Lab Note · 2026-08-28 · Blind-scored gallery

Shore Temple at Mahabalipuram — the six scenes.

Six voxel scenes, two models, three runs each. Identical prompt, reference evidence pack, and sampling budget. All six failed single-shot; every scene here is the repaired version, with the iteration count shown. Scoring is done and the key is unsealed below. Click a scene to orbit it; click the world to ring the bell.

Scene A1 glm-5.3-flash · run_1

repair iterations to working: 1 · tokens: 62,788 · wall: 44 min

blind score: 8/10

open full screen →

Scene A2 glm-5.3-flash · run_2

repair iterations to working: 1 · tokens: 63,086 · wall: 46 min

blind score: 7/10

open full screen →

Scene A3 glm-5.3-flash · run_3

repair iterations to working: 1 · tokens: 55,843 · wall: 41 min

blind score: 8/10

open full screen →

Scene B1 qwen3.8-flash-next · run_1

repair iterations to working: 3 · tokens: 35,820 · wall: 15 min

blind score: 9/10

open full screen →

Scene B2 qwen3.8-flash-next · run_2

repair iterations to working: 2 · tokens: 29,073 · wall: 12 min

blind score: 5/10

open full screen →

Scene B3 qwen3.8-flash-next · run_3

repair iterations to working: 2 · tokens: 25,664 · wall: 11 min

blind score: 6/10 · scored after 1 extra visual-symptom repair iteration the other five artifacts did not get

open full screen →

The blind key, unsealed

Artifacts were scored blind (A/B labels, key sealed until scoring). Scoring is complete; the key is part of the published record. Raw file: KEY.txt.

A = glm-5.3-flash — A1: run_1, A2: run_2, A3: run_3 · blind scores 8, 7, 8 (mean 7.7)
B = qwen3.8-flash-next — B1: run_1, B2: run_2, B3: run_3 · blind scores 9, 5, 6 (mean 6.7)

Teardown · Creative eval report · Battery transcripts · Raw TEARDOWN.md