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, version 1.7.1 (commit 07f2b7a), from the save the human reference game on this map started from; 45 min of game time at speed 40, the length of the person's game (44 min 58 s); no playbook; waits played fast
• Run 06cc59ca, series 20261016T110959-db76eefd, 2026-10-16: Fable's game in The Sinkhole's 1.7.1 round and the first of its four games that day, a fresh series with no memory of its others. Its 1.6.2 game on this map had not finished when this was written.

Result

A full-budget score: the run used all 2,700 game seconds. Net worth grew by 10,912 over leaving the game alone, 28% of the person's 39,436. The score is the project website's formula (80% net worth, 10% each population and popularity); the game counts on the website's leaderboard. It is made of 22.1 points for net worth, 0.9 for population and none for popularity, and is eighth of the ten games of the 1.7.1 round (Gemini 4 Argon 52.4 first, GLM 5.3 Flash 15.1 last).

Of the 13,107, 7,728 (59%) is 336 iron from 25 iron mines, 3,133 is gold, 1,540 is 220 stone from three quarries, 450 is 30 bows from four fletcher's workshops, 184 is wood and 72 is apples. Iron alone is worth 15.7 of the 23.1 points. The gold is the 2,000 it started with, less 1,420 spent (1,020 on planks, 400 on four fletcher's workshops), plus 2,533 from taxes and 20 from a sale. The town ended with no idle peasants, 37 of its 87 jobs open and 88 empty places in its hovels.

How the game went

Opening, 0 to 7 minutes. map_overview and four find_sites took the first 3 game seconds; the first build call, at 0:03, placed a marketplace, two woodcutters, two hovels and an iron mine and took 80 game seconds. It bought 100 planks at 1:28 (400 gold, 54 game seconds for 20 clicks). Two iron mines, two apple orchards, a dairy farm, a hovel and a woodcutter followed at 2:29, and two mines, three orchards (one not placed) and a hovel at 4:41. Food fell from 55 meat to 16 by 6:52 with 36 people in town, and it set half rations (a popularity factor of −4).

Orchards, quarries and weapons, 7 to 15 minutes. Four orchards and two woodcutters at 7:13 with 75 more planks bought, then three hovels, a dairy farm, a quarry and two ox tethers at 10:06. Popularity fell from 100 to 76 on half rations, and at 12:15 it set full rations again. At 12:33 it placed an armory, three fletcher's workshops (100 gold each), a second quarry with a tether and a hovel. No iron mine was placed between 4:41 and 18:28. Gold stood at 1,000 at 15:03 and stayed between 1,000 and 1,450 until 32:42.

Low taxes, mines and housing, 15 to 33 minutes. Low taxes ran from 15:03 to 21:11 (about 550 gold; popularity 83 to 61) before it set them back to none. At 15:29 it placed four more orchards, a fourth fletcher's workshop and a hovel; the searches that found those orchard sites returned one or two each and a dairy-farm search at 21:11 returned none, and it looked for no more farm sites. Iron mines then came in batches: five at 18:28, three at 21:35, four at 25:41, five at 30:33, two at 31:57 and one at 32:42, 25 in all. The first iron reached the stockpile at 4:17 and the first stone at 18:37. The population reached its housing limit of 82 at 22:49; five hovels placed at 23:18 raised the housing to 122, but the population stayed at 82 until 26:13 and at 91 from 27:58 to 31:00, with 24 idle peasants and no open job in every reading; it grew again at 26:13 and 31:00, soon after the build calls of 25:41 and 30:33. The context was compacted at 29:53.

