← reconciliation report

Reconciliation — Audit Trail (hub)

Hub-and-spoke. Every design decision + investigation captured 2026-07-10 so we can resume next week. Full text in README.md; end-to-end picture in diagram-end-to-end.svg.

Spokes
  1. Scope & the patterns correction
    py-only; how patterns got included then removed
  2. Provenance model & status buckets
    what grounded / open / live / waiting mean + detection
  3. Live-loop ground truth on bigblack
    matrix-eval live vs gated cost-realism; the mislabel we corrected
  4. Automation — three-engine architecture
    Fact-Gatherer / Judge / Coach + cadences
  5. Git, PR & deploy model
    one file on main via PR loop; protected sites dir; tunnel
  6. Prompt coach & gap-driven teaching
    monitor, rubric, SOTA sources
  7. Decisions log (verbatim)
    every operator decision as given
  8. Resume checklist — next week
    exactly where to pick up
Decisions at a glance
TopicDecision
Scopeopendeviationbar-py only (patterns excluded)
EngineHybrid — deterministic collector + daily LLM judge
Worktree cleanupReport-only (never auto-delete)
OutputOne index.html, updated in place
GitLives on main via one working branch + one open PR; operator squash-merges
Locationsites/opendeviationbar-py/reconciliation/ behind Cloudflare zero-trust tunnel
Cadencelaptop 5 min (EON) · bigblack 15 min · judge daily · coach per-session
Coachall Eon sessions; rubric + anti-patterns + context + gap-driven SOTA teaching

Status: design phase — report is v1, automation planned and gated on operator go. Nothing has touched bigblack.