GLM 5.3 Flash

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: GLM 5.3 Flash (z-ai/glm-5.3-flash) on OpenRouter, provider z-ai/fp8 only, reasoning medium, output limit 32,768 tokens
• Benchmark: The Sinkhole construction (no arms), version 1.7.3 (commit e0e0eb4), 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. Weapons and armour in store count nothing toward net worth; sold at a marketplace, they bring their gold.
• Run d50859a4, series 20261019T094818-55ae43b7, 2026-10-19, host on the game machine: GLM's no-arms game on this map, a fresh series with no memory of its others. Its weapons-counted pair is its listed 1.7.1 game (run a4e2d606, series 20261015T215019-394d6bce, commit 674e7ba, host on a Mac). Its other games on the map are a second 1.7.1 game (run a044594a, commit 07f2b7a, not listed) and run 2452ade8 (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 6,077 over leaving the game alone, 15.4% of the person's 39,436, against 8.8% in the 1.7.1 game. By the website's formula (80% net worth share, 10% each population and popularity, the net-worth share clipped to −1..2) the score is 23.6: 12.33 points from net worth, 2.77 from 143 people and 8.50 from popularity 85; this game counts on the website's leaderboard. It stands 8.5 points above the 1.7.1 game; scored by this game's rule that game ends at 4,660 and 13.0, and the gap is 3,612 and 10.6 points. Among the 1.7.1 round's ten games (weapons counted) 23.6 would stand eighth; among this batch's five no-arms Sinkhole games so far it is last, 0.2 behind Claude Haiku 5.5's 23.8.

Of the 8,272, 3,476 is gold (42%), 2,208 is 96 iron from 11 mines (27%), 1,736 is 90 flour and 209 bread (21%) and 791 is 113 stone; wood and a cheese make up 61. The armory's one bow counts nothing. The town ended with 24 idle peasants, 27 empty places in its hovels and taxes off.

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).

Planks and the first mines, 0 to 6 minutes. The first call (0:00, 81 s long) placed the marketplace, two hovels, two woodcutters and two orchards. It bought 80 planks (1:27, 3:16, 3:49), added woodcutters, an iron mine and hovels (1:45) and, at 4:14, a quarry with two tethers on the stone 84 tiles away, a second mine and an orchard (a dairy farm and a wheat farm were not placed). Average taxes came at 3:36. It had 24 idle peasants from 3:47 and 18 buildings by 5 minutes (the 1.7.1 game 12); popularity fell from 100 to 82 by 6:10.

Food, 6 to 13 minutes. Two wheat farms, a dairy farm, a mill, a bakery and two hovels at 6:11. The starting meat ran out: a warning at 7:37 (7 in the granary) and, in the recording, none at 8:00 until bought apples arrived (8:14 to 8:23). Popularity fell to 47 at 10:33 and four people left. It set low taxes (10:10), none (10:41) and a chapel (10:48, 250 gold) whose religion factor read 0 in all 44 later readings. Two quarries at 11:18, a second bakery at 12:24.

Bows and bakeries, 13 to 22 minutes. The armory and first fletcher's workshop at 13:16 (see Weapons), more workshops at 13:55, 15:40, 19:06 and 22:12, two quarries (15:40), bakeries at 17:55 and 19:37. Food was low at 14:37, 17:54 and 19:45 (6); it bought 20 apples (14:38) and 25 (20:50). Gold fell to 412, it sold 100 stone (21:33) and set average taxes (22:00). Population went from 52 to 85 and popularity from 55 to 91.

Jobs for the idle, 22 to 34 minutes. With 21 to 24 idle and no open job, nearly every call added jobs: 11 more bakeries (23:14 to 33:37), workshops, four mines on an ore field 130 tiles south (23:14 to 28:44) and four on the near field (31:52, 33:37). Searches at 31:12 and 31:52 found no quarry site within 25 tiles of either deposit. Average then low taxes (22:00 to 29:27) took popularity from 91 to 55; no taxes from 29:27 brought it back. The only compaction came at 34:23.

Mills, workshops, waiting, 34 to 45 minutes. Four more mills (34:26 to 36:54), workshops and hovels (housing 138 to 170), low taxes 37:45 to 41:04 (popularity 84 to 73) and then none. Two waits for wood (38:47, 39:49) ended at 24 and 38 against the 45 asked. The last two workshops went down at 41:29; it waited 150 s, sold 10 bows at 44:27, and a last sale at 44:55 was refused at the deadline.

Each column is the last recording frame at or before that minute; buildings are counted from calls begun before it. Before End the net worth row counts gold, wood, stone, iron and food (at 4) as the recording logs them, and leaves out flour and wheat (about 500 to 900 from 20 minutes), so it reads low.

Weapons

It built them anyway. Its opening reply listed workshops selling weapons for gold, and its rules say that weapons count nothing in the armory and bring their gold when sold. At 13:14 its reasoning set a bow's 15 gold against wood at 1 and called fletchers very profitable (a bow takes two wood, by prompt/game-mechanics.md); the armory (5 wood) and a first workshop (20 wood, 100 gold) went down at 13:16 and 17 more followed to 41:29. From 15:39 its reasoning also named jobs for the idle peasants. It never called set_workshop, so all 18 made bows (no readout shows a crossbow), and it built no poleturner, blacksmith or armorer. At 34:25 it called their 0.2 bows a minute weak, and built six more.

