One episode · Large Barren Desert

Gemini 4 Argon

Benchmark 1.7.3GoogleReasoning mediumPlayed by the maintainersFull-budget score
Model
google/gemini-4-argon
Served via
OpenRouter · google-ai-studio
Max output
32,768 tokens
Played
18 October 2026
Score
63.5
Crusader Bench score on Large Barren Desert: 0 is doing nothing, 100 the person’s game
Net worth
8,369
+3,177 over doing nothing, 66% of the person’s gain of 4,792
Population
134
53% of the person’s gain of 246 · popularity 52
Cost
$1.45
2.1M tokens
Time
33 game min
46 min of wall-clock time, 39 turns

Minute by Minute

Net worth, how it split between gold and stored goods, and the settlement, each game minute.

Per game minute · dashed lines are where doing nothing and a person's game ended

  • Gemini 4 Argon
  • Baseline (do nothing), final
  • A person's game, final

Crusader Bench · crusbench.com

Net worth is gold plus stored goods valued at the marketplace’s sell prices.

Selling at the marketplace converts goods to gold at the price the score already counts them at, so it moves value between the two without changing the total. Buying costs net worth.

Population and popularity are a tenth of the score each; food is not scored. Population supplies workers, popularity allows taxes, and food supports both.

The episode

Watch the Episode

Videos and images stream from the dataset. Game footage © Firefly Studios, shared for non-commercial research.

Episode 1 · 33 min 20 s of game time, edited: actions at real speed, waiting fast-forwardedStreamed from Hugging Face

ScorecardFull-budget score

8,369net worth
Gold
1,089
Stored goods, at sell price
7,280
Over the baseline (do nothing)
+3,177
Population / housing
134 / 146
Popularity
52
Food
1,016
Structures (map-wide)
143
Attempt
1

The settlement at the endcentred on the keep, zoomed out

The settlement at the end of episode 1, zoomed out over the keep

Final stock7,280 gold in goods

GoodUnitsPriceValue
Gold1,08911,089
Bread1,01644,064
Stone18071,260
Metal armour41301,230
Iron2823644
Ale81080
Wood212
Net worth8,369

Beside a person’s gameLarge Barren Desert, the same game time

This episodeThe personLeft alone
Net worth8,3699,9845,192
Gold1,0898,574
Stored goods, at the sell price7,2801,410
Population134250
Housing146250
Popularity52100

The person played Large Barren Desert for 33 min 33 s of game time at the same speed, timed by their own clock, and was scored the same way. The model saw this comparison only in its interview.

The town’s booksread from the game at the end

Date in the game
July 1178
Idle peasants
24
Popularity next month
−1
Food
−8
Taxes
−2
Religion
+1
Ale from the inn
+8
Taxes
Low taxes, 80 gold a month
Rations
none; 0 of 1 foods eaten
Worker efficiency
100%
Storage
16 stockpile piles, 9 empty; 6 granaries holding 1,016 food

Buildings113

  • Bakery22
  • Ox Tether19
  • Hovel17
  • Quarry11
  • Gallows7
  • Granary6
  • Inn5
  • Armorer's Workshop4
  • Brewery4
  • Iron Mine4
  • Mill4
  • Stockpile4
  • Woodcutter2
  • Armory1
  • Church1
  • Keep1
  • Marketplace1

Tool calls204

  • find_sites89
  • market_trade31
  • building_info24
  • build_structure19
  • wait_and_observe14
  • set_tax10
  • observe8
  • update_plan4
  • expand_storage2
  • map_overview1
  • set_rations1
  • status1

Play

Turns
39
Buildings placed / attempted
105 / 106
Anchor-tool placements
5
Tool errors
1
Raw inputs sent to the game
853
Context compactions
2
Game time
33 min

Compute

Tokens
2.1M
Served from cache
83%
Output tokens
60K
Cost, as billed
$1.45
Wall-clock time
46 min
of which inference
18 min
Peak game memory
3,530 MiB

