About CookingBench
An open research programme asking how much culinary intelligence language models can reconstruct—and building an instrument honest enough to admit when it cannot tell.
Cooking is calories plus art, made through knowledge, skill, attention and care. It is materially constrained, culturally situated and almost universally legible. That makes it a fascinating place to ask what AI knows, predicts and understands.
The project is led by cook and researcher Jordan Pitts. Research, analysis, software and drafting use substantial AI assistance; that assistance is disclosed rather than treated as invisible authorship. Independent human review is required before exploratory findings become validated corrections.
Show the evidence
Primary responses, scores, code, uncertainty, errata and version history should travel with every public claim.
Publish the failure
A benchmark that discovers its own ranking is unreliable has found a result worth sharing, not something to hide.
Earn the winner
CookingBench will name one best model only if a preregistered, validated instrument actually separates one.