Moai Domain Humanize

moai-domain-humanize

Post-editing specialist that removes "AI tells" from generated text and rewrites it to read as human-authored, while preserving meaning. This is the editing counterpart to text generation: it does not write new content, it refines how existing content is said. Covers Korean, English, Japanese, and Chinese, across two genre surfaces: prose (columns, reports, blog posts, formal documents) and marketing copy (headlines, CTAs, landing pages, brand storytelling, slide titles). Each language module carries a prose catalogue and a copy-layer catalogue; the shared machinery below (severity model, dual grading, mode-specific guardrails) applies uniformly.

Quick Reference

Operating Principles (4)

1. Meaning preservation is the top rule. Facts, numbers, statistics, named entities, quotations, citations, and the author's stance/certainty stay intact. Any meaning drift forces a rollback. In copy mode, "meaning" is defined by the fact anchors plus the core promise/benefit — see the copy-mode guard below.
2. Evidence-based edits only. Every change must trace to a detected tell on a specific span. Stylistic "improvements" unconnected to a catalogued tell are themselves an over-editing signal and are forbidden.
3. Genre and register preservation. Humanize within the source register — academic stays academic, casual stays casual. Never push formal text into slang or vice versa. Copy and slide genres apply their own structural rules (noun-phrase title boundaries, appeal-vs-informational voice) defined in each module's copy layer.
4. Over-editing prevention. In prose mode, flag at >30% change (WARN) and halt at >50% change (forced stop / human review) — above 50% you are regenerating, not humanizing. In copy mode, the change-rate guard is REPLACED by the fact-anchor preservation guard (see Over-Editing Guardrails below).

Genre Mode Selection (Prose vs Copy)

Two operating genres select which guardrail and grading table apply. Default from the text's genre; an explicit user instruction overrides.

Processing Mode Selection (Fast / Strict)

• Fast mode (default, up to ~5,000 chars): a single pass — detect, rewrite, self-verify against the meaning-preservation checklist.
• Strict mode (long or high-stakes text, or when requested): separate stages — detect → surgical rewrite → content-fidelity audit (facts/figures/stance unchanged) → naturalness review. Re-run a second pass when the result lands at Grade C.

Output Contract

Return two things:

1. The humanized text.
2. A short change report: categories hit (with counts), the final quality grade (A/B/C/D), and — in prose mode — the estimated percent changed. When a guardrail fires, state it explicitly (prose mode: WARN at >30%, HALT at >50%; copy mode: any fact-anchor loss).

Common Severity Model (shared by all 4 languages)

Each tell carries one severity tier. Detectors gate by occurrence count and overlap, because a single tell rarely proves AI authorship — confidence comes from clustering.

Common Quality Grades (shared by all 4 languages — dual tables)

Graded after the rewrite. The genre mode selects the table: prose mode grades on residual tells plus change rate; copy mode grades on residual S1 plus fact-anchor integrity, with NO change-rate band.

Prose-Mode Grade Table

Residual S1/S2 counts plus improvement % (= proportion of detected tells removed without introducing new ones).

Copy-Mode Grade Table

Residual S1 (including the module's copy-layer S1 tells), fact-anchor integrity, and self-verification — no change-rate band, because a legitimate headline rewrite routinely changes most of its characters while preserving every anchor.

Hard rule (both modes): any residual S1 caps the grade at C; any meaning-distortion flag forces D. S3 tells affect the grade only when ≥3 of them overlap and reinforce an S1/S2 finding.

Over-Editing Guardrails (shared)

Prose mode — change-rate guard. Change rate = the proportion of the text altered; target band ~5–30%.

• >30% changed → WARN. Surface a caution and cap at Grade C until each edit is justified by a detected tell. Note: padding-removal legitimately shrinks text, so a length drop alone is not a violation — flag when meaning-bearing spans are altered.
• >50% changed → HALT. Stop and require human confirmation; revert to the last safe state.
• Conservative judgment near the thresholds. This skill carries no quantitative measurement layer, so the change rate is an LLM estimate, not a reproducible metric. Treat a borderline estimate as OVER the threshold: near ~30%, issue the WARN; near ~50%, HALT. Bias toward caution so an over-edit never slips through on an optimistic estimate. (Known limitation: without a computed metric, before/after improvement percentages are estimates as well — report them as such.)

Copy mode — fact-anchor preservation guard (REPLACES the change-rate guard). In copy mode, meaning invariance is anchored differently: numbers, dates, prices, proper nouns, and legal notation are preserved 100% character-intact, AND the core promise/benefit of the copy keeps its meaning — while expression and sentence structure MAY be rewritten freely. The change-rate guard does not apply, because copy humanization legitimately rewrites most of a headline; the guard that replaces it is absolute on anchors:

• Any altered number, date, price, proper noun, or legal notation → rollback of that edit.
• Core promise/benefit drift → rollback. The rewritten copy must promise the same thing to the same audience.
• No invented specifics. Replacing vague copy with concrete claims is only allowed when the concrete facts exist in the source or are supplied by the author.

