Blog

Quality gate: never sync a chapter with broken logic

Quality gate in AI novel writing: PASS 85+, SOFT-FAIL to fix, HARD-FAIL to block — NovelKit's sync threshold on Hermes.

The AI writing quality gate is the layer that protects Publishers from a chapter with broken logic becoming the foundation for the next 50 chapters. NovelKit locks the thresholds PASS=85, SOFT=70 in novelkit_gate and novelkit_sync — not an optional checkbox.

The three verdict levels

  • PASS (≥85) — sync commit, the next chapter gets planned.
  • SOFT-FAIL (70–84) — fix in the sandbox, review again.
  • HARD-FAIL (<70 or a blocking error) — block the pipeline; nothing written to memory.

What does the Quality Auditor check?

Continuity, OOC, timeline, plot thread, vocabulary guard, genre-specific (cultivation, system cheat…). Agent details: Quality Auditor NovelKit.

Circuit breaker and doctor

After 2/3/5 consecutive fails (breaker config), the pipeline pauses. doctor reports canon/index/memory incoherent before continuing. Read the pipeline.

Read more: Memory layer · Writing guide · Studio

SOFT-FAIL workflow in practice

Chapters scoring 72–84 typically lack: foreshadowing, vocabulary guard, a minor timeline slip. Prose Writer fixes it in the sandbox following the review note → Quality Auditor re-runs → only after reaching 85+ does it sync. No "manual fixes outside the system" followed by a sync — that loses the audit trail.

HARD-FAIL and the circuit breaker

CANON_CONFLICT, breakthrough without fuel, SYSTEM_CHEAT — block the pipeline. The breaker opens after N consecutive fails; Lãng Khách / a human must reset it after fixing the root cause. This avoids burning tokens writing on top of already-broken canon.

The PASS 85 threshold — why not 70?

70 is SOFT — a fix is accepted. 85 is the publish bar for a long serial: enough continuity that later chapters do not collapse. Publishers can negotiate their own pilot threshold in a contract, but by default the system protects the catalog.

AI writing quality gate: PASS 85, SOFT 70, breaker 2/3/5

The AI writing quality gate on NovelKit is not a marketing checkbox — it is the novelkit_gate registry on Hermes that blocks the sync commit when a chapter has not met the bar. ≥85 PASS: mutate canon, episodic memory, RAG index. 70–84 SOFT-FAIL: Prose Writer fixes per the review note in the sandbox → Quality Auditor re-runs. <70 HARD-FAIL: block the next chapter (DAG); CANON_CONFLICT, SYSTEM_CHEAT, breakthrough without fuel are usually HARD even with a borderline score.

Why 85 and not 70 for serial publishing? 70 is the bar that can be fixed; 85 is the bar for a chapter to be a safe foundation for chapter 51. Publishers negotiate the pilot threshold in a contract — by default the system protects the catalog.

SOFT-FAIL workflow in the field

Chapters scoring 72–84 typically lack: foreshadowing, vocabulary guard (Sino-Vietnamese xianxia), a minor timeline slip, weak DNA Depth on one field. Workflow: read the review flags → fix in the sandbox → re-review → only PASS then sync. No "fixing it in Word outside" then syncing — that loses provenance, and memory no longer matches the review.

Hermes circuit breaker

  • hard×2 — two consecutive HARD-FAILs → breaker opens
  • soft×3 — three consecutive SOFT-FAILs → breaker opens
  • total×5 — any five consecutive fails → breaker opens

record_result updates state; Lãng Khách / a human resets after fixing the root cause; resume continues the task — no duplicate sync (P12).

The gate within the 9-step pipeline

Write stage output → review stage (novelkit_gate) → only a PASS enters the sync stage (novelkit_sync commit) → doctor. Parallel bootstrap is fine; write/review/sync run sequentially per chapter per novel. A missing tool registry → bootstrap.verify_registry() fails — the surface will not start.

Publishers negotiating thresholds and provenance

By default PASS ≥85 protects the catalog — a chapter scoring 70 as the foundation for chapter 51 is a continuity risk. Publishers can negotiate a pilot threshold in a B2B contract; the system still logs the full review artifact. Provenance: who reviewed, score, verdict, flags, timestamp — a transparent dossier before publishing to a platform.

Genre-specific HARD-FAIL triggers

  • Xianxia: POWER_INFLATION, breakthrough without fuel, missing cultivation texture
  • Isekai/system: SYSTEM_CHEAT with no cost/cooldown
  • Romance: resetting the relationship graph with no canon event
  • Every genre: CANON_CONFLICT with PROJECT_DNA or database/*
Breaker ruleThresholdReset
hard×22 consecutive HARD-FAILsLãng Khách / human after root cause
soft×33 consecutive SOFT-FAILsFix the recurring error pattern
total×55 any consecutive failsDoctor + resume, no duplicate sync

Hermes cron: style_audit every 10 chapters catches drift before it accumulates into SOFT-FAILs; rolling_seed after sync keeps generation continuity. Lãng Khách orchestrates via delegate_tool in hub-and-spoke — the Quality Auditor is a separate reviewer, not the Prose Writer grading itself.

Read more: Quality Auditor, the 9-step pipeline, canon, the guide, copyright.

Quality gate FAQ

Fix it in Word outside then sync?
No — that loses provenance, and memory no longer matches the review artifact.
What do I do when the breaker opens?
Fix the root cause → Lãng Khách/human resets → resume, no duplicate sync.
Publish a 70-point chapter to a platform?
That is the bar that can be fixed — safe serial publishing needs PASS ≥85 as the foundation for later chapters.

Lãng Khách does not bypass the gate — the Prose Writer and Quality Auditor are separated hub-and-spoke via delegate_tool. The DAG blocks chapter N+1 while chapter N has not synced a PASS; bootstrap.verify_registry() protects the tool chain.

The AI writing quality gate is the novelkit_gate registry on Hermes — PASS 85 publish-ready, SOFT 70 sandbox fix, breaker 2/3/5 protecting a 100+ chapter serial catalog.

The gate blocks the sync commit when the bar is not met — protecting memory and the RAG index.

SOFT-FAIL workflow in practice

Chapters scoring 70–84 are not discarded — Prose Writer fixes them in the sandbox following the review flags, and Quality Auditor runs again. Publishers can set a pilot threshold higher than 85 for premium titles. HARD-FAILs usually come from a continuity break or a cultivation violation — you must fix the outline beat, not just polish the sentences.

Mapping the gate to retention

A chapter that passes the gate = lower risk of readers dropping off over a plot hole. Platforms measure retention by cohort — NovelKit's catalog aims to reduce "a broken chapter ruining the arc" compared with pure chat. The gate is insurance, not a replacement for a final human editor if the contract requires one.

Related: Auditor, pipeline, the guide.

Negotiating thresholds in a B2B contract

By default PASS 85 protects the catalog. A premium pilot can be 88+. A mass-market experiment can be 82 with 100% human spot-checking. Document it in the SLA — do not change the threshold silently between batches.

Breaker 2/3/5 explained

2 fails → warning. 3 → pause, suggest doctor. 5 → breaker opens, needs a human to unblock. This avoids burning tokens on a broken outline or DNA conflict.

Auditor, Pipeline.