Skip to content

Raise, fix and recheck a defect

Principles in play: Evaluate corrections.

The defect ledger connects a detected failure to its repair and to a targeted recheck of the original problem. A later review that does not repeat the complaint does not establish that the correction worked. Who raised the finding, who repaired it and who rechecked it stay on record separately.

  1. Raise it, quoting what is wrong, and name everything it reaches.

    Say to your agent
    The overall share in the results note is the mean of the yearly percentages, not total cases over total patents. Open a defect row for it in my name, quoting that line. The same code builds the summary table, so list both.

    If you returned the work, your return already opened the row, with the outputs you named. A finding from a model review that you reproduced gets its row too, taken from the review’s report, before anyone fixes it.

  2. The owner repairs it.

    Say to your agent
    Fix the defect in my results note and mark the row fixed in my name.

    The commit that makes the fix names the row after Fixes, so the report can join the correction to its recheck.

  3. The raiser runs a targeted recheck, with their own evidence.

    Say to your agent
    Recheck the fix against the original criterion: recompute the overall share from the two count columns of the table.

    Each output the row lists gets its own recheck.

  4. Read the recheck back.

    Say to your agent
    Show me the recheck field by field before you record it.

    A field you did not state reads not stated, and nobody asks you for more.

  5. Confirm it.

    Say to your agent
    Run it.

    The row reads verified once every output it lists has a passing recheck; until then it says how many are rechecked and which still wait.

  • Closed is not verified. closed means a later review did not raise the finding again; verified means a recorded targeted recheck supports the correction. A row someone closes by hand is a repair claim, and it waits for a recheck.
  • Ids follow your project. A new project’s row ids carry the prefix DEF, unless ai/project.json sets another.
  • A self-check is labelled. A checker who also made the correction is recorded as a self-checker.
  • Verified is not accepted. A verification does not accept the work or authorize publication.
What the agent runs
Terminal window
python3 scripts/defect_ledger.py open --label TEXT --severity high|medium|low \
--quote "..." --source <file> --by <name> --object <output> --object <output>
python3 scripts/defect_ledger.py ingest --run ai/reviews/<run> --finding <id> --matches <file> # a review's finding
python3 scripts/defect_ledger.py mark --id <ID> --fixed --by <fixer> --note "<the fix>"
python3 scripts/defect_ledger.py draft-recheck <ID> --checker NAME --person --object <output> \
--method "..." --outcome "..." --limitations "..." --said "..." # one per listed output
python3 scripts/defect_ledger.py verify <ID> --evidence <file>
python3 scripts/defect_ledger.py check

This page describes TRACE-OS 0.1.0-rc23, a release candidate prepared on 2026-10-15.