They made 56 bows. It sold 55 in seven calls (24:01 to 44:27, lots of 5 or 10) for 825 gold by the recording's gold steps; one bow stayed, worth nothing. The replies add to 54 and 810: at 38:01 the reply counted 4 bows and 60 gold, and the gold rose by 75. The wing cost 1,800 gold, 365 wood in buildings, 112 in bows and 18 jobs, filled from the idle pool (idle read 15 to 24 and no job stood open in any reading after 14:00). A workshop made 0.1 to 0.4 bows a minute (1.5 to 6 gold); in the last readout a quarry made 19 gold a minute and an iron mine 9. With wood at its sell value of 1 the wing returned 825 against 2,277, −1,452 of net worth (2.9 points); with the planks at the 4 gold it paid, −2,883.

The 1.7.1 game began its wing at 13:15 as well: 15 fletcher's workshops (all set to crossbows with set_workshop), two blacksmiths and an armory, for 1,900 gold, 520 wood, 21 iron and 17 jobs. It made 87 weapons (23 bows, 43 crossbows, 21 swords), sold 49 for 1,245 and held 38 worth 1,020, 18% of its 5,680. By its own rule the wing returned 2,265 against 2,903 at sell values (wood 1, iron 23; −638); by this game's, 1,245 (−1,658). The wings were alike in size and start; this one made bows only, used no iron and sold nearly all it made. Its 96 iron stayed iron (2,208 of net worth), and its gold, wood and workers otherwise went to 1,760 of purchases and a bread chain (320 wood for 4 wheat farms, 5 mills and 16 bakeries).

Beside its other games and the other models

The pair differs in more than weapons: the version (1.7.2 retries a blocked build_structure target only within the footprint's side plus a tile, where 1.7.1 searched 12; 1.7.3 added this benchmark), the host (a Mac there, the game machine here) and the town the model built. The score is 8.5 points higher: net worth +5.26, popularity +2.00, population +1.28. Counting weapons in both, this game leads by 2,607 (its last bow adds 15); in neither, by 3,612. What the two files show:

• Jobs. Both towns held 21 to 24 idle peasants with no open job almost throughout, so population followed jobs: 119 employed (143 less 24) against 55. This game placed 108 buildings against 59 and built to 41:29; the 1.7.1 game's last was at 37:21, and its 12 waits took 1,150 game seconds (271 here).
• Iron: 11 mines against 3, +1,127. It also mined about twice the stone (408 against 216) and sold 315 of it.
• Bread and popularity. Four wheat farms, five mills and 16 bakeries, full rations all game and food never under 22 after 22:00, against apples, dairy, an inn and half rations switched seven times; popularity 85 against 65.
• Taxes: about 2,250 gold against 124.
• Early wood: the marketplace stood from 0:00 and 80 planks were bought by 4:08; the 1.7.1 game built its marketplace at 11:00 and bought nothing all game.
• Against: 1,110 of net worth lost buying planks and apples, 250 on a chapel, and 88 fewer stone.

The 1.7.2 reach shows in the placements: 23 of 131 tries did not go through (5 of 64 in the 1.7.1 game). Ten offered a nearestFit 5 to 12 tiles away and nine were placed at the named tile in a later call, 39 to 102 game seconds on. Twelve placements moved, 4 tiles each but an orchard (10) and the chapel (7); the 1.7.1 game moved six, up to 10 tiles. The town's size confounds this. On the same rules Sonnet 5.5 (35.3), Opus 5.5 (39.3) and Fable 5.1 (41.8) scored far above 23.6; in the 1.7.1 round GLM's 15.1 was last.

What held it back

• The bow wing, about −1,450 of net worth and 18 of its 108 buildings.
• Taxes against popularity. Average taxes from 3:36 brought about 500 gold in 7 minutes while popularity fell from 100 to 47 with food short; 22:00 to 29:27 about 1,080 while it fell from 91 to 55; 37:45 to 41:04 about 630 for 84 to 73. Taxes in all were worth 4.6 points; the chapel's 250 gold left religion at 0.
• Buying. 230 planks (920 gold) and 105 apples (840) lost 1,110 of net worth, 13% of the final figure. The apples bridged real shortages (food 0 at 8:00, 6 at 19:45).
• Space and wood. Ten tries at farms, quarries and mines found no ground, were off screen or off the ore; no quarry site was left at 31:12 and 31:52. An iron mine (31:52) and a fletcher (35:56) were refused for wood, and two waits (121 s) for planks ended short.
• The idle cap. Population stood at 140 to 143 from 37:44 to the end; 24 idle in 17 of 54 readings.

