Gemini 4 Argon

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: Oasis by the Sea construction (no arms), version 1.7.3 (commit e0e0eb4), 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; sold at the marketplace they bring their gold.
• Run 1003de73, series 20261018T093657-80aeaf5f, 2026-10-18: Gemini's no-arms game of the 1.7.3 batch, a fresh series with no memory of any other game, played again after the first attempt (2026-10-17) failed on an OpenRouter key limit. The series records e0e0eb4, not b967c65 (the commit of its weapons-counted game, run ffe69c40): 1.7.3 relocked with a fix to the compaction retry, play and scoring unchanged.

Result

A full-budget score: the run used all 2,000 game seconds. Net worth grew by 16,170 over leaving the game alone, 88% of the person's 18,293. With 191 people and popularity 60 the score is 86.2 by the project website's formula (80% net worth, 10% each population and popularity, the net-worth share clipped to −1..2); this game counts on the website's leaderboard. It is 18.8 above its weapons-counted game of the same version, 18.2 above its 1.6.3 game (no weapon buildings) and 35.2 above its 1.5.1 game, and the highest of the batch's 16 finished games (next: GPT-6.1 Sol's weapons-counted 82.3).

The stored goods are 265 iron (6,095), 385 stone (2,695), 495 bread (1,980), 46 wood (46), 2 flour (20) and 2 bows, which count nothing. Gold is 38% of net worth, iron 35%, stone 15%, bread 11%. The town ended with 22 idle peasants, 51 empty places in its hovels and 495 food, all bread.

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 call keeps it running for 12 to 79 seconds (median 40).

0 to 6 minutes. At 0:00 the marketplace, two hovels, a woodcutter and an iron mine; 100 planks at 0:28; iron mines in twos from 1:02, ten by 5:03 (first iron at 2:26); 250 planks by 4:43. At 5:46 no iron site was left within 25 tiles of (419, 341) and the granary held 6 food for 52 people. Gold was 0 at 5:47 and stayed under 560 from 3:00 to 16:00.

6 to 16 minutes. Half rations at 6:10; 35 iron sold for planks in four lots (6:13 to 11:55); at 6:56 a wheat farm, a mill and three bakeries; six quarries and five ox tethers from 7:51 to 9:04; full rations at 10:57, low taxes at 11:01. At 11:55 its reasoning took 8 bread to a flour (corrected to 4 at 22:08), priced a bakery at about twice an iron mine and planned 20 bakeries; 15 stood after the 14:32 call. Double rations at 15:20 and mean taxes at 15:26: the +8 and −8 kept popularity at 62 to 69.

16 to 22 minutes. The 16:52 call placed three hovels, two gallows and the armory; workshops came from 18:00 (see Weapons) beside bakeries (23 after the 18:00 call, 35 after 21:27) and a fifth wheat farm at 20:29.

22 to 33 minutes. At 22:08 it dropped the workshops for wood-only jobs: 30 more ox tethers (39 in all), hovels, three mills, three woodcutters. Taxes went to extortionate at 22:46 and back to mean at 24:10 as popularity fell from 69 to 57 (its lowest). From 24:10 it placed a batch, waited 45 to 60 game seconds and read the result; the last build was at 31:30. Compactions came at 9:55, 16:49, 20:08 and 23:31. It sold 24 spears at 32:07 and 3 at 33:09.

Each column is the last reading at or before that minute. Before the End column the net worth rows count gold, wood, stone, iron and food (at 4) as the recording logs them; flour and wheat are left out, and neither game's rows count weapons (the weapons game's 1,272 is in its End value only).

Weapons

It did not leave weapons alone: one armory, seven poleturner's workshops and one fletcher's workshop, 27 spears sold, and weapons in its reasoning from the opening plan ("weapon workshops (selling crafted weapons at the Marketplace)").

When and why. In the second handoff (16:49) and the reply after it (16:52) it priced them: weapons count nothing in the armory, but a spear sells for 10 from 1 wood (+9), a pike for 18 from 2 (+16), a crossbow is "a 10x value multiplier"; it estimated 1.5 to 2 pikes a workshop a minute and a payback in about 4 minutes. (The no-arms rules leave prices out; these are the sell prices in the market_trade description.) The armory went up in the 16:52 call (its poleturner was rejected as too close to a signpost, its fletcher found no room); one poleturner followed at 18:00, a fletcher and two poleturners at 19:20, two at 20:29, two at 21:27. At 19:20 it also gave them as jobs: 20 idle peasants, and new ones stop at 24. At 22:08 it had one spear and wrote "building more 100-gold weapon workshops with 11 minutes left would LOSE net worth"; the fourth handoff (23:31) says to build no more (distant ones deliver about 0.1 to 0.3 weapons a minute) and to sell all weapons at the end. It never called set_workshop: the poleturners made spears (10 gold, 1 wood), the fletcher bows.

