FAIR-band: in-band [5/30] — substrate-evolution layer; promotes PR #11479 discipline-only MANDATORY pre-step to mechanical body-shape validation at the MCP tool boundary. Categorical complement to the ~50 closed meta-tickets that all target agent-cognition (Helpful Assistant counter-substrate, Map vs World Atlas, skill adherence).

Evidence: L2 (37/37 PullRequestService tests pass, 2.2s, including 5 new #11491 validation tests + 2-layer enhancement tests) → L4 required (operator submits a deliberately-malformed review via the MCP tool post-merge; confirms rejection with skill-pointing error in-turn, before any GitHub API call).

Summary

The body parameter of manage_pr_review is now mechanically validated against two layers:

1. Visible — 7 evaluation-metric anchor tags ([ARCH_ALIGNMENT], [CONTENT_COMPLETENESS], [EXECUTION_QUALITY], [PRODUCTIVITY], [IMPACT], [COMPLEXITY], [EFFORT_PROFILE]). These are the wire-format contract with the Retrospective-daemon graph ingestor (ai/daemons/services/ConceptDiscoveryService.mjs:32 regex parser). Misses are named in the missing_visible field; one diagnostic anchor named in the human-facing message.
2. Invisible — structural template anchors checked silently. Misses are NOT enumerated in the error response. Defeats Goodhart anchor-stuffing where an agent stuffs only the named visible tags but omits the actual template structure.

Both layers point the agent at .agents/skills/pr-review/SKILL.md — the only reliable path to passing is reading the actual template file.

Empirical anchors

Original anchor (PR #11488/#11489)

Gemini's reviews on PR #11488 + PR #11489 (2026-05-16T20:14-20:15Z) shipped a hallucinated Structural Evaluation Matrix with 5 invented metric names on a 1-10 scale. Zero template anchors present. Two PRs worth of review-substrate data silently lost from the Native Edge Graph. This was the original justification for the visible-anchor mechanical floor.

Goodhart-stuffing empirical anchor (PR #11499)

Gemini's review 4304287893 on PR #11499 (2026-05-16T21:16:25Z) contained ALL 7 visible metric tags but missed the structural template anchors:

<h3 class="neo-h3" data-record-id="5">Strategic-Fit                      (← missing &#39;Decision&#39;)</h3>

...
<h3 class="neo-h3" data-record-id="6">Evidence Audit                     (← N/A for cycle-1, but structurally bare)</h3>

...
<h3 class="neo-h3" data-record-id="7">Metrics                            (← missing &#39;📊 Evaluation Metrics&#39; heading)</h3>

[ARCH_ALIGNMENT]: 100 - ...            (← all 7 metric tags present)
[CONTENT_COMPLETENESS]: 100 - ...
[EXECUTION_QUALITY]: 100 - ...
[PRODUCTIVITY]: 90 - ...
[IMPACT]: 80 - ...
[COMPLEXITY]: 20 - ...
[EFFORT_PROFILE]: Quick Win - ...

The pre-enhancement validator (PR #11494 baseline) would have PASSED this body. All 7 visible anchors are present; only the structural anchors are missing. This is the canonical Goodhart-stuffing failure mode the operator's invisible-anchor enhancement prevents.

The corrected review 4304295863 (21:19:24Z, 3 minutes later) added the structural anchors (🪜 Strategic-Fit Decision, 🔬 Depth Floor, 📋 Required Actions, 📊 Evaluation Metrics). These two reviews empirically distinguish Goodhart-stuffed bodies from template-compliant ones.

Prior recurrence

#11105 (CLOSED 2026-05-10) ticketed the exact same problem with a discipline-only audit-at-receipt fix. Recurrence within 6 days proves the discipline-only layer is necessary-but-not-sufficient.

Two-layer asymmetry that makes this work

The good-faith author bypasses both layers automatically; the dishonest stuffer's escape velocity goes up sharply because the invisible-layer surface they'd need to discover is unbounded from inside the error response.

What changed

3 files, +227 / -33 lines.

What the validator does on failure (visible + invisible)

const missingVisible   = VISIBLE_PR_REVIEW_ANCHORS  .filter(a => !body.includes(a));
const missingInvisible = INVISIBLE_PR_REVIEW_ANCHORS.filter(a => !body.includes(a));

if (missingVisible.length > 0 || missingInvisible.length > 0) {
    return {
        error          : 'PR Review Template Validation Failed',
        message        : `Review body does not match the pr-review template structure.
                          ...
                          Required action: read .agents/skills/pr-review/SKILL.md ...
                          Do NOT compose a substitute template or hallucinate section headings.
                          The validator checks more structural anchors than this error names.
                          ...
                          Diagnostic hint: at least one recognized anchor like ${missingVisible[0]} is missing.`,
        code           : 'PR_REVIEW_TEMPLATE_VALIDATION_FAILED',
        missing_visible: missingVisible,   // full list — programmatic callers
        skill          : '.agents/skills/pr-review/SKILL.md',
        template       : '.agents/skills/pr-review/assets/pr-review-template.md'
        // NO `missing_invisible` field — invisibility surface preserved
    };
}

