External Review — golden-rule

Cycle: 2026-06-04 Reviewer: External Review Partners (retained) Subject audit: reports/auditor/2026-06-04-golden-rule-audit.json Agent source: .claude/agents/golden-rule.md Rubric version: a00e133872b463ce

1. Audit Quality Assessment

Arithmetic

Score: 1 + 1 + 0 + 0 + 0 = 2/10. Correct. Max is 10 (five dimensions × 2 points). No arithmetic error.

Floor rule application

eval=absent triggers non-conformant regardless of total. Rubric line 79 reads exactly: “eval absent => non-conformant regardless of other scores.” Applied correctly. Verdict stands.

Cited clauses

All five cited_clause values appear verbatim in rubric.a00e133872b463ce.yml:

• 01:one-sentence-job — present in charter dimension
• 02:versioned-header — present in prompt dimension
• 03:toolspec-doc — present in tools dimension
• 04:eval-suite-exists — present in eval dimension
• 05:model-lock — present in currency dimension

No invented clause IDs. No false citation.

pre_canon flag

pre_canon: true is correct. golden-rule.md is a single file at .claude/agents/golden-rule.md with no charter directory, no TOOLSPEC.md, no eval suite, no CHANGELOG.md. The agent predates the multi-artifact canon layout. Flag is appropriate; this is a predates situation, not defies.

False-Present check

No false-Present verdicts. Both partials (charter, prompt) are substantiated by what is genuinely present in the agent source. The three absents (tools, eval, currency) are correctly scored — there are literally no corresponding artifacts.

Understatements and framing problems

Two understatements that do not change scores but do mislead readers:

Understatement 1 — charter notes dismiss structure that exists. The auditor wrote: “Job statement embedded in prompt body; no dedicated charter artifact.” This is technically accurate but creates a false impression. Lines 17–25 of golden-rule.md are a structured Job Description block: Title, Unit, Reports to, Expertise, Purpose, Collaborates with, Comms, Abides by. The block names the Reports-to chain (“Principal / Unit Director”) and names the collaborators (“all units”). This is meaningfully more organized than a job statement “embedded in prompt body.” The score of 1 is still right — measurable success criteria, rung justification, review cadence, and the sign-off gate are all absent — but the note should have distinguished what is present from what is missing. As written, it reads as though nothing structured exists.

Understatement 2 — prompt notes understate the constraints coverage. The auditor found 02:versioned-header as the cited clause for partial. That is correct as the first gap. However, the audit note says “guardrails section not formally structured as <constraints>.” The five-bullet Guardrails section (Never finalize without approval, Preserve foundational documents, Don't over-produce, Surface conflicts, Stay honest) is substantive constraint coverage — it just lacks the XML wrapper. Similarly, the <role> block absence is noted but the structured Job Description block does much of what a <role> block would do. The auditor’s assessment of partial is right, but the note undersells the quality of what is present by framing the gaps as if the content is missing rather than structurally informal.

Neither understatement warrants a score change. They do affect how a training plan should prioritize: the work is primarily structural reformatting, not content authoring.

2. What the Auditor Missed or Understated

Beyond the framing problems above, there are two substantive gaps in the audit’s characterization.

2a. The co-create sequence is load-bearing and unusual

The audit noted “instructions, output-spec, and altitude are strong” but did not call out what makes this agent distinctively well-designed for its scope: the ideate-optimize-approve sequence (Steps 1–6 in the Process section) is a mandatory, enforced workflow. The agent self-describes it as “a Golden Rule, GR-3” and names it a hard constraint under Guardrails. This is an agentic governance pattern — approval gating before any file mutation — that is operationally more sophisticated than what the rubric dimensions directly measure. The audit’s characterization that the agent has “strong” instructions is accurate but treats this as a quality-of-prose observation rather than noting it as the agent’s core design mechanism.

2b. Tier reasoning in the audit is correct but the tier itself was not pre-assigned per pipeline protocol

pipeline.md section “Tiering guide” states: “Assign tier before commissioning the audit. Without a tier, all pre-canon agents produce undifferentiated Non-conformant verdicts with no signal on what to do.” The audit emits T2 in tier_reason, which is the right tier, but the audit appears to have determined tier during the audit rather than receiving it as a pre-assigned input. The pipeline protocol requires it the other way around. This is not an error in the audit findings, but it is a process gap worth flagging: if tier is post-assigned, the audit’s non-conformant verdict is structurally missing calibration context. For golden-rule specifically, the result is still correct. For agents where tier would change which requirements apply, the ordering matters.

