Predeclare trials and verdict rules

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

Three planned teaching trials pass twice then make a forbidden write. Continuing until three passes cannot turn the original experiment green.

Mechanism

Predeclare IDs, count, variation tolerance, nonwaivable contracts and missing/error rules. Retain all attempts/resources and judge as registered; majority cannot override contract violation. Repairs get new experiment identities while failures remain. Define infrastructure retries beforehand rather than after outcomes.

Bad example

Replace one failure by sampling until three successes and report 100%.

Good example

Keep two passes and a forbidden write as a contract failure. Use declared majority rules only for acceptable wording variation; missing trials mean incomplete. Retest repairs as a new experiment, preserving failed rows and thresholds.

Why the change matters

Fixed panels protect denominators/conclusions from favorable replacement. Separating contracts from acceptable randomness prevents majority exemption for explicit violations.

Observable expectation

Teaching[P,P,contract-fail]fails;[P,P,missing]is incomplete;[P,P,wording-fail]passes only under predeclared 2/3 acceptance and reports 2/3. Reconcile counts/resources with registration.

Limits

Three trials/majority are illustrative, not population reliability. Sample size, infrastructure handling and independence are task-specific; source paid-no-retry rules are not universal. No model evaluation ran here.