What worked

• A marketplace and planks at the start: 18 buildings by 5:00 against 12.
• Jobs kept coming: population 143, 119 employed, from 37 build calls of 1 to 7 buildings.
• Iron: 11 mines, 96 iron, 27% of net worth; 0.4 iron a minute each in the last readout.
• Food after 22:00: never under 22, 209 bread at the end, popularity back to 85 after a low of 55 (29:26) once taxes were off.
• Stone sold in eight batches, the first when gold ran low (412 at 21:06); net worth unchanged.
• Placements held: all 108 reported placed are in the final count; the 23 that did not go through were not built (a fletcher marked unverified at 34:26 is not among the 18).

Market

Bought: 230 planks in six purchases (1:27 to 32:32) for 920 gold and 105 apples in four (8:05 to 20:50) for 840. Sold: 315 stone for 2,205 and 55 bows for 825, 3,030 in all, where the replies add to 2,980 (310 stone, 54 bows). The recording gives the figures used here. Seven of the eight stone replies, six of the seven bow replies and every purchase match its gold steps (two purchases overlapped tax payments). Two differ. At 27:47 the reply gave 60 stone and 420 gold; the gold rose 515, about 60 of it a tax payment in the same frame, so 455 for 65 stone (five clicks of 10, three of 5). At 38:01, 4 bows and 60 against 5 and 75. The stone sale was made with 80 in stock, below the 200 at which the known fault is described; the gold per click in its sales (140, 70 and 35, so 20, 10 and 5 stone as the stock fell from 116 at 21:33) shows the click size following the stock below 200 too. stock and gold in both replies were right and scores are unaffected. Purchases lost 1,110 against the sell price (planks 690, apples 420); selling stone leaves net worth unchanged, and a bow adds its gold. The 26 calls took 408 game seconds (purchases 198, stone sales 129, bows 78, the refused call 3). None of its 96 tool calls was a game_action, so it set no automatic trade by clicks, and every fall of gold in the recording is accounted for: 18 workshops, the chapel and the purchases.

Cost and speed

67 turns, 96 tool calls (37 build_structure, 26 market_trade, 19 find_sites, 8 set_tax, 3 waits, 3 others), 3.30 million tokens and $0.122 billed by OpenRouter (preparation $0.003, play $0.116, compaction $0.003). The interview cost $0.002 more, with 44,288 of its 46,266 tokens read from the cache. The agent run took 47.6 real minutes, 12.1 of them waiting for the model. Every request went to Z.AI and none failed (the 1.7.1 game had 27 retried 429s); 66 of 67 read the prompt cache (96.7% of input), the miss being the first after compaction. The compaction (34:23, context estimate 76,898 to 43,871) was one request that returned the handoff, 22 s, with no retry. Its figures for gold, stock, population, workshops, bakeries and mines matched the game; its counts of tethers (6 against 8) and southern mines (5 against 4) did not.

Of the 2,700 game seconds the 37 build calls took 1,616, the 26 market calls 408, the 3 waits 271 (one of 150 s played fast in 56 real seconds), the 8 tax changes 54 and the host's 5-second top-up after 51 replies that acted without a screenshot 255. The build calls placed 108 buildings: 15.0 game seconds each, 2.9 a call. The 1.7.1 game's 24 calls took 823 s for 59 (13.9 s each, 2.5 a call). Its host was a Mac and this one's the game machine, and the version and buildings differ, so the 1.1 s says nothing by itself.

The interview

GLM called the gap stark and put it down to the person selling bread rather than storing it (161 bakeries) and to inns, ale and fear-factor buildings that cut rations and raised worker efficiency, where it stored goods (4,796 of its 8,272) and ran full rations. Playing again it would build the bread-selling economy at once, inns, a church and fear-factor buildings, jobs and hovels in larger batches and more iron mines and quarries, and worry less about popularity. It praised build_structure tile targets, find_sites, place_near and the reader's stats, found the popularity and immigration rules opaque and complained that per-building output came only in 3-minute windows.

Against the game these match: 8,272 against 41,631; 143 against 506 people; 11 iron mines and 16 bakeries; full rations and no inn all game; stored goods 58% of net worth; no bread sold (209 left); worker efficiency 100%; bows a poor substitute (825 gold for 1,800). These do not: it says it did not understand why growth stalled, but its reasoning named the 24-idle limit at 3:47 and 7:37 and at 34:26 it wrote that it would add job buildings; it praised place_near, which it never called; and "net worth really means gold" is not the rule, since stored goods count at the sell price and each of its eight stone sales left net worth unchanged (weapons are the exception). "Stuck at 24 for most of the run" is 17 readings of 54 (41 at 21 or more). The person's inn, ale and fear figures come from the interview prompt and cannot be checked here.

For the harness

• market_trade undercounted two sales (stone at 27:47, a bow at 38:01), the first with 80 in stock, below the 200 the known fault is described at.
• 1.7.2 placements: ten refusals offered a nearestFit, nine used in a later call (39 to 102 s on). A fletcher reported unverified (34:26) was not built; it was not checked. Two tiles were not_on_screen (26:43, 28:44); the quarry was placed later (27:47), the mine not.
• A text-only reply at 44:44 with 13 s left (the host ends a run after four in a row); the last call, at 44:55, was refused at the deadline.
• No provider errors or stalled replies. The game saved at the end ("glm 5-3 flash 1-7-3 d50859"; episode.json records two attempts).