Precursor Timeline

Engine: PMLC = python3 <skills-root>/learnings-coach/scripts/pmlc.py (this skill needs the learnings-coach skill installed alongside it).

Gather candidates

PMLC timeline <ws> --before incident

This lists every dated mention in findings and evidence (<ws>/timeline_candidates.json), with snippets and sources. Many are duplicates or context dates (e.g. a commissioning year). The engine finds dates; you decide which are signals.

Curate

For each candidate keep it only if it is a signal: something that could have told someone the risk was rising and that they could actually see then (see the knowable-then test in learning-principles.md §1a). Examples: an inspection that found degradation, a repeated repair on the same item, a skipped or overdue maintenance schedule, an audit recommendation left open, a procedure published but not used, a configuration change not documented. Record:

Add the event sequence on the day itself as a final block if the narrative gives it.

Learn

Close with three to six signal learnings: patterns in how signals were handled (no escalation path, closure without verification, schedules that slide when a job is late, repeated repairs treated as normal). Each one names the entity and tier that must change the way it handles signals, and a leading indicator.

Output

Write the pack <run>/outputs/precursor-timeline.json (schema: learnings-coach/references/pack-schema.md), then:

PMLC check-pack <ws> <run>/outputs/precursor-timeline.json   # fix every problem it lists
PMLC pack <ws> <run>/outputs/precursor-timeline.json         # -> precursor-timeline.html (primary, vertical timeline) and .md

Sections and the kinds that fill them:

• Curated signals: one timeline. Each row of the Curate table above becomes an event:
  • Date → date
  • Signal → text; do not write "(N months before)": the renderer computes it from the incident date and shows year-only dates as "c. YYYY (about N years before)"
  • signal: true
  • Received by → received_by
  • Response → response ("None recorded" is an answer)
  • Knowable then → knowable_then (optional text, or true or false)
  • Evidence → cites
• Event sequence on the day: a second timeline, in its own section.
• Signal learnings: learning ×3–6, with owner = the entity and tier that must change. Follow each with its indicator.

Put data-quality corrections in data_quality, findings left out on purpose in exclusions, and same-condition findings in supporting_cites.

{"date": "2022-03", "text": "…", "signal": true, "received_by": "Materials Management (F06, T1)",
 "response": "None recorded", "cites": ["FND-…", "EVD-…"]}