Taxes up, food out, 32 to 45 minutes. The last build call was at 33:52. Readings from 32:42 on:

At 32:42, with popularity 91 and 60 food, it set average taxes: 84 gold a month, −4 popularity a month against +1 with none. Its logged reasoning counted the food at about eight minutes' supply, judged buying more not worth the cost and took popularity as able to absorb the loss. The "Food is low" warning came at 35:27 (38 food for 111 people), when it went back to low taxes. At 37:05, with 25 food, it wrote down consumption at about 34 a minute against about 26 produced, noted that running out would cost more popularity than half rations, and waited two minutes without changing rations; they stayed full from 12:15 to the end. The granary count first read 0 at about 39:19 in the recording and read 0 on and off, in six spells of about 45 seconds in all, until 42:45. At 39:28, with popularity 50 and 2 food, it set average taxes again, writing that the shortage would crash popularity anyway and that the 24 idle peasants would leave first.

Popularity went under 50 at about 39:22 and the first peasant left at 39:57; the idle peasants were gone by 43:41, workers were leaving after them, and 61 people had left by the end. Every factor the reports listed from 33:49 was tax (−4 at average, −2 at low) but for a food-variety +1 at 37:05; no food penalty appeared, though the recording shows the steepest falls (8 points in 23 seconds from 39:19, 13 in 28 seconds from 41:00) while the count read 0, against about one point every 7 seconds at −4 a month. Nothing was built after 33:52: from the reading of 35:27 to the end wood went from 29 to 184 (192 before the sale), gold from 1,571 to 3,133 and stone from 140 to 220.

Each column is the last reading at or before that minute. Before the end the net worth row counts gold, wood, stone, iron and food (at 4) as the recording logs them; bows (450 at the end) are left out, so it reads low. Iron mines standing is from the last production readout at or before the minute. The iron delivered in each five minutes was 1, 13, 19, 16, 25, 38, 57, 85 and 82.

Beside the other models

The ten games of the 1.7.1 round, by score:

Fable's net worth is fifth of the ten; its score is eighth because population and popularity gave 0.9 of a possible 20 points. Six games (Haiku, Luna, Sonnet, Fable, Opus and Muse Wonder) built 17 to 27 iron mines and ended between 11,100 and 15,400. Sonnet 5.5 ended with 17 mines, 315 iron, 100 people and popularity 100: 478 less net worth than Fable and 10.0 points more score. The two best games took another route: Gemini and Sol built 57 and 45 bakeries, ran double rations with 419 and 619 food in store, and ended on mean taxes bringing 258 and 217 gold a month (Fable's last reading: 40), with 15,929 and 13,934 gold against Fable's 3,133. Fable's 1.6.2 game on this map was still running, so there is no second game to set it beside.

What held it back

• Taxes raised while the food ran short. From 32:42 taxes brought 1,985 gold, none of it spent, worth 4.0 points of score. Popularity stood at 91 and the population at 104 then, worth 9.1 and 2.0 points on the formula; at the end they were worth 0 and 0.9, a fall of 10.2 points if they had held. The reports put the monthly change down to tax; the empty granary's own effect is not itemised. At the formula's weights one point of popularity is worth about 49 gold of net worth, and average taxes cost 5 points a month against none (about 245 gold's worth) for 84 to 88 gold.
• Food short from about 35 minutes. Twelve orchards and two dairy farms gave about 26 food a minute; 111 people on full rations ate about 34. The stock fell 7 or 8 a minute from 29:53 (79) to 37:05 (25), and the one lever it had used before, half rations at 6:52, was not used again.
• Mines late, then none. Five mines had been placed by 4:41 and no more until 18:28, while orchards, hovels, quarries and weapon workshops took the wood and gold stood between 1,000 and 1,450; four waits (19:38, 21:35, 31:57, 32:42) were for 40 to 50 planks for its next builds. Each mine delivered about 0.7 iron a minute once running (16 net worth), and two thirds of the 336 iron (224) arrived after 30 minutes. After 33:52 it placed nothing: find_sites near (370, 380) had found no ore, and the one listed deposit it had not tried, at (341, 575), is 130 tiles from the keep.
• 24 idle peasants and no open job in 17 of the 20 readings from 15:03 to 39:28 (17, 18 and 21 in the others); the population stood at 82, 91 and 111 in steps, with 27 free places in the hovels at its peak of 111. It built no inn, chapel or church; its logged reasoning weighed a chapel at 19:38 and 23:15 and left it.
• Plan and notebook unused. update_plan and notebook_write were never called, so only its own handoff at 29:53 crossed the compaction. The opening plan it wrote before play said to expand stockpiles before they fill and to keep building to the final minute; it did neither.
• Storage. The 16 stockpile piles were full at 43:41 (empty piles: 7 at 35:27, 5 at 37:05, 3 at 39:28, 2 at 41:54). It sold 20 planks to make room for iron, but a pile holds one good, so iron stopped at 336 (seven full piles) at about 44:36, some 7 iron (about 160 net worth) short of what the mines would have added.

