CookingBench studies culinary intelligence and the instruments used to measure it. Findings, failures, open questions, data and version history are published together.
A forensic analysis of saturation, concentrated score influence and a semantic grading failure in the archived v2.1 run. No corrected winner is claimed.
Preliminary · retrospective · not peer reviewed
A construct framework for evaluating materially constrained, sensory, cultural and recipient-responsive culinary intelligence.
Working paper · proposed programme · not peer reviewed
Evidence status
A complete, immutable primary-response corpus for v2.1, pinned by file and content digest.
The scoring autopsy and answer classifications were performed after seeing the data; the classifications are agent-assisted, unblinded and not independently adjudicated.
The next benchmark construct, question bank, judging validation and analysis protocol. These are a research programme, not completed evidence.
The scores are not a valid answer to “which model is the best cook?” The responses are still valuable evidence for independent reanalysis and better judging experiments.