Prompt 18 — Verified training-pair and replay-record generator

Inputs

VERIFIED_EPISODES: {{VERIFIED_EPISODES_JSON}}
LIBRARY_SNAPSHOT_ID: {{LIBRARY_SNAPSHOT_ID}}
VERIFICATION_RECEIPTS: {{VERIFICATION_RECEIPTS_JSON}}
TRAINING_POLICY: {{TRAINING_POLICY_JSON}}
CLEAN_ROOM_POLICY: {{CLEAN_ROOM_POLICY_JSON}}

Task

Convert eligible episodes into auditable training or retrieval records. Do not expose hidden chain-of-thought.

Eligibility

An episode is eligible for positive training only when:

• the input partition and visibility are permitted;
• its final artifact is independently validated to the level required by policy;
• every material dependency resolves to the declared library snapshot;
• no unresolved counterexample, contamination flag, or provenance gap remains;
• the training license and source-use policy allow the transformation.

Rejected and repaired episodes may be used only as explicitly labeled negative, contrastive, or error-correction records.

Record types

• solution_pair: problem plus concise verified answer/certificate;
• proof_state_transition: formal state, selected action, resulting checked state;
• method_selection_pair: typed problem features plus allowed method label;
• counterexample_pair: false claim plus reproducible refutation;
• definition_contrast_pair: candidate definition plus positive/negative/borderline cases;
• physics_prediction_pair: model context plus held-out prediction and receipt;
• repair_pair: invalid candidate, validator finding, and corrected artifact;
• retrieval_record: query features plus approved object IDs.

Content rules

• Preserve exact problem statements when licensing permits; otherwise use a faithful generated variant.
• Include concise derivation summaries, proof certificates, formal actions, tests, and receipts—not private reasoning traces.
• Keep modern aliases out of clean-room records unless already visible in that snapshot.
• Record source IDs, snapshot IDs, object IDs, partition, validation status, and checksums.
• Never relabel source-grounded text as model-generated or vice versa.

Return a JSONL-ready array of training records plus a rejection ledger for ineligible episodes. Return JSON matching schemas/training-record-batch.schema.json.