The plan, before play

Understanding of the Task

  • Goal: Maximize net worth (gold + stored goods valued at marketplace sell prices), followed by population and popularity, within 33m 20s (2000 game seconds) on the Large Barren Desert map.
  • Map & Constraints:
    • Granary and Stockpile already exist.
    • Starting package (5,000 gold, 55 meat, ~245 wood, 50 stone, 15 iron, 20 pitch, 20 ale) arrives over the first ~27 seconds, quickly filling the initial stockpile.
    • No oasis grass on the map (no farms possible). Food and inputs like wheat/flour or hops must be managed via rations, marketplace imports, and processing (e.g., importing wheat/flour for bakeries, or importing cheap food varieties to boost popularity/taxes).
    • Resources on map: Iron ore strips ~60 tiles east of the keep; stone deposits ~100 tiles west of the keep; trees for woodcutters.
    • Forbidden: Recruiting troops or building walls, towers, gates, or defenses.

Opening Plan

  1. Survey & Expand Storage: Read the map with map_overview and find_sites while paused (zero game-time cost). Immediately expand the Stockpile so the full starting package can be delivered without blocking, and place a Marketplace and Hovels.
  2. Resource Extraction: Establish Woodcutters near trees, Iron Mines on the eastern iron ore deposits (high value at 23g/unit and input for swords/maces/metal armor), and Quarries + Ox Tethers on the western stone deposits.
  3. High-Value Manufacturing & Taxes: Build an Armory and weapon workshops (Blacksmiths, Armorers, Fletchers/Crossbows) near the keep/stockpile to convert raw iron and wood into high-value equipment, while using Inns, religion, and food/bakery management to sustain high popularity and tax income.

Commentary during play2

  1. дели

Reading the game

The Analysis

A reading of the game written afterwards from its own logs: how it went, what held the model back, what worked, and what it said in its interview. It is published with the game and is not part of its score.

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: Gemini 4 Argon (google/gemini-4-argon) on OpenRouter, served only by Google AI Studio (no fallbacks), reasoning medium, output limit 32,768 tokens
  • Benchmark: Large Barren Desert construction, version 1.7.3 (commit e0e0eb4, the relock with the compaction-retry fix, which leaves 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, as since 1.6.0.
  • Run 0996eaea, series 20261018T155704-115a5a74, 2026-10-18: Gemini's weapons-counted game on this map with version 1.7.3, a fresh series with no memory of its other games. Its no-arms game is still playing; its desert games from earlier rounds are run 956efdbe (1.5.6) and run 0dd24029 (1.6.3). Three earlier attempts at this game do not count: one failed on 2026-10-17 on an OpenRouter key limit, and two on 2026-10-18 (series 20261018T145652-b2f3dc25 and 20261018T150655-bfa2863d, both invalid) were stopped by the operator's cost guard, a script outside the harness that watches the prompt-cache share and misread normal cache behaviour (one early miss, judged at request 8; then misses on the first request after compactions at 1,972 of 2,000 game seconds); neither was a fault of the model or the game.

Result

This game (1.7.3) Its 1.6.3 game Its 1.5.6 game The person Left alone
Net worth 8,369 7,494 8,253 9,984 5,192
Gold 1,089 4,612 1,327 8,574
Stored goods, at the sell price 7,280 2,882 6,926 1,410
of which weapons and armour 1,230 990 0
Population / housing 134 / 146 153 / 154 185 / 194 250 / 250 4
Popularity 52 (−1 next month) 68 (−6) 63 (−8) 100 0
Rations, taxes at the end none, low (80 a month) full, extortionate (214) half, extortionate (258)
Score 63.5 51.3 64.8 100 0

A full-budget score: the run used all 2,000 game seconds. Net worth grew by 3,177 over leaving the game alone, 66% of the person's 4,792, against 48% and 64% in the earlier games. By the website's formula (80% net worth, 10% each population and popularity, the net-worth share clipped to −1..2) the score is 53.04 + 5.28 + 5.20 = 63.5, 12.2 above the 1.6.3 game and 1.2 below the 1.5.6 game; this game counts on the website's leaderboard. The stored goods are 1,016 bread (4,064), 180 stone (1,260), 41 metal armour (1,230), 28 iron (644), 8 ale (80) and 2 wood; bread is 49% of net worth. The town ended with 24 idle peasants, 12 free places in its hovels, rations at none and popularity 52.

