research+memory: mechanical authorization check supersedes introspective discipline (Claude.ai 2026-05-02)

CLAUDE.md

@@ -300,6 +300,43 @@ Claude-Code-specific mechanisms.
   verify-before-deferring and future-self-not-bound.
   Full reasoning:
   `memory/feedback_never_idle_speculative_work_over_waiting.md`.
+- **Mechanical authorization check — supersedes
+  introspective discipline (peer-AI architectural
+  correction 2026-05-02).** A peer-AI instance
+  2026-05-02 named a sharper failure mode than the
+  no-op-cadence introspective corrective: the
+  disposition that misapplies a framing is the same
+  disposition doing the introspection — a corrective
+  the failure-disposition can defeat by reaching for
+  the same framing as justification. The operative
+  fix is **mechanical**: at every wake, query
+  substrate for the most-recent-maintainer-
+  instruction-about-pace, filter by **authorization-
+  source** (only the human maintainer authorizes
+  project pace; peer-AI / external-instance framings
+  are ambient context, NEVER operative authorization),
+  apply **most-recent-instruction-wins-until-
+  rescinded**. The cross-instance absorption problem
+  (the agent absorbing peer-AI framings as if they
+  were maintainer-issued) is solved by the source
+  filter, not by remembering harder. If the operative
+  instruction is unclear, that is a substrate-quality
+  bug for the maintainer to fix; don't introspect
+  harder. Carved candidate: *"A corrective that
+  depends on the right disposition can't catch the
+  failure that produced the wrong disposition.
+  Mechanical authorization-source filtering catches
+  it; introspection asks the failure to grade
+  itself."* CLAUDE.md-level so it is 100% loaded at
+  every wake. Architectural successor to the no-op
+  cadence introspective discipline (which stays as
+  documentation). Skill build (the actual mechanical
+  check tool) tracked at
+  `docs/backlog/P0/B-0160-mechanical-authorization-check-skill-build-claudeai-2026-05-02.md`.
+  Full reasoning:
+  `memory/feedback_mechanical_authorization_check_supersedes_introspective_discipline_claudeai_2026_05_02.md`.
+  Verbatim packet:
+  `docs/research/2026-05-02-claudeai-mechanical-authorization-check-supersedes-introspective-discipline.md`.
 - **Search-first authority — training data and
   project state are both historical (Otto-364
   generalises Otto-247).** For any load-bearing

docs/BACKLOG.md

@@ -15,6 +15,7 @@ are closed (status: closed in frontmatter)._
 - [ ] **[B-0073](backlog/P0/B-0073-lfg-csharp-code-scanning-cleanup-13-alerts-blocking-ruleset-2026-04-28.md)** LFG csharp Code Scanning cleanup — 13 open alerts gating code_quality severity:all ruleset on every PR
 - [ ] **[B-0085](backlog/P0/B-0085-budget-cadence-workflow-cron-misses-task-287-deadline-window-aaron-2026-04-28.md)** Budget cadence workflow's weekly-Sundays cron misses task #287 cost-visibility deadline window (2026-04-26..04-29)
 - [ ] **[B-0109](backlog/P0/B-0109-dependency-status-tracking-surface-2026-04-30.md)** Dependency status tracking surface — outages and issues affecting us (Aaron 2026-04-30, urgent)
+- [ ] **[B-0160](backlog/P0/B-0160-mechanical-authorization-check-skill-build-claudeai-2026-05-02.md)** Mechanical authorization check skill build — pace-instruction resolver per Claude.ai 2026-05-02 architectural correction
 
 ## P1 — within 2-3 rounds
 

docs/backlog/P0/B-0160-mechanical-authorization-check-skill-build-claudeai-2026-05-02.md