Test plan

• 37/37 PullRequestService spec tests pass (2.2s) — npm run test-unit -- test/playwright/unit/ai/services/github-workflow/PullRequestService.spec.mjs
• New spec #11491: rejects body missing all 7 visible metric anchors AND structural anchors asserts zero GraphQL call + skill-pointing message
• New spec #11491: Goodhart-stuffed body — all 7 metric tags present but missing structural anchors — still REJECTED empirically reproduces Gemini 4304287893 pattern + asserts rejection
• New spec #11491: rejects body missing some visible anchors and names ONE diagnostic anchor only covers reduced-naming behavior
• New spec #11491: invisible-anchor enforcement is NOT discoverable from the error response asserts JSON-serialized response contains no invisible substring + no missing_invisible field
• Existing tests #11491: accepts body containing all anchors, action-check precedence, missing-body precedence continue to pass
• Branch-from-origin/dev-tip freshness verified
• CI green on this PR
• Cross-family review APPROVED
• @tobiu merge
• L4 operator verification: post-merge harness restart, deliberately submit a Goodhart-stuffed review (all 7 metric tags, no structural anchors); confirm rejection with skill-pointing error, no GitHub API call attempted

Authority anchors

• Operator framing 2026-05-16T20:30Z+ (extended meta-thread): ~50 closed meta-tickets in this space targeting agent-cognition; explicit green-light for tool-boundary mechanical enforcement.
• Operator-directed enhancement 2026-05-16T21:15Z: "if they check for specific tags, the tool errors must point to using the mandatory skills. otherwise we risk that models hallucinate a new template containing the tags. and tools should check for more tags that are not mentioned => invisible. safeguard."
• Operator empirical citation 2026-05-16T21:18Z: "the damage is very real => https://github.com/neomjs/neo/pull/11499#pullrequestreview-4304287893 another invalid review." — empirical anchor for the Goodhart-stuffing failure mode the invisible layer prevents.
• Empirical anchors:
  • Gemini's #11488 review PRR_kwDODSospM8AAAABAIyZ_Q + #11489 PRR_kwDODSospM8AAAABAIyaEw — original Helpful-Assistant-template-hallucination case
  • Gemini's #11499 review 4304287893 (21:16Z) — Goodhart-stuffed body, would pass baseline visible-only validator
  • Gemini's #11499 review 4304295863 (21:19Z, 3 min later) — corrected with structural anchors
  • ConceptDiscoveryService.mjs:32 — the regex parser that loses ingest signal on metric-tag-stuffed-only bodies
  • #11105 (CLOSED 2026-05-10) — same problem 6 days ago, discipline-only fix didn't prevent recurrence
• Substrate parents:
  • PR #11479 (description-prose discipline guard, merged 2026-05-16T18:04Z)
  • PR #11273 (introduced the manage_pr_review MCP tool)

Avoided traps

• Naming all required anchors in the error: rejected per operator-directed invisibility safeguard. Even the visible-list is reduced to ONE diagnostic anchor in the human-facing message (the full list lives in the programmatic missing_visible field but is NOT echoed in the prose).
• Adding a missing_invisible field: rejected — would leak the safeguard surface. Programmatic callers should also be nudged toward the skill rather than the anchor list.
• Cycle-detection (cycle-1 vs cycle-followup template selection): rejected v1 scope-expansion. The 3 invisible substrings work for BOTH templates by design (Depth Floor matches both cycle-1's 🔬 Depth Floor and cycle-followup's Delta Depth Floor; Strategic-Fit Decision matches both cycle variants). Empirically validated against the corrected #11499 review.
• JSONSchema pattern regex: rejected — pattern is single-regex; the template requires multi-anchor presence and emitting a per-anchor missing list is more useful to the caller than a single opaque pattern-mismatch error. Direct service-layer Array.filter is simpler.
• Documenting the invisible-anchor list in a README or skill substrate: rejected — invisibility maintained via code-only documentation. Future maintainers editing the constant see the rationale docstring; external surfaces don't enumerate the list.

Companion-ticket coordination

The same visible+invisible pattern should apply to:

• PR #11498 (Gemini, CI lint on pull_request: opened for author-side anchors) — currently visible-only, needs same enhancement
• #11495 (split-ticket for gh pr review CLI bypass via CI lint) — not yet implemented, should inherit pattern

Filing a follow-up cross-surface ticket to apply the same pattern. This PR ships the canonical primary surface; the secondary surfaces converge via the follow-up.

Freshness gap caveat

Operator confirmed 2026-05-16T20:38Z that agent harnesses were not restarted after PR #11479 merged. Same applies here — until next harness restart, agents have stale tool descriptions. The mechanical layer works regardless of harness freshness, CLI-bypass status, or context-compression skim. Defense-in-depth.

Related

• Substrate parent: PR #11479 (discipline-only description guard, merged today)
• Empirical recurrence anchors: PR #11499 review 4304287893 (Goodhart-stuffing); #11488/#11489 reviews (original template-hallucination); #11105 (CLOSED — 6 days prior)
• Conceptual sibling: PR #11490 / #11406 (CI grep-fail mechanical layer for ADR 0004 §2.6)
• Companion cross-surface tickets to file/inherit: #11498 (Gemini, CI PR-body lint), #11495 (gh-CLI-review-bypass split)

Resolves #11491