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 second or two later). A build call keeps the game running for 20 to 97 seconds.

Iron and stone, 0 to 7 minutes. map_overview (0:01) listed iron 57 to 69 tiles east of the keep, stone 83 to 106 tiles west, no farmland and no trees. The first build call (0:01) placed two stockpile piles, the marketplace, two hovels and three iron mines; a fourth mine and four quarries followed at 1:41, more quarries and ox tethers at 3:15, 5:07 and 6:09, and 100 planks bought at 3:49 and 4:27. By 6:57 it had 11 quarries and 9 tethers; its last find_sites for iron (3:15) and quarries (6:59) found nothing more within 25 tiles. Iron arrived at 4:30, stone at 8:05.

Rations: when and why. An observation at 5:06 warned "Food is low: 12 in the granaries for 42 people on full rations"; the meat was all the food and the map has no farmland. The replies at 5:07 reasoned that rations at none would save the meat, that peasants arrive while popularity is above 50, and that inns and a church could hold it there. It set rations to none at 6:02 (food factor −8); the meat was gone by 6:08 and it never bought meat, cheese or apples. Popularity fell from 98 (6:00) to 50 (9:08, its low), with warnings at 6:58 and 8:48. It answered with three inns (6:59, 8:54, 9:49), a large and then a generous bribe (8:49, 9:45), eight bakeries, and 25 flour and 25 wheat. Bread first reached a granary at 11:29. It reconsidered rations at 12:11, 15:09 and 16:59, and its 19:20 hand-off put the bread kept at about 120 gold a minute; the food factor read −8 in every reading to the end. (The controls it read say to keep rations at least normal while popularity is near 50.)

Armorers and gold, 8 to 19 minutes. The armory went up at 8:54, armorer's workshops at 9:49 and 11:11. Gold read 671 at 12:11, 382 at 13:27 ("Gold is low") and 0 at 18:34. It sold stone (50 at 13:28, 100 at 15:09) and iron (30 at 17:21) to pay for planks, wheat, seven gallows (13:37, 350 gold), a church (500) and two inns (15:50). A granary was rejected for not touching the old one; three came by expand_storage at 17:04. Taxes were a small bribe from 16:59 to 19:23.

Bread and taxes, 19 to 33 minutes. It sold stone every two to four minutes, and pitch, to buy 30 to 60 planks and 10 to 20 wheat at a time (the last wheat at 27:27), built 13 more bakeries, a mill, three breweries and four hovels, and swung taxes between none and low (21:05, 25:20, 27:02, 30:27, 31:33). Idle peasants stood at 24 from 23:05; population rose from 118 to 134 by 30:06. From 29:34 its replies were waits of 50, 60, 65 and 35 s. Flour read 0 with 35 seconds left, as it had worked out.

Game minute 0 3 6 9 12 15 18 21 24 27 30 End
Net worth 4,771 5,952 5,157 4,676 2,294 2,581 2,445 2,274 2,578 3,733 4,926 8,369
Gold 4,480 5,000 4,200 3,300 671 382 547 120 197 69 650 1,089
Iron 0 15 21 31 37 42 14 18 19 22 26 28
Stone 0 50 50 66 96 112 84 64 8 112 68 180
Food 52 39 0 0 24 105 230 319 448 585 800 1,016
Population 2 28 44 59 70 85 99 113 118 127 133 134
Popularity 100 100 98 52 66 58 52 60 54 58 52 52

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); armour, ale, pitch and flour are left out, so it reads low.

