[hhff]

PR 5 of 5 for the inspectable OODA loop (spec and plan in #193). Stacked on #203 (base ooda-knowledge). It closes the owner's question: is the loop making the system better informed, more autonomous and more trusted, week by week?

Adoption gate: same as #203

This PR adds two top-level observation fields, week and figures. Merge it only when every instance has v0.37.0 or later. Today stacksbot is on v0.33.2.

What changes

• The weekly learning record. Each week, observe-metrics writes one observation for the past week, and on first run it back-fills one per past week. Each record has kind: learning, source: metrics, subject: project, its week (yyyy-Www) and source_key: learning:<week>. Its figures are counts, never shares. Its refs hold up to 60 of the records it counted, and its detail is one sentence. Records can't be back-dated, so each one carries today's at and its figures come from replaying the log up to the end of its week (withLogAt).
  • Informed: accepted knowledge; recurring signals formed and explained; starts, and starts given a lesson.
  • Autonomous: ladder calls, and how many were made on trust; owner questions; merged tickets.
  • Trusted: answers with a recommendation, and how many agreed; calls, and corrections.
  • The definitions are in the plan's "PR 4 and 5 review fixes", item 5. Only observe-metrics writes these records; reading the dashboard never does.
• The chart at the top of the Loop. Six small charts, two per column, with weekly points:
  • Informed: accepted knowledge; share of signals explained, and of starts given a lesson.
  • Autonomous: share of ladder calls made on trust; questions per merged ticket.
  • Trusted: agreement; corrections per 100 calls.
  • A point whose denominator is under 5 is drawn hollow, titled "too few to say".
  • Retros and rung changes appear as dashed vertical marks.
  • Each point opens its week's learning sidepeek.
  • With fewer than three records, the Loop shows tiles and a sentence saying trends need a few weeks.
  • lineChart gains hollow and linked points and xmarks; its existing callers draw exactly as before.
• The learning sidepeek: the week's figures, every record it counted, and lesson efficacy. For each lesson given to a start, it lists the observations on its evidence's keys in the four weeks before the lesson was first given and since then. This is a list with ids, not a score.
• Docs: supervisor.md §3, and a note in operator.md's daily review.

For the operator's review (#190 follow-up)

The review can read:

• log --type observation --json: the learning records, with figures and week
• dashboard-data: ooda.learning
• peek learning <id> --json: the counted records and lesson efficacy

Tests

• test/ooda-learning.test.mjs:
  • back-fill over three weeks, with figures counted by hand
  • writing twice adds nothing
  • dashboard-data never writes a record
  • the peek's lists and efficacy rows
  • the chart code run in vm: tiles, six charts, hollow points, links, marks, unchanged old output
• One Chrome check in test/dashboard.test.mjs.
• log-tolerance.test.mjs now uses a different made-up field (insights), since figures is real.
• G1 golden re-recorded; only ooda keys were added.
• Full suite on Node 24: 427 passed, 0 failed, 1 existing skip.
• Speed: on a 14,000-line, 45-week log, the first back-fill adds about 1.3 s, once.

One-way doors

None. Additive fields, behind the adoption gate.

🤖 Generated with Claude Code

[hhff]

The weekly learning record and the Loop's chart, OODA part 5 of 5. Review fixes from the author's pass: the line fits LINE_MAX, the back-fill is capped at 104 weeks, writeLearning runs under a lock with a dedupe on read, nested figures are left open for forward compatibility, malformed figures are guarded, and only the supervisor's records count as written. Its CI flake was the dashboard refresh stall diagnosed separately, not this code. CI 9/9 at df1cc9c.