What they made. The armory held 1 weapon at 21:13, 2 at 22:52, 5 at 25:03, 13 at 27:57 and 23 at 31:32; the best readout was 0.4 spears a workshop a minute, against the 1.5 to 2 assumed. Six poleturners and the fletcher stood 55 to 67 tiles (straight line) from the armory at (443, 354); only the first poleturner, at (457, 364), was 17 away. It sold 24 spears for 240 at 32:07 and 3 for 30 at 33:09; the sale of 2 bows (30 gold) was refused at the session deadline.

Cost against return. The eight workshops cost 800 gold and 90 wood, the armory 5 wood, the spears and bows 31 wood: 800 gold and 126 wood for 270 gold, 656 of net worth behind (2.9 points of the 86.2). The two gallows cost 100 more and showed no fear factor or change in worker efficiency (100%). Held, the 800 gold would have counted 1:1 (3.5 points), and gold limited the town only until about 18 minutes (1,258 at 19:22, 6,554 unspent at the end). It priced a church (500, +1) and cathedral (1,000, +2) at 16:52 and judged them at 20:08 to "barely break even", and chose a fifth wheat farm over a hops farm, two breweries and four inns; ox tethers (5 wood) and woodcutters (3) gave jobs for wood alone, and became its main jobs from 22:08.

Beside its other games and the other models

Against its weapons game this one ended 4,216 higher: gold +4,000, bread +1,900, stone +189, against iron −483, wood −56, flour and wheat −62 and the weapons game's 1,272 in weapons. The gold: about 2,790 more income apart from sales (taxes, about 7,580 against 4,790, reckoned as final gold less the 1,000 start, plus purchases and gold-cost buildings, less sales), 1,600 less spent on buildings (900 against 2,500: 8 workshops and 2 gallows against 22 workshops and 3 inns), 210 less on purchases and 600 less from sales (1,075 against 1,675). The tax gap is the settings: mean taxes with double rations from 15:26 against low or average taxes (level 4 or 5, apart from mean at 3:07 to 4:35, none at 11:35 to 12:43 and high from 32:35) with full rations. The bread gap is food: the weapons game fell from 170 food at 23:49 to 20 with 14 bakeries; this one rose from 253 (23:32) to 495 on 38. With its weapons counted at nothing the weapons game would stand at 11,902 (61.8), still 24.4 points below this game. Against its 1.6.3 game (68.0) the gain is 3,787: stone +1,638, gold +2,124, iron +115, wood −80, flour −10.

Beside the other models. In the eight finished pairs of the batch the no-arms game scored higher in six (by 3.0 to 18.8) and lower in two (Sonnet −5.1, Sol −19.9); the weapons held in the weapons-counted games were worth 45 to 1,272 (0.2 to 5.6 points), the largest being this model's. For this model weapons do not explain the move: 68.0 (1.6.3, none built), 67.4 (1.7.3, 22 workshops, 1,272 held), 86.2 (1.7.3, 8 workshops, 270 sold). Each is one game, and the towns differ in taxes, food and housing as well.

What held it back

• Popularity 60. Mean taxes against double rations held it at 57 to 69 from 15:26, 40 under the person (up to 4.0 points). No church, ale or working fear building.
• Ten iron mines. The strip was full at 5:46; its weapons game fitted 12 on it and its 1.6.3 game 13.
• Ox tethers late. Six quarries stood by 9:04 with 5 tethers, and its own reading at 14:32 was that tethers limited them. It had 9 at 16:51 (13 stone a minute) and 16 at 23:32 (13); delivery rose to 21 at 27:55 and 37 at 29:44 with 30.
• Wood for placements. Eight placements failed for want of 1 to 20 wood (47 in all), seven of them with 1,463 to 3,748 gold in hand; it had stopped buying planks after 21:12 (the failure text names the fix).
• Five wheat farms. At 22:08 it counted the open oasis grass as full (5 wheat farms, 3 orchards, a dairy farm) and wrote that the chain was at its limit; the person had 10 wheat farms.