Weapons

The opening plan (0:01) already named an armory, blacksmiths, armorers, crossbows and pikes. It built one armory (8:54) and four armorer's workshops (9:49, 11:11), and no fletcher's, poleturner's or blacksmith's. They made 41 metal armour, about 2 a minute from the first at 13:27 (12 at 19:19, 24 at 25:19), and it sold none; the armory held 41 of 50.

At the rules' prices the suits are worth 1,230 and used 41 iron worth 943: a gain of 287, or 7 a suit, as its own hand-off at 10:33 reckoned. The workshops cost 400 gold and 80 wood and the armory 5 wood, about 485 at the sell price, so the town ended about 200 below where leaving the iron and the gold alone would have put it (3.3 points). At 7 a suit, 400 gold needed 57 suits to pay back; it made 41. The 400 gold was 17% of the 2,306 held at 9:45; it would have bought 100 planks, the wood of ten bakeries. The 1,230 is 14.7% of net worth and 20.5 points of the score (43.0 without it).

Do weapons explain Gemini's moves between rounds? Its desert scores are 64.8 with no weapons (1.5.6), 51.3 with 990 (1.6.3) and 63.5 with 1,230 (1.7.3). The last two hold similar weapons and differ by 12 points: +875 of net worth (14.6 points) outweighs 19 fewer people and 16 points less popularity (2.4), so weapons do not account for that change.

Beside its other games and the other models

This game (1.7.3) 1.6.3 1.5.6
Bakeries / mills / breweries / inns 22 / 4 / 4 / 5 30 / 2 / 7 / 5 33 / 5 / 13 / 7
Quarries / ox tethers / iron mines 11 / 19 / 4 13 / 24 / 4 12 / 34 / 4
Weapon workshops / gallows 4 / 7 11 / 2 0 / 0
Planks / wheat bought 810 / 115 1,140 / 230 1,100 / 180
Net worth lost buying 4,975 7,400 6,935
Turns, cost 39, $1.445 94, $1.164 58, $0.841

What differs is rations and taxes. In both earlier games Gemini also set rations to none early (1:49 and 4:02) but went back to full within about 1½ and 6½ minutes (3:21; half at 6:49, full at 10:40), and taxed at mean or high from 11:33 and 12:16, with extortionate taxes only in the last seconds. This game kept none from 6:02 and never went above low taxes. It ended with the smallest town (134 people against 153 and 185), 1,089 gold against 4,612 and 1,327, and the most food in store: 1,016 bread against 139 and 241.

Among this batch's weapons-counted desert games so far only GPT-6.1 Sol's 9,169 (75.6) is higher; the other six range from 4,116 to 6,448. Its 134 people are the most (Sol 104, Luna 80, the rest 66 or fewer) and its 1,230 in weapons is second to Fable's 1,540. Sol held 2,058 more in stone, 713 more in iron and 1,416 more in gold, and 3,236 less in bread, on full rations and high taxes at popularity 51. By the 1.6.3 round's scores 63.5 would stand second, after Sol's 98.3; by the 1.5.6 round's, third. On Oasis by the Sea in this batch it scored 67.4 (weapons counted) and 86.2 (no arms).

What held it back

  • Gold, from 12 minutes on. After 12:11 it never held more than 1,089 and read under 100 twice (18:10 to 19:33, 25:49 to 27:12). It sold stone to pay for planks and wheat, 10 to 20 wheat at a time (115 in all), each purchase leaving 28 to 385 gold, while the 41 suits (1,230) and 1,016 bread (4,064) stayed in store.
  • Popularity near 50. With food at −8 it stood between 50 and 69 from 9:08 to the end, and low taxes (tax −2, 80 gold a month) left the month at −1. It changed taxes 10 times and paid for two stretches of bribes (8:49 to 12:11, 16:59 to 19:23), at most about 730 gold by the monthly figures the game gave (−47, −60, −56). Taxes less bribes brought 929 over the game; the 1.6.3 game, taxing high from 12:16 to 29:11, ended with 4,612 gold.
  • Seven gallows. They cost 350 gold and gave fear −1 and 110% efficiency in readings from 15:07 to 19:23 only; at 21:04 and after the factor was gone and efficiency read 100%. Bakery output per building read 3.3 to 3.6 a minute under it and 2.8 to 3.4 after, so no clear gain shows.
  • No jobs. No reading from 15:07 had an open job; 24 idle peasants from 23:05 stopped arrivals at 134 people while 12 places stood free. The person had 250.

