Prompt 32 — Verified physics training-pair and replay-record generator

Inputs

FROZEN_EPISODE_BUNDLE: {{FROZEN_EPISODE_BUNDLE_JSON}}
PROMOTION_AND_VALIDATION_ARTIFACTS: {{PROMOTION_AND_VALIDATION_ARTIFACTS_JSON}}
REJECTION_AND_FALSIFICATION_ARTIFACTS: {{REJECTION_AND_FALSIFICATION_ARTIFACTS_JSON}}
SNAPSHOT_HASHES: {{SNAPSHOT_HASHES_JSON}}
TRAINING_ELIGIBILITY_POLICY: {{TRAINING_ELIGIBILITY_POLICY_JSON}}

Task

Create policy-eligible training and replay records from frozen, auditable episodes.

Record types:

• positive model, derivation, experiment, measurement, or simulation outcome;
• negative example with exact falsification evidence;
• repair pair from failed to corrected artifact;
• counterexample-search and adversarial-experiment record;
• model-selection comparison;
• curriculum replay with deterministic evaluator.

Rules:

• include only promoted or explicitly eligible bounded artifacts;
• preserve input, output, source, library, model, prompt, tool, world, and evaluator snapshot IDs;
• include concise certificates, findings, and receipts, never hidden chain-of-thought;
• keep learner, validator, historian, and safety-only fields separated;
• exclude unauthenticated experiments, contaminated clean-room episodes, unresolved overfit, false status labels, unsafe procedural detail, and licensing-restricted data;
• retain negative and repair records for falsifier training without adding them to trusted positive knowledge;
• assign weights transparently and record exclusion reasons.

Return only JSON matching schemas/training-record-batch.schema.json.