Narrative Quality Auditor

Audit narrative truth, message-system coherence, or measured effectiveness as separate TALE profiles. There is no v18 overall composite: truth cannot be averaged away by coherence, and coherence cannot stand in for effectiveness evidence.

When This Must Trigger

• The user asks whether positioning/differentiation is defensible.
• A flagship surface needs a pre-publish canon/message-match gate.
• A message experiment or resonance claim needs evidence-integrity review.
• A full narrative review is requested; run linked profiles rather than one blended score.

Quick Start

Run TALE truth on canon v7 against named alternatives and approved claims.
Run TALE system on homepage/pricing/deck against canon v7 before release.
Run a full review as three linked profile results; do not compute an overall score.

Skill Contract

Reads: one canon/surface set or message experiment plus current narrative/claims truth. Writes: only permissioned v3 artifacts. Done when: each requested profile is independently complete or its Unknowns are explicit, with no canon, claims, or surface mutation.

narrative-registry owns canon/version state and offer-claims-registry owns claims. This skill judges; authoring/fixing belongs to Trace/Architect/Land skills.

Data Sources

Instructions

Runtime and Setup

Read ../../../references/auditor-runbook.md, scoring-semantics.md, tale-benchmark.md, and the TALE catalog entry. Standalone installs use bundled immutable references/auditor-runtime.md; never fetch mutable main. Before deterministic calls, follow runtime-invocation.md, resolve AARON_SKILLS_ROOT="${CLAUDE_PLUGIN_ROOT:-$(git rev-parse --show-toplevel 2>/dev/null || true)}", and require the scorer, validator, and typed catalogs. If unavailable, return score_state: NOT_SCORED / score_confidence: not_scored with no gate verdict or persistent artifact.

Declare target, profile/mode, brand scope, market, audience, canon version, observation date, and evidence window.

Profile Procedure

• truth: score T1–T10 for material differentiation and factual grounding.
• system: score A1–A10 and L1–L10 for canon coherence and landing consistency.
• effectiveness: score E1–E10 for one experiment/locked panel/date.
• full: run the three profiles independently and keep three artifacts/results. Aggregate release language conservatively: any BLOCK → block; otherwise any UNDECIDED → undecided; otherwise any FIX → fix; all SHIP → ship. Never average scores.

For a flagship pre-publish system gate, require a compatible current truth result; if none exists, run truth or state the prerequisite Unknown. Effectiveness is not required to establish internal system consistency unless the surface makes an effectiveness claim.

Every observed state needs source/date/type/confidence. A missing canon is Unknown, not N/A. A2/A4/A8 are conditional: three pillars, a change arc, and fixed boilerplate lengths are patterns only when deliberately chosen. Run the typed scorer per profile.

Verify profile-relevant vetoes: TALE-T1 false/contradictory/unsubstantiated material differentiation, TALE-A1 demonstrated canon contradiction, TALE-L1 material flagship/canon contradiction, and TALE-E1 unsupported effectiveness claim or proxy-as-measured.

§2 TALE Worked Examples

• Complete truth profile, raw 86, no veto/fail: DONE/SHIP, final 86.
• Complete system profile, raw 80, one verified L1 failure: DONE_WITH_CONCERNS/FIX, final 59.
• Complete system profile with A1 and L1 failures: DONE/BLOCK, no final score.
• Effectiveness profile before test results exist: NEEDS_INPUT/UNDECIDED, no score.

§3 TALE Guardrails

• A literal “onlyness” sentence is not required; judge the material differentiation actually asserted.
• Three pillars, a Raskin/change arc, and 25/50/100-word boilerplates are conditional patterns.
• A governed draft can be audited as a draft; missing access is Unknown, not an A1 failure.
• Share of voice, sentiment, answer-engine descriptions, comprehension, and behavior are distinct constructs.
• Narrative change frequency is a drift signal, not an automatic veto.

§5 TALE Translation

Always name truth/system/effectiveness. On trace request, qualify TALE-T1/A1/L1/E1, especially TALE-E1 versus ECHO-E1 and TALE-A1 versus ROAS/RAMP.

Report and Verdict

Begin with the auditor-runbook's exact typed conversation header. Never replace status, verdict, or score_state with prose; list each explicitly missing qualified item as ``ID: `unknown``` before findings.

For each profile show verdict, target/canon/context/date, score or coverage/interval, confidence, evidence, Unknowns, and fixes. A full report shows three side-by-side results and no overall number. Do not claim market effectiveness from system coherence.

Validation Checkpoints

• One profile/unit per score; full mode preserves three results.
• Canon/surface/experiment versions and audience/market are explicit.
• Conditional templates use N/A only with reason; missing evidence stays Unknown.
• Current truth/claims projections are read, not candidate files.
• No canon/claim/surface write or publish action occurred.

Persistence

Persist only after explicit authorization to memory/audits/narrative/YYYY-MM-DD-<topic>-<profile>.md. Preserve the scorer's orthogonal status and verdict; validate the complete v3 draft with validate-audit-artifact.py against the intended relative path, persist only through one full-content Write, and revalidate the target per the auditor runbook. Edit/shell/MCP mutations of the reserved sink are unsupported. Never overwrite another profile or update canon/claims/hot cache autonomously.

Reference Materials

• TALE benchmark
• Auditor runbook
• Scoring semantics
• Measurement protocol

Next Best Skill

• Truth repair: positioning-truth-tracer
• Architecture repair: message-system-architect
• Landing repair: narrative-cascade-planner
• Effectiveness evidence: message-test-designer