What worked

• Bread. 38 bakeries and 7 mills on 5 farms made 495 bread (1,980) while double rations ran for 18 minutes; food rose from 131 at 15:27.
• Taxes against rations. Mean taxes with double rations kept popularity above 50 and brought 229 gold a month at the end; population reached 191 of 242 places.
• Ten iron mines by 5:03, 300 iron dug (265 stored, 35 sold).
• Placements. 165 of 180 requested went through and every one reported placed is in the final count; the 15 that did not were 8 for wood, 4 for no room and 3 rejected as too close to a signpost.

Market

550 planks (4 each, worth 1) were bought for 2,200 gold, all by 21:12: 1,650 of net worth lost against the sell price, 7.2 points. 35 iron sold for 805 and 27 spears for 270. Nothing else was bought: the 22:08 reading of 1 wheat (23) to 1 flour to 4 bread (16) made buying wheat or flour a loss. The 16 market calls took 252 game seconds.

Cost and speed

80 turns, 321 tool calls (205 find_sites, 31 build_structure, 16 market_trade, 7 waits), 4.19 million tokens and $2.159 billed by OpenRouter, all to Google AI Studio (preparation $0.044, play $1.963, compaction $0.152). The run took 62.9 real minutes: 29.9 waiting for the model (80 replies, 22 s on average, longest 67 s) and 2.1 on four compactions; fast-forwarding six waits saved 98 s.

All four compactions succeeded on the first request: each sent the whole conversation with the handoff instruction (92 to 95% of its input read from the cache), stopped normally and wrote 4.3 to 6.7 KB; none used the text-only retry. All 80 gameplay requests read the prompt cache (91.5% of input tokens); the first after each compaction re-read a 15,895-token prefix.

Tool calls took 1,963 of the 2,000 game seconds: build_structure 1,253 (31 calls, 165 buildings placed, 7.6 s a building), market_trade 252, waits 333, the rest 125; the reading tools took none. The 15 build calls that moved the camera (24 moves) took 49 s against 32 s for the other 16 (9.7 against 5.8 s a building). Its weapons game came to 6.9 s a building, its 1.6.3 game 8.7 and its 1.5.1 game 6.5; the last two were hosted on the Mac, this game and its weapons game on the game machine. It placed 5.3 buildings a call.

The interview

It ran 90 s after the last reply: 4 answers, $0.029, 62,554 of 67,905 tokens read from the cache. Gemini put its result at about 89% of the person's and named three differences: fear buildings and religion (worker efficiency 140% against its 100%), buying wheat to keep the bakeries running, and selling stored goods to fund a church, a cathedral and about 50 fear buildings. Playing again it would do those, and fit 10 wheat farms, 10 quarries and 13 iron mines. It praised the paused reply, tile targets and the production readout, and asked for build calls to be ordered by camera position or the camera cost stated. It did not mention weapons.

Against the game these match: the 89%, 191 against 202, the 2,100 gap, popularity 60 against 100; 5 wheat farms, 6 quarries, 10 iron mines, 3 orchards, 38 bakeries; the 10 × 10 orchard footprint; stone rising from 13 to 37 a minute; the wheat arithmetic (23 gold in, 16 out), reached at 22:08 and never traded on; the camera cost. Its 264 iron and 499 bread are the 33:11 reading (265 and 495 at the end). Three statements do not match: "33 ox tethers" is the person's count, it has 39; the gallows were not "early" (placed in the 16:52 call, half way) and nothing in the logs shows it seeing that "more than 2" were needed, since no reading showed a fear factor; and selling stored goods to fund investments fits the first half, when gold was under 560 until 16 minutes, not the second, when gold sat idle from about 19 (it sold 35 iron in all). Two cannot be checked: that fear raises the yield per bought wheat (efficiency stayed 100%) and that its own buildings blocked iron-strip footprints (consistent with 10 mines against 12 and 13 elsewhere).

For the harness

• The granary line. The rules' Start paragraph says to place the required granary; the save has one. No granary call was made (the first reasoning saw granaries: 1); two more went up by expand_storage at 15:04.
• Compaction. All four took the single-request path, so the e0e0eb4 fix to the text-only retry was not exercised. An empty reply after the fourth (0 output tokens, 23:34) made the host let 8 game seconds pass.
• The deadline refused the last sale (2 bows) with "Session deadline". The build_structure text says the host moves the camera but not what that costs.
• The save needed two attempts (game_save.attempts in episode.json); the second succeeded.
• The reader's watchdog forced one collection after a 20 s stall of the game's garbage collector at about 31:30; every build_structure result reported feedbackCoverage: reader_active.