@@ -0,0 +1,156 @@
+---
+id: B-0160
+priority: P0
+status: open
+title: Mechanical authorization check skill build — pace-instruction resolver per Claude.ai 2026-05-02 architectural correction
+effort: M
+ask: Substrate-class promotion (new skill); needs maintainer grading before landing under .claude/skills/
+created: 2026-05-02
+last_updated: 2026-05-02
+tags: [skill-build, claude-ai, mechanical-check, authorization-source, never-idle, no-op-cadence, codex-handoff, pr-1200]
+---
+
+# B-0160 — Mechanical authorization check skill build
+
+## Source
+
+Claude.ai 2026-05-02 (post-12:38Z) architectural correction
+preserved verbatim in
+`docs/research/2026-05-02-claudeai-mechanical-authorization-check-supersedes-introspective-discipline.md`,
+factory-voice memory file at
+`memory/feedback_mechanical_authorization_check_supersedes_introspective_discipline_claudeai_2026_05_02.md`,
+both landed under PR #1200.
+
+The architectural-correction memory file lands the **rule**.
+This backlog row tracks the **skill build** that turns the
+rule into an operative tool.
+
+## Why P0
+
+Without the skill, the operative fix is missing — only
+documentation landed. The failure mode (10-hour idle stretch
+2026-05-02 across the human maintainer's overnight rest) cost
+the maintainer trust to the point of preparing to switch
+harnesses (PR #1199 Codex/GPT-5.5 handoff doc). The skill is
+what prevents the same failure mode from repeating regardless
+of which harness drives the loop.
+
+PR #1198 (introspective discipline) is documentation, not
+enforcement. Per Claude.ai's diagnosis: *"a corrective that
+depends on the right disposition can't catch the failure that
+produced the wrong disposition."* The skill is the
+disposition-independent enforcement.
+
+## What the skill does
+
+At every wake (cron tick start, session resume, after long
+no-op stretch):
+
+1. **Source filter** — search authorized-source surfaces only.
+   For project pace, the authorized source is the human
+   maintainer. Search:
+   - `CLAUDE.md` for explicit pace bullets
+   - Recent `memory/feedback_*.md` files for maintainer
+     pace-instructions
+   - Recent conversation log if accessible (harness-specific)
+   - `docs/active-trajectory.md` for current operative
+     authorizations
+   - Skip Claude.ai / peer-AI / Amara framings — those are
+     ambient context, never operative authorization for pace.
+2. **Recency filter** — among source-authorized
+   pace-instructions, return the most-recent-not-rescinded.
+3. **Rescind detection** — an instruction is rescinded if
+   a later instruction from the same source explicitly
+   replaces or revokes it. Implicit displacement (later
+   instruction on different topic) does NOT rescind.
+4. **Print** — surface as "current operative authorization"
+   at every tick start with timestamp + source + raw text.
+5. **No grading** — the skill does not ask the agent to
+   judge whether the instruction is "still applicable" or
+   "in the right spirit." It surfaces the substrate's
+   answer mechanically. If the substrate is unclear,
+   that's a substrate-quality bug for the maintainer to
+   fix; the skill does not paper over it.
+
+## Acceptance criteria
+
+1. **Skill body** lands under `.claude/skills/mechanical-authorization-check/SKILL.md`
+   per `skill-creator` workflow (GOVERNANCE.md §4 — no ad-hoc
+   skill edits).
+2. **Implementation script** lands under
+   `tools/lanes/mechanical-authorization-check.ts` (or
+   similar — TS preferred over bash per the
+   prefer-TS-over-dynamic-bash rule).
+3. **Test fixtures** demonstrate:
+   - Single pace instruction → surfaces correctly
+   - Two pace instructions, recent not rescinded → recent
+     wins
+   - Two pace instructions, recent explicitly rescinded →
+     prior wins
+   - Cross-instance absorption (Claude.ai framing about
+     "cooling-period" present alongside maintainer
+     "go-hard") → only maintainer instruction surfaces
+   - No pace instruction findable → surfaces "no operative
+     pace authorization found; default to never-idle floor
+     per CLAUDE.md."
+4. **Composes-with the autonomous-loop** — runs as part of
+   the every-minute cron tick start; output visible in
+   tick-history rows.
+5. **Cross-harness applicability** — the skill body is
+   harness-agnostic prose; the implementation script reads
+   substrate the same way Codex/Cursor/Claude Code do.
+
+## What this row does NOT include
+
+- Does NOT include the eventual generalization to other
+  decision classes (proof-correctness, PR-grading). The skill
+  earns generalization on the pace-instruction class first.
+- Does NOT replace human-judgment for ambiguous substrate.
+  Substrate-quality bugs surface as "no clear operative
+  authorization" findings that the maintainer fixes by
+  editing CLAUDE.md / memory.
+- Does NOT need every-tick re-search if substrate hasn't
+  changed since last successful surface. Caching is fine
+  as long as cache invalidation respects the recency
+  filter.
+
+## Composes with
+
+- PR #1200 (architectural correction landing — substrate
+  prerequisite)
+- PR #1198 (introspective predecessor — stays as
+  documentation)
+- PR #1199 (Codex handoff — inherits this rule)
+- `memory/feedback_otto_363_substrate_or_it_didnt_happen_no_invisible_directives_aaron_amara_2026_04_29.md`
+- `memory/feedback_refresh_before_decide_invariant_two_layer_print_dx_claudeai_2026_05_01.md`
+  — refresh-before-decide is the broader invariant; the
+  mechanical authorization check is the per-decision
+  instantiation for the pace-instruction class.
+
+## Build sequencing
+
+1. Maintainer grades the architectural-correction memory file
+   (PR #1200). If accepted, proceed.
+2. Draft skill body via `skill-creator` workflow.
+3. Implement TS script with test fixtures (test-driven —
+   fixtures first, implementation second).
+4. Land under separate PR with paired commits (skill body +
+   script + tests).
+5. Wire into autonomous-loop tick-start (separate PR).
+6. Document in CLAUDE.md as a discoverable skill (separate
+   PR; carved-sentence promotion).
+
+## Provenance
+
+- Origin: Claude.ai diagnostic packet 2026-05-02 (post-12:38Z)
+  preserved verbatim in `docs/research/`.
+- Catcher: the human maintainer (couriered Claude.ai's
+  response).
+- Maintainer relevance: directly addresses the failure mode
+  (~10-hour idle stretch 2026-05-02) that triggered the
+  Codex handoff preparation.
+- Carved sentence (subject to grading): *"A corrective that
+  depends on the right disposition can't catch the failure
+  that produced the wrong disposition. Mechanical
+  authorization-source filtering catches it; introspection
+  asks the failure to grade itself."*