2c. No flag on the protected-document guardrail

The agent’s guardrail “Preserve foundational documents. Never alter docs/concept.md or anything under archive/” is a hard constraint with named protected paths. This is relevant from a risk perspective — it means the agent has a self-enforced scope limit on destructive writes. The audit did not note this as a risk-mitigating factor in the tier reasoning. For a T2 characterization (“low risk”), this confirmation of active guardrails is material to the justification.

3. Tier Recommendation

T2 — confirmed.

Evidence:

• Scope: rules drafting only. Output = text blocks destined for CLAUDE.md or CONTRIBUTING.md.
• Tools: Read, Write, Edit, Grep, Glob — file-system only, no external integrations.
• State changes: bounded to local repository files that the user explicitly approves before write.
• No security system access, no governance system writes, no external API calls.
• Active guardrail protecting docs/concept.md and archive/ from mutation.
• Approval gate (GR-3) prevents any write without explicit human sign-off.

This matches the T2 profile in pipeline.md: “Narrow scope, low risk.” Canon requirements at T2 are lightweight: versioned prompt header, one-sentence job, minimal stated tools with justifications, named owner. The audit applied the full five-dimension rubric, which is the correct instrument, but the upskilling recommendations must be calibrated to T2 requirements, not T1.

4. Upskilling Recommendations

Calibrated to T2 (lightweight canon). T1 artifacts — full eval harness (run.py), model.lock, CHANGELOG.md, RUNBOOK.md, CI regression gate — are not required at this tier. Do not impose T1 build costs on a T2 agent.

Recommended order:

P0 — Minimal viable canon compliance for T2

1. Add versioned prompt header to golden-rule.md. Format: v0.1 — <model> — last-eval: none. One line; near-zero effort; satisfies 02:versioned-header and provides the anchor for future CHANGELOG entries. This is the cheapest fix with the highest marginal signal.

2. Add a named owner and review cadence to the Job Description block. “Reports to: Principal” is present but cadence (e.g., “reviewed quarterly by Principal”) is absent. One line addition; satisfies the 01:owner-cadence clause at T2 weight.

3. Add per-tool justification for the five tools listed in the frontmatter (Read, Write, Edit, Grep, Glob). At T2 this is an inline table or annotated list, not a TOOLSPEC.md. Example: “Grep — locate existing rules before drafting; prevents duplicates.” Addresses the tools dimension without imposing T1 toolspec overhead.

P1 — Structural conformance

4. Lift the Guardrails section into a formal <constraints> block (rename or wrap). Content is already present and strong. This is a structural rename, not content work. Satisfies 02:constraints-section.

5. Lift the Job Description block into a formal <role> block (rename or wrap). Same — content is present, structure is informal. Satisfies 02:role-section.

6. Add a lightweight verification artifact: two or three golden examples of golden-rule output (one good rule, one rejection of an untestable idea, one conflict-surface case). Not a run.py; a markdown fixture file is sufficient for T2. This addresses the eval dimension at T2 weight — the floor rule will continue to flag eval=absent until any verifiable test artifact exists, even a minimal one.

P2 — Future cycle

7. Tighten the one-sentence job in the frontmatter description (currently multi-sentence, spans lines 3–13 of the frontmatter block). Distill to one sentence; reserve the expansion for the Job Description section body. Satisfies 01:one-sentence-job exactly.

8. If the agent graduates to multi-file layout (e.g., a dedicated golden-rule/ directory), migrate the Job Description into charter.md at that time. Not warranted now given single-file scope.

Do not do at T2:

• Full run.py eval harness
• model.lock file
• CHANGELOG.md with held-out eval deltas
• RUNBOOK.md
• CI gate

These are T1 requirements. Imposing them on a T2 agent wastes build time and produces ceremony without commensurate risk reduction.