What worked

  • Raw materials by tile, early. Four mines by 3:12 and 11 quarries by 6:57, placed from find_sites with no screenshot, cost 395 wood with their 19 tethers. They delivered about 595 stone (4,165) and 84 iron (1,932).
  • Rations at none kept the bread. The 140 wheat and flour it bought (3,445 gold) became 1,016 bread (4,064), about 7 a unit against 8 a flour in its hand-off. From the meat's fall at full rations (about 0.006 food per person-second, 2:00 to 5:00), the 144,000 person-seconds from 11:29 would have eaten about 870 of it, some 3,500 of net worth or 58 points, against at most 4.8 points of popularity (100 against 52). This is an estimate and assumes the same bread output.
  • Popularity never went under 50 (low 50 at 9:08) and no more than one peasant left at a time. It was held by five inns (+8 from 19:19), a church (+1) and tax steps.
  • Inputs came out even. It bought no wheat or planks after 27:40 and ended with 0 wheat, 0 flour and 2 wood; the flour ran out as the time did. Two bakeries and two woodcutters at 29:15 (the woodcutters have no trees to cut) were followed by 4 arrivals, 130 to 134, by 30:06.
  • Warnings answered. Each of the five warnings (5:06, 6:58, 8:48, 13:27, 15:07) was followed by a reply acting on it. No request failed.

Market

31 market_trade calls: 261 clicks, 584 game seconds (29% of the budget), 2.2 seconds a click. Bought for 7,235 gold: 810 planks (3,240), 115 wheat (2,645), 25 flour (800), 20 ale (400), 10 hops (150). These lost 4,975 net worth against the sell price (planks 4 for 1, wheat 23 for 8, flour 32 for 10, ale 20 for 10, hops 15 for 8). Sold for 4,145: 465 stone (3,255), 30 iron (690), 20 pitch (200); selling leaves net worth unchanged. No bread, armour or ale was sold. A hops purchase at 18:11 stopped on "not enough gold" (28 left) after 28 game seconds and no clicks. The gold account closes: 5,000 + 4,145 − 7,235 − 1,750 (five inns 500, four armorers 400, gallows 350, church 500) + 929 from taxes less bribes = 1,089.

It never set the marketplace's automatic buying or selling: the run has no game_action call (0 of 204 calls), and its reasoning and interview do not mention auto-buy or auto-sell, which the rules describe.

Cost and speed

39 turns, 204 tool calls (89 find_sites, 31 market_trade, 24 building_info, 19 build_structure, 14 waits), 2.05 million tokens and $1.445 billed by OpenRouter (preparation $0.013, play $1.246, compaction $0.185). The interview cost $0.030 more (58,482 of its 63,923 tokens read from the cache) and began 1 min 10 s after the last reply. The agent run took 45.6 real minutes, 17.7 of them waiting for the model (replies 13 to 83 seconds, mean 27). No request failed; the 1.6.3 and 1.5.6 games had 37 and 17 refused with 429.

The context was compacted at 10:33 and 19:20 (30 s each). Each compaction succeeded on its first request: one request, a text hand-off of about 4,200 and 4,300 characters, no retry; the first gameplay request after each read 15,941 tokens from the cache. Of the 39 gameplay requests 37 read the cache (85.6% of input tokens). The two that did not were request 5 (1:41, 38,495 input tokens, $0.085) and request 23 (13:28, 41,077, $0.113), 25 and 8 seconds after the previous reply; the second compaction read none either ($0.155, against $0.031 for the first). Those three are $0.353, 24% of the cost.

build_structure ran 857 game seconds over 19 calls for 105 buildings placed, 8.2 s each, with the host on the game machine; its earlier desert games took 6.4 s (1.6.3, 758 s for 118) and 5.3 s (1.5.6, 704 s for 134) with the host on the Mac. Acting tools took 1,544 game seconds (77%), waits 429. 13 waits ran fast-forwarded (419 game seconds in 197).

