Blog

Quality Auditor: no faulty chapter moves forward

Quality Auditor NovelKit is an automated post-review layer — continuity, OOC, timeline — before sync commit allows the next chapter.

In the AI novel writing pipeline, the Quality Auditor (Khí Linh review) is an agent you cannot skip. Unlike spell-check, this is multi-dimensional logic review with thresholds PASS 85+, SOFT 70+, and HARD-FAIL blocking the pipeline — running on Hermes before every sync.

What does the Quality Auditor check?

  • Continuity — timeline, location, events that have already happened
  • OOC — behavior deviating from the Character Profile / Want-Need-Lie
  • Plot thread — open threads left unresolved or resolved illogically
  • Cultivation / system — realm, stat, cheat flags (genre-specific)
  • Prose style — terminology, dialogue voice, style guard

Scores and verdicts

PASS ≥ 85: sync is allowed. SOFT-FAIL 70–84: fix in the sandbox, then re-review. HARD-FAIL: blocks the following chapter; circuit breaker after repeated failures. Tool: the unified novelkit_gate registry on Hermes.

Before sync commit

Only a passing chapter enters Memory.md, RAG reindex, and episodic commit. doctor checks coherence after sync. Details: Quality gate, Pipeline.

Read more: Memory layer · Canon file-first · The system

Scoring rubric and genre extensions

The overall score aggregates multiple dimensions: continuity, character, plot, style, genre-specific (cultivation, system interaction, and more). Meta Genre has a Pitfall Watchlist: SYSTEM_CHEAT, POWER_INFLATION, EMOTION_DEFICIT. Cultivation stories have a breakthrough process check.

Human-in-the-loop when needed

A SOFT-FAIL can be fixed by hand in the sandbox before re-review. Publishers can require human sign-off on a pilot before auto-scaling. The gate is still the condition for sync — you cannot bypass it with chat.

Review artifact output for Publishers

Every synced chapter ships with a review file: score, verdict, flags, suggested fix. Publishers assessing it do not need to reread the entire thing — they look at the score distribution and the HARD-FAIL rate of the pilot.

Quality Auditor NovelKit: scoring rubric and genre extensions

The Quality Auditor NovelKit (Khí Linh review via novelkit_gate) is the quality gate before sync — not "AI grading itself for fun." The score aggregates continuity, character, plot, style, plus genre-specific: cultivation process (cultivation), system interaction (system), romance chemistry (romance), hard-SF consistency (sci-fi).

Thresholds: ≥85 PASS — sync allowed; 70–84 SOFT-FAIL — fix in the sandbox per the review note, then re-run; <70 HARD-FAIL. An explicit verdict (PASSED/SOFT-FAIL/HARD-FAIL) always beats an ambiguous score band. Meta Genre Pitfall Watchlist: SYSTEM_CHEAT, POWER_INFLATION, EMOTION_DEFICIT, CANON_CONFLICT.

Review artifact output for Publishers and provenance

Every synced chapter ships with a JSON + MD review file: score, verdict, flags, suggested fix, DNA Depth Audit. Publishers assess a pilot through the score distribution and HARD-FAIL rate — no need to reread the whole manuscript. Provenance records who reviewed, the timestamp, and a link to the chapter version.

Human-in-the-loop when needed

SOFT-FAIL: the Prose Writer fixes it in the sandbox → re-review → reaches 85+ before sync. Publishers can require human sign-off on a pilot before auto-scaling. Do not bypass the gate by copy-pasting outside the system — it loses the audit trail and pollutes memory.

Quality Auditor within the Hermes circuit breaker

record_result updates the breaker: hard×2, soft×3, total×5 consecutive failures → pipeline blocked; Lãng Khách / human resets after the root cause. This avoids burning tokens writing chapter 47 on top of canon broken by a HARD-FAIL in chapter 46.

Detailed rubric and genre extensions

The score aggregates continuity, character, plot, style — plus genre-specific extensions: cultivation process (cultivation), system interaction (system), romance chemistry (romance), hard-SF consistency (sci-fi). The explicit verdict PASSED/SOFT-FAIL/HARD-FAIL always beats an ambiguous score band. Meta Genre Pitfall Watchlist: SYSTEM_CHEAT, POWER_INFLATION, EMOTION_DEFICIT, CANON_CONFLICT.

Publishers read the review artifact instead of the full manuscript

Every synced chapter ships with a JSON + MD review: score, verdict, flags, suggested fix, DNA Depth Audit. Publishers assess a pilot through the score distribution and HARD-FAIL rate — no need to reread 50 chapters. Provenance records who reviewed, the timestamp, and a link to the chapter version — a copyright dossier when needed.

VerdictScorePipeline action
PASS≥85novelkit_sync commit allowed
SOFT-FAIL70–84Fix in sandbox → re-review
HARD-FAIL<70DAG blocked, the next chapter does not start

SOFT-FAIL workflow: the Prose Writer fixes per the review note → the Quality Auditor re-runs via novelkit_gate → reaches 85+ before sync. Do not bypass the gate by copy-pasting outside the system — it loses the audit trail and pollutes the 5-layer novelkit_memory.

Cluster: quality gate, pipeline, canon, long-form fiction, copyright.

Quality Auditor FAQ

Can the Prose Writer grade itself?
No — Khí Linh reviews separately via novelkit_gate, hub-and-spoke under Lãng Khách.
Can a score of 84 be published?
No sync — a SOFT-FAIL is fixed in the sandbox, then re-reviewed to reach ≥85.
Publisher sign-off on the pilot?
You can require a human before auto-scaling — the artifact is still mandatory.

The review artifact records score, verdict, flags, DNA Depth Audit — a B2B and copyright dossier. record_result updates the breaker; this avoids writing chapter 47 on top of canon from a HARD-FAIL in chapter 46.

The Quality Auditor is Khí Linh review, separated from the Prose Writer via Lãng Khách delegate_tool — hub-and-spoke, no self-grading. Genre extensions: cultivation, romance, system, hard-SF in the novelkit_gate rubric.

Quality Auditor NovelKit is the provenance gate before sync — Publishers read the artifact instead of the full manuscript.

PASS ≥85, SOFT 70–84, HARD <70 — an explicit verdict in the review JSON of every synced chapter.

The Meta Genre Pitfall Watchlist catches SYSTEM_CHEAT, POWER_INFLATION, CANON_CONFLICT on Hermes.

Genre-specific checks

Cultivation: realm, fuel, dao heart. Romance: chemistry, conflict pacing. Isekai: world rule, cheat cost. System: stat consistency, quest logic. Each genre squad injects a checklist into the review bundle — not one generic list for every title.

Review artifact for legal / licensing

The review file stores verdict, score, flags — proof of the QA process before publishing. Useful when registering copyright or in a dispute over the origin of AI content with human oversight.

Quality gate, copyright.