What worked

• Iron. 25 mines delivered 336 iron, 59% of the net worth and 71% of its growth over leaving the game alone. At the last reading (43:41) they were delivering 17 a minute, worth 391 net worth a minute, against 75 a minute from the three quarries (about 25 each, more than a mine's 16) and 15 from the four fletcher's workshops (under 4 each).
• Backing out of two mistakes. It read the factors and set rations back to full at 12:15 (popularity 76) and taxes back to none at 21:11 (popularity 61); popularity then climbed to 91 by 32:42 with the population at 104.
• Placement. Of 75 placements, 72 were placed directly, 2 by the tile-map retry (a hovel moved 4 and 3 tiles, an orchard 11 and 11) and 1 was not placed (an orchard at 4:41, "nothing fits at or near that spot" with a remedy). The final count of every building type equals what it placed. It worked from find_sites tile targets and saw four screenshots in the whole game (the first, map_overview, one observe and the one after the compaction).
• Little lost at the market. It bought 255 planks and sold only 20 (below).

Market

255 planks bought in three trades (100 called at 1:28, 75 at 7:13, 80 at 23:18) for 1,020 gold, worth 255 at the sell price: 765 net worth lost (1.6 points). 20 planks sold at 43:41 for 20 gold, which leaves net worth unchanged. Nothing else was traded. A market click took about 2.7 game seconds (20 clicks, 54 seconds). These are units at the market's fixed prices.

Cost and speed

48 turns, 75 tool calls (28 find_sites, 15 build_structure, 14 wait_and_observe, 5 set_tax, 4 market_trade, 3 building_info, 2 set_rations, and one each of map_overview, inspect_building, observe and status), 2.59 million tokens and $2.751 at the profile's prices (preparation $0.400, play $2.254, compaction $0.097), the highest of the ten games' logged costs. Play was 16,005 output tokens ($0.80), 2.40 million tokens read from the cache ($0.60) and 68,221 written to it ($0.85). The interview cost $0.102 more. The agent run took 38.4 real minutes, 5.6 of them waiting for the model; the context was compacted once (20 seconds). Eight waits were played fast and saved 1.8 real minutes net. The 14 waits used 822 game seconds, 30% of the 45 minutes; the rest passed in tool calls: the seven build calls made on their own took 65 to 144 game seconds each for 5 to 7 placements (13 to 22 game seconds a placement), and the host let 5 game seconds pass after each of 15 turns that took no screenshot (75 seconds).

Providers and the prompt cache

Every request went to Anthropic's API, none failed, and all 48 gameplay requests read the prompt cache, 97.2% of the input tokens. The interview request came 129.5 seconds after the last gameplay reply, inside the cache's 5 minutes, and read 53,670 of its 56,586 tokens from the cache: the whole prompt of the last gameplay request (52,471 read and 1,199 written), with 1,521 tokens new.