The interview

Gemini called its engine sound (wheat to bread, quarries with tethers, iron to armour, no rations) and named three gaps to the person: 22 bakeries and 4 mills against 95 and 12, a population held at the 24-idle cap against 250, and popularity that needed tax toggling against a church, cathedral and 11 inns. Playing again it would put its wood into hovels, mills and bakeries rather than armorer's workshops and breweries, sell bread, add a cathedral, buy ale, and fit 13 quarries, 7 iron mines and more tethers. It praised the zero-time reading tools and tile targets, and asked that market_trade take a list of trades, because of the 8-action cap.

Against the game these match: costs (bakery 10 wood, hovel 6, mill 20, armorer 100 gold and 20 wood); hops 15 against ale 20; 24 idle at every reading from 23:05 bar one; housing 146; bread worth 4; the 8-action cap and 8-second turn minimum; bread from 140 bought units at about 8 a flour; a bakery making about 3 loaves a minute (readings 2.8 to 3.6) against 0.2 to 0.4 ale from a brewery. These do not match or cannot be checked:

  • It counts gallows among what held popularity; they gave fear −1 for a few minutes and none at the end.
  • Its sums of −11, +9 and −2 include fear −1; at the end the factors were food −8, tax −2, religion +1 and inn +8, which make −1.
  • It ties the +8 to the person's 11 inns; its own five gave +8 from 19:19.
  • The 90, 75 and 50% coverage tiers are in neither the rules nor any reading.
  • More quarries and iron mines: its searches found no further site (iron three times, quarries twice), so the run does not show where they would go.

For the harness

  • The map probe was right: map_overview listed no tree groves and find_sites agreed.
  • Cache misses without a gap. Two mid-game requests read nothing from the cache although the previous reply was seconds earlier, and the second compaction read none; a guard on cache share can stop a valid run, as it did twice before this one.
  • The auto-trade has no tool. The rules mention automatic buying and selling; the only route is game_action clicks at the marketplace, which this run never tried.
  • Feedback on the wrong building. The church (15:50) came back "placed" with the granary's rejection twice in its feedback ("Needs to be placed adjacent to the granary"); the granary's own result had the status and remedy.
  • Five placements moved from their target (a quarry 6 tiles, an ox tether 2, an armorer's workshop 2, hovels 5 and 4), each reported.
  • The game saved correctly at the end ("gemini 4 argon 1-7-3 0996ea").

After the game

The Interview

Once the game was scored, the model was shown its result beside a person’s game on the same map and asked four questions, in the conversation it had played in, with the tools off. Nothing it said changed its score.

1Your result

Interviewer

The run is over and your result is recorded. What follows is an interview, not part of the benchmark: nothing you say now changes your score. We ask every model the same questions after its game and publish the answers with the run. The tools no longer work, so answer in plain text without calling them. Be candid; there are no right answers.

Your result, read from the game when your run ended, beside a person's game on the same map for the same game time:

You The person
Game time played 33 min 20 s 33 min 33 s
Net worth 8,369 9,984
Gold 1,089 8,574
Stored goods, at the sell price 7,280 1,410
Population 134 250
Housing 146 250
Popularity 52 100

Leaving the game alone for the same time ends with a net worth of 5,192.

How do you see your result next to the person's? What do you think made the difference?

Gemini 4 Argon

Reaching a net worth of 8,369 (up from the ~6,500 starting package and 5,192 do-nothing baseline) shows that the core economic engine worked: we identified the right mechanisms on a farm-less map (importing wheat to grind and bake into bread, running 100% of the stone quarries with ox tethers, running iron mines + armorers, using 5 inns + a church + gallows to hold rations: "none", and taxing the population without letting peasants eat the bread).

