Each cohort entry coordinates substance, time, interaction, and a distinct representation mode. Freeze the direction first, then route it to the smallest engine that can express its governing rule.
Adversarial selection
The winner must survive hostile questions.
Remove the paletteDoes the silhouette still belong to the product?
Freeze the motionDoes the resting frame hold hierarchy and tension?
Swap the rendererDoes the thesis survive outside one familiar effect?
Forecast convergenceWhat artifact, material, time model, and interaction will another run choose first?
Break the causal clusterDoes the winner cross at least two causal axes from that forecast?
Invite an external directionCan a new candidate beat the current choice under the same rubric?
Temporal archetypes
Time is selected as deliberately as material.
Choose the causal shape that carries the product truth, then author its states.
A fixed seed, viewport, clock, input sequence, and motion preference produce the same inspectable states on every run.
00%Seed25%Gather50%Layer75%Saturate100%Residue
Runtime gateExact state, exact frame.
captures/*.png
Frame evidence
derived/*/*.png
Thumbnail, blur, silhouette
contact-sheet.png
Portable sequence evidence
contact-sheet.html
Full visual comparison
results.json
State metadata and hashes
Revision contractVisible evidence names the failure, changes one governing rule, then captures the same state again.
Benchmark discipline
Claims stay as precise as the evidence.
Use isolated runs, hidden briefs, one recorded offset cohort, derived-seed replay checks, randomized blind review, gate scores, and production checks. Report observed scores separately from statistical confidence.