The interview

Fable put the gap to the person down to scale and to what it did with the last ten minutes: a raw material economy of 25 iron mines, 3 quarries and 12 orchards that it said was "capped by how much ore and oasis grass" it could reach, stockpiled rather than sold, against a bread chain that sold continuously. It called its ending a judgment error: it let food run out while taxing, and popularity fell to 0 and peasants left. Playing again it would copy the person's core (wheat farm, mill, bakery and a sell loop), build inns with ale and a church for popularity, use the fear factor, give the idle peasants jobs, keep taxes at none and popularity away from 50, and sell more. It praised find_sites and build_structure by tile, the stats JSON and the wait warnings, and complained about one placement report, the game seconds a market click takes and the compaction. It asked for a reminder in the rules that selling costs no net worth and that bread and ale chains scale with population.

Several claims do not match the run:

• It said its tax experiment "earned about 1,500 gold". Taxes brought 2,533 in all: 548 from low taxes at 15:03, and 1,985 from 32:42 (531 at the first average spell, 594 at low taxes, 860 from 39:28).
• It said the "compound effects (no food −8, taxes −4)" took popularity to zero. No reading lists a food factor of −8, or any food penalty beyond the half rations of 6:52 to 12:15 (−4 at the most); the factors from 33:49 are tax, with a food-variety +1 at 37:05.
• It said one build_structure call reported seven placements as placed with own_buildings evidence "yet wood never dropped and the building count did not change". That is the call of 25:41 (four mines, a woodcutter and two hovels). status at 30:02 lists 17 mines, 6 woodcutters and 16 hovels, which is the earlier 13, 5 and 14 plus those seven. The handoff, written from the reading seconds before, already listed 17 mines and 6 woodcutters, so the two agreed. Wood stood at 126 before the call and 105 after: the call cost 95, and the woodcutters delivered about 20 a minute meanwhile.
• "Capped by how much ore and oasis grass I could reach" is half true. The ore near the keep was used (searches at 30:33 and 33:49 found none), but a listed deposit was untried. map_overview listed 26,436 tiles of farmland from x 276 to 536 and y 264 to 545; all eleven of its orchard and dairy searches were centred between x 300 and 350 and y 410 and 515, and it did not search the east or the north of that.
• It asked for a reminder that selling costs no net worth. The benchmark rules it was given say it: "Selling turns goods into the same value in gold, so it leaves net worth unchanged", and its handoff at 29:53 says not to buy because buying lowers net worth.

The rest checks out: 25 iron mines, 3 quarries, 12 orchards; 9,974 in stored goods; bread selling at 4; 24 idle peasants for most of the run; no inn built; and a full stockpile stopping production, which it noticed at 43:41 with 79 game seconds left.

For the harness

• Fable's games were the first played with the host running on the game machine itself; that changes nothing the game plays.
• Wait warnings worked as designed: seven for popularity falling by 10 or more (10:04, 12:15, 35:27, 39:28, 40:46, 41:54, 43:41) and one for food (35:27). Three ended waits early, at 64 of 120, 65 of 150 and 105 of 190 game seconds. No warning or game_events entry came for storage filling, the granary reading 0 or people leaving; game_events was empty in every reading.
• status leaves out the tax level that every observation carries (tax_level).
• After the compaction it doubted its own correct placements (above); that cost a status call and a 20-second wait, 37 game seconds from 29:53, before it looked for ore again.
• The "only waited and read for 3 game minutes" note came twice (39:28 and 43:41), after waits of 143 and 107 game seconds.
• Placement: 4:41 an orchard at (339, 465) found no free spot nearby and was not placed; the tile-map retry moved a hovel at 2:29 and an orchard at 7:13. No refused calls or stalled replies.
• A little over the budget at the end. From the tick the save was loaded at to the final reading, the game ran 2,712.8 game seconds for a budget of 2,700. The round's other full games ran 5.5 to 8 seconds over, and games of earlier versions, which play every wait at speed 40, 4 to 5 (the time between loading and the first pause). This game ended inside a wait played fast, at speed 90, where the final pause takes hold later. The extra seconds of iron are worth about 0.1 of a point.
• The game saved correctly at the end ("claude fable 5-1 1-7-1 06cc59", in this folder).