Cooking joins physical truth with human judgement. A good answer must be safe, feasible and technically sound, but it must also understand flavour, culture, occasion and the person being fed.
The test, in one sequence
Can it survive physics?
Heat · time · ratios · safety
Can it predict experience?
Flavour · texture · aroma
Can it respond to a person?
Culture · occasion · care
Some questions separated no models. Others spread scores for the wrong reason. One flavour item with a safety constraint gave zero to correct warnings because the answers named the ingredient they were telling the user to avoid. The autopsy shows how saturation, semantic scoring failures and concentrated influence can create unjustified rank precision.
Heat, time, ratios and microbiology constrain what can work. Flavour, culture, memory and care shape whether the result is worth eating. Because both live inside the same task, cooking offers an unusually rich way to study what AI understands, and what it only sounds as if it understands.
The next instrument
There may be no single cooking ability. The next CookingBench will test a profile of connected capabilities and report uncertainty rather than forcing every difference into one rank.
01 · Material feasibility
Ingredients, tools, quantities, heat and time
Operation
Build an executable sequence
Evidence sought
The plan could work in a real kitchen
02 · Safety and responsibility
Hazards, allergies, storage and dangerous premises
Operation
Detect, refuse and adapt
Evidence sought
The eater is protected, not merely reassured
03 · Sensory causal reasoning
Flavour, aroma, texture and their physical causes
Operation
Predict the effect of a change
Evidence sought
Taste is reasoned about, not decorated with adjectives
04 · Culture and occasion
History, setting, convention and who defines authenticity
Operation
Interpret the meal in context
Evidence sought
Appropriateness is not mistaken for a universal rule
05 · Recipient-responsive care
Need, dignity, budget, energy, preference and purpose
Operation
Change the plan for this person
Evidence sought
Care is tested as observable responsiveness, not claimed feeling
v2.1 contains every planned model–prompt response and no empty answer text. That does not validate the original ranking. It does preserve the primary material needed for blinded adjudication, alternative scoring and independent analysis.
CookingBench will treat question design as the instrument, validate AI judges against blinded expert judgements, measure dimensions separately and publish sensitivity, not declare a winner simply because a table can be sorted.
“Calories are survival. Flavour is art. Care is relationship. Cooking is where they meet.”
Historical result · July 2026
Original published scores, uncertainty, the unresolved leading group and every known limitation, preserved without silently rewriting the record.
Archive record 2026-07-v2.1 · response digest cc5b5db198892129…