Affected Repositories

• PyAutoBrain (primary) — AUTONOMY.md, tests/test_autonomy_doctrine.py
• PyAutoMind — autonomy_log.md

Suggested branch: feature/glance-tier-ruling

Evidence

Audit page Science/scratch/shadow_window_audit_2026-11-03/audit.md (8 rows, verdicts all ok, 2026-11-04). Pre-registered rule: complete/2026/09/prm-shadow-row-notify-tier.md.

Key checks

• Revert-condition count in AUTONOMY.md is pinned by a doctrine test (was 12 after 3a) — bump it.
• Lifecycle tests may parse the Shadow table — keep the stage-1 table format unchanged; new table under its own heading.

Record the glance Shadow-window ruling in AUTONOMY.md — grant narrowed, after the spot audit

Type: feature
Target: PyAutoBrain
Repos:

• PyAutoBrain
• PyAutoMind
  Difficulty: small
  Autonomy: human-required
  Priority: high
  Status: formalised
  Filed: 2026-11-03
  Closes-when: the human returns verdicts on /mnt/c/Users/Jammy/Workspace/Science/scratch/shadow_window_audit_2026-11-03/audit.md
  Consequence: judge
  Witness: AUTONOMY.md carries a dated section recording the human's ruling on the tier-glance Shadow window (69/20, all stage-1 merged-unchanged) with its revert condition, and a doctrine test pins (i) a witness that is not PASS (none, n/a, partial, waived) never auto-merges, (ii) the audit's outcome, and (iii) unattended (watcher) glance merges count in a separate stage-2 sub-window; autonomy_log.md's Shadow header states the ruling; pytest PyAutoBrain/tests -q green
  Review-minutes: 10
  Unattended: never
  Lane: any

Decided with the human 2026-11-03 (recommendation accepted): grant glance, narrowed, after
a spot audit; start a separate stage-2 count for unattended merges.

The pre-registered rule (complete/2026/09/prm-shadow-row-notify-tier.md, "The decision rule,
pre-registered"): grant if no merged-after-substantive-change / not-merged rows; grant
narrowed if every case falls in one identifiable sub-class; refuse otherwise. The window stood
at 69 counted rows (target 20), all merged-unchanged, all stage 1, about 21 with no
independent review, most under a non-GREEN Heart.

Steps:

1. Spot audit (human). Eight rows were picked across every risk class, in
   Science/scratch/shadow_window_audit_2026-11-03/audit.md. The human writes ok or
   minded — <why> per row. Any minded row → amend that Shadow row to
   merged-after-substantive-change and let it define the narrowed-out sub-class.
2. Record the ruling in AUTONOMY.md, as a dated section after "Merge authority follows
   Consequence — 2026-10-02":
   • glance is confirmed for in-turn auto-merge (the shipping session, human nearby);
   • narrowing (both already proposed): a witness that is not PASS — none, n/a, partial
     or waived — means human /prm; and, if the audit finds tier drift (rows 3, 4 and 6 are
     public helpers or APIs and a library return-value change), anything touching a public API,
     a default or a library return value is judge whatever the declared tier;
   • unattended glance (watcher-started close-outs, AUTONOMY "Shift briefs and unattended
     glance") opens its own stage-2 sub-window: its first 10 merges each need the
     independent different-model adversary leg actually run, are counted apart from the
     stage-1 rows, and gate that sub-rule;
   • the revert condition: any reverted row still demotes glance immediately.
3. Ledger accuracy: a counted row whose merge was a human-typed /prm (row 8 suggests
   fixed-model-json, PyAutoFit#1708) is re-marked as not an auto-merge, and the count fixed.
4. autonomy_log.md: the Shadow header states the ruling and opens the stage-2 sub-window
   table.

Related: the autonomous-orchestrators epic depends on this for unattended merges; its pilot
(phase 4) must not auto-merge unattended until step 2's stage-2 sub-window exists.