Findings Summary

Pipeline Handoff

{
  "pipeline_handoff": {
    "subject_agent": "golden-rule",
    "cycle": "2026-06-04",
    "recommended_tier": "T2",
    "findings": [
      {
        "ref": "F1",
        "area": "audit-methodology",
        "severity": "Low",
        "finding": "Charter notes dismiss structured Job Description block that is present in lines 17-25; framing misleads training plan readers about how much content work is needed vs. structural reformatting",
        "stream": "tooling"
      },
      {
        "ref": "F2",
        "area": "audit-methodology",
        "severity": "Low",
        "finding": "Prompt notes understate Guardrails quality: constraint content is present and substantive, only the <constraints> XML wrapper is absent — distinction matters for P0 vs P1 prioritization",
        "stream": "tooling"
      },
      {
        "ref": "F3",
        "area": "audit-missed",
        "severity": "Low",
        "finding": "Co-create/approval-gate sequence (GR-3) is the agent's core agentic governance mechanism; characterized only as 'strong instructions' rather than as a named design pattern",
        "stream": "tooling"
      },
      {
        "ref": "F4",
        "area": "pipeline-protocol",
        "severity": "Medium",
        "finding": "Tier post-assigned during audit rather than pre-assigned per pipeline.md protocol; creates calibration risk on future agents where tier changes which requirements apply",
        "stream": "process"
      },
      {
        "ref": "F5",
        "area": "audit-missed",
        "severity": "Low",
        "finding": "Protected-document guardrail (docs/concept.md, archive/) not cited as risk-mitigating factor in tier justification despite being a named hard constraint in the agent source",
        "stream": "tooling"
      },
      {
        "ref": "F6",
        "area": "agent-charter",
        "severity": "Medium",
        "finding": "Review cadence absent — the one T2-applicable charter gap; owner (Principal) is present but cadence is not stated",
        "stream": "agent-development"
      },
      {
        "ref": "F6b",
        "area": "agent-charter",
        "severity": "Low",
        "finding": "Measurable success criteria, rung justification, sign-off gate absent — T1-tier items, not T2 requirements; noted for completeness only",
        "stream": "noise"
      },
      {
        "ref": "F7",
        "area": "agent-prompt",
        "severity": "Medium",
        "finding": "Versioned header absent; role and constraints sections structurally informal (no <role>/<constraints> blocks) despite substantive content being present",
        "stream": "agent-development"
      },
      {
        "ref": "F8",
        "area": "agent-tools",
        "severity": "Medium",
        "finding": "Five tools listed in frontmatter with no per-tool justification; T2 requires annotated list at minimum, not a full TOOLSPEC.md",
        "stream": "agent-development"
      },
      {
        "ref": "F9",
        "area": "agent-eval",
        "severity": "High",
        "finding": "No verification artifact of any kind exists; eval-floor rule blocks conformant/provisional verdict until any testable artifact (even markdown golden examples) is present",
        "stream": "agent-development"
      }
    ],
    "upskilling_items": [
      {
        "id": "TP-1",
        "dimension": "prompt",
        "lever": "Add versioned header to golden-rule.md: 'v0.1 — <model> — last-eval: none'",
        "objective": "Satisfy 02:versioned-header; establish anchor for future changelog entries",
        "priority": "P0",
        "success_measure": "Header line present at top of prompt section; matches pattern vN.N — model — last-eval-score"
      },
      {
        "id": "TP-2",
        "dimension": "charter",
        "lever": "Add named owner and review cadence to Job Description block",
        "objective": "Satisfy 01:owner-cadence at T2 weight; one line addition (e.g., 'Review cadence: quarterly by Principal')",
        "priority": "P0",
        "success_measure": "Job Description block contains explicit cadence field; owner already named (Principal)"
      },
      {
        "id": "TP-3",
        "dimension": "tools",
        "lever": "Add per-tool one-line justification for Read, Write, Edit, Grep, Glob as inline annotated list",
        "objective": "Satisfy 03:distinct-purpose at T2 weight without requiring full TOOLSPEC.md",
        "priority": "P0",
        "success_measure": "Each of the five tools has a one-line purpose statement explaining its role in the rules-drafting workflow"
      },
      {
        "id": "TP-4",
        "dimension": "prompt",
        "lever": "Rename or wrap Guardrails section as <constraints> block; rename Job Description as <role> block",
        "objective": "Satisfy 02:constraints-section and 02:role-section; content is already present, this is structural reformatting only",
        "priority": "P1",
        "success_measure": "Prompt file contains <role> and <constraints> labeled sections with existing content preserved"
      },
      {
        "id": "TP-5",
        "dimension": "eval",
        "lever": "Add a markdown fixture file with 2-3 golden examples: one good rule output, one rejection of an untestable idea, one conflict-surface case",
        "objective": "Provide minimum verifiable artifact to break eval-floor; T2 does not require run.py harness",
        "priority": "P1",
        "success_measure": "At least one golden example exists that a reviewer can check actual agent output against; eval dimension moves from absent to partial"
      },
      {
        "id": "TP-6",
        "dimension": "charter",
        "lever": "Distill frontmatter description to a single sentence; move expansion to Job Description section body",
        "objective": "Satisfy 01:one-sentence-job exactly; current description spans 11 lines across the YAML frontmatter",
        "priority": "P2",
        "success_measure": "Frontmatter description field is a single sentence of 25 words or fewer; full context retained in body"
      }
    ]
  }
}