apples 560). A plank purchase at 4:45 reported not_traded; the recording shows no gold spent. Sold: 280 iron (50, 130, 100) for 6,440, 100 wheat for 800, 200 planks for 200 and 100 stone for 700: 8,140 in all. The replies add to 7,030 (250 iron, 40 stone). Two calls differ, the known 1.7.3 fault at 200 or more in stock: iron at 32:55 (215 in stock; gold rose 1,150, then four times 460, so 130 units, 30 more than asked) and stone at 44:14 (288 in stock; two rises of 350, so 100 units, 50 more than asked). Their stock and gold were right; the other sales (stock under 200) match the recording. Selling left net worth unchanged. None of its 120 tool calls was a game_action, so it set no automatic trade by clicks, and every fall of gold in the recording sits in a purchase's clicks (3,040 of the 3,140 show as separate steps; three of 20, 40 and 40 fall in frames where gold also rose). The 18 calls took 452 game seconds for 175 clicks, 2.6 a click.

Cost and speed

44 turns, 120 tool calls (37 find_sites, 24 build_structure, 19 waits, 18 market_trade, 8 set_tax, 5 place_near, 9 others), 2.60 million tokens and $3.331 at the profile's prices ($10 per million input, $50 output, $0.25 cache read, $12.50 cache write): preparation $0.404, play $2.821, compaction $0.105. The 1.7.1 game cost $2.751: this one wrote more (22,107 output tokens against 16,005; 89,675 written to the cache against 68,221) in fewer turns (44 against 48). The interview cost about $0.110. The agent run took 38.9 real minutes, 6.9 of them waiting for the model. All 44 requests read the prompt cache (96.4% of input tokens) and none failed. The context was compacted once (22:51, 22 seconds), in one request that succeeded on the first. Nine waits were played fast and saved 4.2 real minutes (three ended early on warnings); the 1.7.1 game's eight saved 1.8.

Of the 2,700 game seconds the 24 build calls took 1,154, the 19 waits 730 and the 18 market calls 452. One build call (5:22) was refused for asking for more than 8 actions in a reply; the other 23 placed 85 buildings: 13.6 game seconds each, 3.7 a call (place_near added 6 in 87 s). The 1.7.1 game's 15 calls took 1,339 s for 74 (18.1 each, 4.9 a call). The host was the game machine in both, so the 4.5 s difference mixes the version, the mix of buildings and how many a call carried, and says nothing by itself. It saw five screenshots (the first, two map_overview, one center_on, one after the compaction).

The interview

Fable put the gap to the person down to a fixed-size engine of iron mines and taxes against a scalable bread chain, to idle peasants it had no jobs for, and to spending popularity on taxes where the person spent it on cheaper rations. Playing again it would build the wheat, mill and bakery chain early, add inns and a chapel, keep taxes at none, reserve ground around the stockpile and verify each batch of placements. It praised the tile tools, status, wait_and_observe and market_trade's net-worth change; it disliked a batch of placements it thought wrongly reported, no warning that the stockpile was closing in, and the memory lost in the compaction.

Against the game these match: 18,294 against 41,631 and 124 against 506 people (4.1 times); about 24 mines and taxes at 4 to 5; 24 idle peasants and population stuck at 124; quarry and iron sites used up; expand_storage finding no room and iron sold to keep the mines delivering; apples bought late (34:55); bread at 4 against wheat at 8.

These do not. "An early batch of placements that the harness reported as placed but never appeared (wood wasn't deducted)": the batch of seven was at 21:27, and all of it stands (see above), as in its 1.7.1 interview. "My economy stopped growing around the 30-minute mark": building and population did, but net worth went from about 7,700 (7,118 in the table, without wheat) to 18,294 after it. "A stockpile boxed in by hovels": the result says no free ground touches it and does not name what is there, so this is unchecked. "Never seriously considered cutting rations": it weighed halving them at 4:16 and bought cheese instead. It named wait_and_observe with until among the tools that helped; it used until once (33:31, food at least 0, met at once). Selling to make room, which it counted among its mistakes, lost nothing.

For the harness

• The market_trade reply undercounted two sales (iron at 32:55: 100 units and 2,300 gold against 130 and 2,990; stone at 44:14: 40 and 280 against 100 and 700, with a note saying it stopped for lack of units after selling 50 more than asked): the known 1.7.3 fault. stock and gold were right, its later reasoning uses those, and scoring is unaffected.
• No room beside the stockpile at 31:17 and 32:38, although 1.7.2 keeps a spot beside it: no placement reported blocks_stockpile, and the result does not say which building took the room. No warning came before the last pile filled.
• Placements: five not placed (5:28, 12:26, 14:54, 23:14 twice), each no_free_spot_nearby with no nearestFit; two place_near hovels stopped short (1:06 for 3 wood, 6:50 for room); one call refused at 5:22 for more than 8 actions, re-sent at 5:28. After the compaction it doubted placements that had held, as in the 1.7.1 game.
• Wait warnings: three for popularity (7:33, 10:47, 22:51) and three for food (16:08, 34:55, 40:22); the last two cut waits to 75 of 150 and 90 of 120 game seconds.
• No provider errors, and no forced garbage collection in 457 memory samples. The game saved at the end ("claude fable 5-1 1-7-3 31bbbb").