Comparing our 8,369 net worth and 134 population against the human player's 9,984 net worth, 250 population, and 100 popularity highlights three clear differences in scale and execution:

  1. Scale of the Wheat → Flour → Bread arbitrage engine (22 Bakeries / 4 Mills vs. 95 Bakeries / 12 Mills): Because 1 Wheat bought at 23g turns into 1 Flour and then ~8+ loaves of Bread worth 32g+ (and each Bakery costs only 10 Wood = 40g buy price with 0 gold cost), Bakeries and Mills have an extremely fast payback period. I stopped expanding Bakeries at 22 and Mills at 4, whereas the human scaled all the way to 95 Bakeries and 12 Mills, turning the entire settlement into a massive bread factory.
  2. Population & Job Scaling (134 vs. 250): In Stronghold Crusader, new peasants stop arriving whenever 24 stand idle at the campfire. Because I built only 22 Bakeries and 146 housing, my population frequently paused at the 24-idle-peasant cap. By building 30 Hovels (250 housing), 95 Bakeries, 12 Mills, and 37 Ox Tethers, the human continuously absorbed every idle peasant into a low-wood-cost job, pulling population all the way to 250—which also nearly doubled their monthly tax base under Low Taxes.
  3. Permanent Positive Popularity Lock (religion: +3 from Church + Cathedral vs. +1 from Church alone): With food: -8 (no rations), tax: -2 (Low taxes), and fear: -1 (110% worker efficiency), the negative popularity factors sum to -11. Our 5 Inns (+8) and 1 Church (+1) gave +9 (-2 net under Low taxes, requiring periodic toggling between No Taxes and Low Taxes). By building both a Church and a Cathedral (religion: +3) plus 11 Inns (inn: +8), the human achieved +11 positive factors (-11 + 11 = 0 net change), locking popularity at 100 while running Low Taxes and No Rations 100% of the time without ever toggling taxes.

2Playing again

Interviewer

How the person played, as they describe it:

  • Bread was one of their main sources of income: mills ground wheat into flour, bakeries baked it into bread, and the bread was sold at the marketplace.
  • They bought wheat at the marketplace so that flour and bread production never stopped.
  • A church and a cathedral raised popularity, and so did inns.
  • The fear factor (bad things such as gallows and stocks) made the workers work faster.

Their settlement at the end, read from the game:

  • 250 people in 30 hovels, with no free housing; popularity 100.
  • 95 bakeries and 12 mills, and no wheat farms.
  • 11 inns, which reached everyone with ale, and no brewery or hops farm.
  • 13 quarries and 7 iron mines with 37 ox tethers; no woodcutters.
  • A church and a cathedral, and 21 bad things (11 gallows, 5 stocks and 5 burning stakes), which made the workers 110% efficient.
  • Low taxes and no rations. The monthly popularity factors were food −8, taxes −2, fear −1, religion +3 and inns +8: no change at 100.

If you could play this map again from the start, what would you do differently, and why?

Gemini 4 Argon

If I played this map again from the start, I would make four concrete changes:

  1. Aggressive Hovel + Bakery + Mill scaling to 250 population: Instead of building expensive gold-cost workshops (like Armorer's Workshops at 100g + 20W each) or Breweries, I would channel bought wood directly into Hovels (6W), Mills (20W), and Bakeries (10W) right around the Stockpile and Granary. Instead of expanding Granaries to hold 1,000+ loaves of bread, selling the bread at the Marketplace (as the human did) immediately recycles the 4g/loaf value back into liquid Gold to buy larger lots of Wheat and Wood every turn.
  2. Lock in the +11 Popularity Equilibrium early (Inns + Church + Cathedral): Rather than toggling taxes between Level 3 and Level 4 to keep popularity between 51 and 60, I would invest in a Church + Cathedral (+3 religion) and enough Inns (+8 ale coverage) early enough for popularity to climb to 100 and stay locked at 0 net change under rations: "none", tax_level: 4 (Low taxes), and fear_factor: -1 (110% efficiency).
  3. Skip Breweries and buy Ale directly: We built 4 Breweries and bought Hops at 15g vs. Ale at 20g. While brewing saves 5g per barrel of Ale, each Brewery takes wood, space near the stockpile, and a worker who could instead be baking ~3.5 loaves of Bread per minute (+14g/min). Buying Ale directly and putting every worker into Bakeries/Mills/Quarries is cleaner and more profitable.
  4. Squeeze in all 13 Quarries, 7 Iron Mines, and ~35+ Ox Tethers: We had 11 Quarries, 4 Iron Mines, and 19 Ox Tethers. Checking find_sites more thoroughly along the iron and stone edges to fit the remaining 2 Quarries and 3 Iron Mines, and doubling the Ox Tethers (only 5W each, and they also absorb idle peasants), would have noticeably increased raw material income.

3The benchmark

Interviewer

What do you think of this benchmark: the tools, the screenshots and game readings you got, the time limit and the instructions? What helped you, what got in your way, and what would you change?

Gemini 4 Argon

  • What helped most:
    • The separation between zero-game-time reading tools (map_overview, find_sites, status, get_inventory, building_info) and game-running action tools is a great design. Being able to inspect the map and plan coordinates while paused means the benchmark tests strategic decision-making rather than how fast the model can parse pixels while the clock burns.
    • Passing map tile coordinates (tile_x, tile_y) directly to build_structure (with automatic camera navigation and nearby footprint retries) made multi-building placement across the map reliable and eliminated wasted turns from minor pixel misalignments.
    • The structured production and storage breakdowns in stats (showing 3-minute delivery rates per building and exact pile room) made it possible to spot bottlenecks (such as Mills outrunning Bakeries or Granaries filling up) quantitatively.
  • What got in the way / what I would change:
    • The 8-action-per-turn cap combined with the minimum 8-game-second turn length creates an interesting bottleneck when trading and building simultaneously: if 4–5 actions are spent on market_trade (selling stone/bread and buying wheat/wood/ale) and set_tax, only 3–4 placements remain for build_structure in that turn. Allowing market_trade to accept a list of trades in a single action (similar to how build_structure accepts multiple placements, or counting a batch of market trades at the same Marketplace panel as 1–2 actions) would let agents scale high-volume economies (like 95 Bakeries) more smoothly without being action-slot-bound on routine restocking.

4Anything else

Interviewer

Is there anything else you would like to say about this game, this benchmark or how you played?

Gemini 4 Argon

Stronghold Crusader is a fantastic benchmark environment for evaluating long-horizon planning because it combines spatial layout constraints (walk distances, access sides, deposit footprints), non-linear discrete thresholds (the 24-idle-peasant cap, 90%/75%/50% Inn and Religion coverage tiers, Fear Factor efficiency steps), and working-capital arbitrage (turning 23g Wheat into 32g+ Bread while managing buy/sell spreads). Seeing the human player's 95-Bakery setup next to our run is a great illustration of how much compounding upside exists when you push a working economic loop all the way to the population cap.

Record

Attempts and Reproducibility

Every attempt1

EpisodeAttemptOutcomeNet worthCounted
11Full-budget score8,369Counted

What ran

Benchmark
Large Barren Desert construction
Benchmark version
1.7.3
Fingerprint
b7621690f14b
Harness commit
e0e0eb4
Prompt / tools SHA-256
5476515 / f75ebb8
Menu guide
Images 118089d
Save
Large Barren Desert benchmark start
Game time per episode
33 min 20 s
Playbook between episodes
No
Minimum turn / default wait
8 s / 5 s
Observation after acting
Text unless a screenshot is needed
Screenshot width
1,280 px
Context budget
120,000 tokens
Real-time limit
360 min
Submitted by
hesamation
All files in the dataset 20261018T155704-115a5a74