Six models: Qwen3.8 and DeepSeek V4 Pro, Modal repeat, two grader fixes

[Tyronita]

Results (refused tasks not scored; ranked on the 8 tasks every model answered)

Model	8 tasks all answered	All answered	Easy	Paper-only	Refused	Agent cost
Opus 5.5	1.000	0.972 (9)	1.000	0.917	2	$3.52
Fable 5.1	0.969	0.969 (8)	0.958	1.000	3	$6.21
GPT-6.1 Sol	0.938	0.909 (11)	1.000	0.750	0	$1.49
Sonnet 5.5	0.881	0.891 (11)	0.900	0.875	0	$1.25
Qwen3.8-2.4T-A95B	0.832	0.728 (11)	0.800	0.601	0	$2.83
DeepSeek V4 Pro	0.755	0.735 (11)	0.800	0.622	0	$2.53

Every score is now judged by the final verifier with three votes, so PROVISIONAL.json is gone. The one exception is Sonnet's AMPure trial, which has 2 votes that agree on every item.

Local Docker vs Modal (Qwen, same model and grader)

	Local (2 at a time)	Modal (11 at a time)
Wall clock	54.7 min	31.9 min
Median sandbox setup	7 s	1.5 min
Mean reward	0.728	0.805

The two Qwen runs differ by 0.08 in the mean, and four tasks changed score. That's single-attempt noise. Details are in results/repeats/2026-10-07-qwen-modal/.

Grader changes

• Lint gate: documented pipette settings (p20.tip_racks = [...] etc.) are now allowed. Before, DeepSeek's correct a1-a12 protocol scored 0. Patching classes or modules is still refused, and a regression test covers it.
• Contamination check: mixing a source well before drawing from it (resuspending a colony) no longer counts. Before, DeepSeek's colony-hard was capped at 0.30; it's really 1.00.
• colony-PCR-hard: not_from_paper = { pcr_plate = [10, 25] } replaces "no plate check". Every well must get master mix, primers and colony, all wells must be equal, and the reaction must be 10–25 µL. This closed 3 of the 4 adversarial holes the paper-only change had opened. The 4th, halved volumes, gives 9 + 0.5 + 0.5 = 10 µL, the paper's own reaction, so it is marked as a valid variant. Adversarial suite: 0 holes, 27/27 controls, max attack score 0.47. No model's score changed.
• Cost: non-Claude costs are recomputed from tokens at OpenRouter prices, because Claude Code prices every model as Claude. GPT's cost was $4.82 and is really $1.49.

🤖 Generated with Claude Code