History Coverage Audit

This audit tracks how much of posttrainllm’s historical work is normalized into the structured attempt ledger, classified as technique inventory, or intentionally left as narrative/reference material.

The goal is precision, not completeness theater. If an old note lacks enough context to recover a real attempt, it receives a final narrative/reference classification rather than becoming fabricated evidence or unfinished work.

Current Structured Coverage

Source of truth: attempts.json.

Confidence Coverage

Normalized Families

Classified Non-Ledger / Partial Surfaces

These docs contain useful evidence or lessons, but not every row is a run attempt. The table records their exact treatment so we do not inflate the attempt ledger with implementation facts or learning notes.

Classified Historical Sources

Backfill Rule

Add an old attempt to docs/attempts.json only when all fields can be filled:

• id
• name
• family
• status
• evidence
• failure_reason_confidence
• lesson
• next_action
• evidence_sources

For failed, regressed, worked-with-caveat, and inconclusive, also require failure_reason.

If the source docs conflict on a number, do not hide the conflict. Either quote both values in evidence and use inferred, or leave the attempt out until the original run artifact resolves it.

Current Honest Readout

The structured history now covers every known evidence-backed experiment in the factory, browser, Mac-runtime, model, external-runtime, and benchmark lanes. Broad technique rows remain classified in docs/techniques/audit-inventory.md. Learning-journal and archive narrative entries are intentionally not attempts unless they point to a real run, eval, artifact, or PRD decision; that is a final classification, not an implied experiment backlog.