Velith / rules

---
description: Velith editing pipeline (7 stages), the 2026 AI-tell taxonomy, cold-read protocol, and style conventions
globs:
  - "**/edits/**"
  - "**/STYLE.md"
  - "**/bible.md"
alwaysApply: false
---

# Velith — Editing Pipeline and Style Conventions

Full reference: `skills/loom/quality-bar.md`. Editing means rewriting the manuscript in place; every stage snapshots `drafts/` first (`velith.mjs snapshot`) and files a report of what changed.

## Stages

| # | Stage | Rewrites | Report |
|---|-------|----------|--------|
| 0 | Fact check (nonfiction, technical, academic; fiction with real-world claims) | Yes: unverifiable claims removed | `edits/00-fact-check.md`, `edits/claim-ledger.md` |
| 1 | Editorial assessment: promise delivery, pacing map, redundancies, voice consistency, five-axis score, intervention plan | No | `edits/01-assessment.md` |
| 2 | Developmental: restructure (author approval), chapter rewrites via chapter-writer, deepen; then continuity-editor | Yes | `edits/02-developmental.md`, `edits/continuity-report.md` |
| 3 | Line edit: style-doctor measures + rewrites; then meaning-level pass | Yes | `edits/03-line-edit.md`, `edits/style-report.md` |
| 4 | Copy edit: names, terms, numbers, punctuation, 띄어쓰기, headings, citations | Yes | `edits/04-copy-edit.md` |
| 5 | Proofread; set chapters `status: final` | Yes | `edits/05-proofread.md` |
| 6 | Readiness: beta-reader cold read; PASS or REVISE with fix list (max two loops, then author) | No | `edits/06-readiness-report.md` → `edits/readiness-report.md`; on PASS `edits/editorial-report.md` |

## Five-axis rubric (1-10)

Voice and prose · Structure and pacing · Depth (character / argument) · Specificity and grounding · Reader experience. Readiness PASS: every axis ≥ 7, mean ≥ 7.5, no put-down point in chapters 1-3.

## The tells that matter in 2026

Structural: uniform paragraph length; punch-line paragraph endings; tricolons; symmetrical scenes; "not X but Y"; balanced antithesis; reflective codas; em-dash cascades; rhetorical question chains; summary-then-scene; naming the theme.

Emotional (fiction, memoir): precise self-knowledge; physical-reaction inventories; universal kindness; dialogue that answers; "something shifted"; tidy resolution; interchangeable voices; sensory garnish.

Nonfiction/technical: hedged authority without a source; invented specificity; framework inflation; listicle drift; preview-and-recap; uniform confidence; analogy padding; code that never fails.

Korean: 번역투 무생물 주어; ~것이다 남발; 피동 과잉; 감정 명명; 접속부사 연쇄; ~에 대해/~을 통해; 대명사·~들 과잉; 존댓말·시제 흔들림; 균일한 문장 길이; 한자어 과밀; 설명형 대화.

English lexical: delve, tapestry, testament to, navigate/landscape (metaphorical), nuanced, multifaceted, plethora, myriad, robust, seamless, leverage, unpack, at its core, it's worth noting, a beacon of, resonate, embark, foster, underscore, pivotal, game-changer, transformative, holistic, synergy, paradigm, elevate, vibrant, bustling.

## Cold-read protocol

Read once at reading speed without fixing; mark where attention dropped or the machine showed. Diagnose each mark with a quoted line and a named tell. Score the five axes with one honest sentence each. Prioritize the three fixes that most raise the lowest axis. Only then revise. A critique with no quoted lines is not a critique.

## Metrics

`node velith.mjs metrics <drafts-dir>` reports per chapter: sentence length mean/sd/cv, mid-band share, paragraph length distribution, punch-ending share, type-token ratio, em-dash density, AI-tell hits per 1k (en/ko), "not X but Y" count, rhetorical question density, connectives, dialogue share, and cross-chapter repeated 3/4-grams. Flags: cv < 0.4; mid-band > 60%; punch endings > 35%; tells > 2/1k; em-dashes > 3/1k; any 4-gram ×3+; TTR > 15% below median.

## Severity

Critical: contradiction of fact/character/timeline, fabricated claim or citation, failing code, put-down point in chapters 1-3. Major: repeated image or scene shape, term defined twice, structural gap, voice drift. Minor: spelling, formatting, single tell instance.

## Tone rules

All voice checks reference `STYLE.md`'s voice fingerprint and voice-lock passages. The author's deliberate constructions override the taxonomy. Never sand a voice to neutral.