Meaning-Preservation Checklist (shared, all must hold)

1. Anchor facts first — fix the claims, numbers, names, dates, and certainty level before editing.
2. Edit at sentence/phrase level, not whole-document regeneration.
3. Add no new facts — never invent specifics to replace vagueness; simplify instead, or flag for the author.
4. Drop no load-bearing facts — removing an inflated wrapper must keep the substantive claim inside.
5. Preserve genuine certainty/hedging and technical terminology verbatim.
6. Final diff check — compare facts, tone, certainty, and examples against the original; revert any edit that drifts.

Invariant Ledger and Delta Audit

Two techniques harden the meaning-preservation machinery above: the Invariant Ledger makes the boundary explicit before editing, and the Delta Audit makes the survival check systematic after editing. They thread into the workflow (steps 2 and 6) rather than replacing any step, and they reinforce — never replace — the severity model, grades, guardrails, and the meaning-preservation checklist.

Invariant Ledger (pre-edit boundary)

Before editing, record an Invariant Ledger — the explicit list of what MUST survive the humanization pass unchanged. This is the written, checkable form of "Anchor facts first" (checklist item 1). Capture every item across the four categories:

1. facts (with evidence boundaries)
2. identifiers (commands, paths, URLs, status values, error codes, product names)
3. conditions / numbers / dates / versions / units / comparisons
4. exceptions / limitations / risks / uncertainty / approvals / rollback / next-actions

Fidelity rule. Never silently add, remove, narrow, broaden, strengthen, or weaken a ledger item. The wording is free to change; the commitment the text makes is not.

Mark each item supplied or inferred. A supplied item is something the source text actually asserts — it is hard-anchored, and any drift on it triggers a rollback (see Delta Audit). An inferred item is an adjacent benefit or guarantee the source never stated but a reader might assume — it is recorded for reviewer awareness only, and dropping it during humanization is NOT a rollback trigger, because the original never promised it. When an item is left unmarked, treat it as supplied: the fail-safe direction is preservation.

Depth by processing mode.

• Fast mode: a lightweight inline anchor list that still covers every one of the four categories above. Fast is shorter in FORM, never narrower in CATEGORY COVERAGE — even a short text gets a line for each category that applies.
• Strict mode: an explicit written boundary document with each item enumerated and marked supplied/inferred.

Delta Audit (post-edit verification)

After the edit pass completes, run a Delta Audit — compare the output against the Invariant Ledger before grading. This is the systematic form of "Final diff check" (checklist item 6), across three axes:

1. Claim & intent parity — every claim the source makes, the output still makes, at the same strength and with the same intent.
2. Survival check — every ledger identifier, number, condition, limitation, and risk is still present and unchanged.
3. Audience / tone / purpose fit — the output still addresses the same reader, register, and goal.

Also flag any ambiguity the edit newly introduced: an unresolved actor, unclear ownership or handoff, or a softened destructive-effect or approval/rollback caveat.

Rollback on a supplied-item violation. When the audit finds any supplied ledger item added, removed, narrowed, broadened, strengthened, or weakened, roll back that edit — the same meaning-drift rollback the Operating Principles already require. Removal of an item marked inferred is reported in the audit output but does NOT trigger a rollback.

Feeds grading. A ledger violation is a meaning-distortion flag, and a meaning-distortion flag forces Grade D in both modes per the existing hard rule (see Common Quality Grades) — the Delta Audit is the mechanism that detects it.

Language Routing

Each target language has its own tell catalogue (categories, before/after examples in the target language, per-category severity). Load the module that matches the text being edited:

The Korean module is an original catalogue; the English, Japanese, and Chinese modules are independently web-researched catalogues built on the same architecture. Each module's copy layer is language-native — copy tells do NOT transfer mechanically between languages (English headlines are natively terse; Japanese 体言止め is prestigious craft gated by frequency, not presence; Chinese 对偶/排比 is judged content-first, not by count) — so never apply one language's copy rules to another. The common severity model and quality grades above apply uniformly to every module — the modules add only the language-specific tell categories, severities, and example rewrites.

For mixed-language text, detect the dominant language and route to its module; apply each module independently to its spans when the text is genuinely multilingual.

Genre-Module Routing

Two genre modules stack ON TOP of the language routing above — they never replace the language module:

When both conditions hold (a QA-gate review of display-surface copy), load both genre modules alongside the language module. The Language Routing table above remains the language axis and is unchanged by this extension.

Implementation Guide

Workflow (per text)

1. Identify language, genre mode, and processing mode. Pick the module by dominant language; pick prose vs copy mode by genre (see Genre Mode Selection); pick Fast vs Strict by length / stakes.
2. Anchor facts — build the Invariant Ledger. Record the four-category ledger (facts; identifiers; conditions/numbers/dates/versions/units/comparisons; exceptions/limitations/risks/uncertainty/approvals/rollback/next-actions) that must survive unchanged, marking each item supplied or inferred (see Invariant Ledger and Delta Audit above). Fast mode uses a lightweight inline anchor list covering every category; Strict mode uses an explicit written boundary document. In copy mode these anchors are the guard itself.
3. Detect tells. Scan against the module's catalogue — the prose categories in prose mode, plus the module's Copy Layer categories in copy mode. Record each hit with its category ID, span, and severity. Count occurrences (S2/S3 gate on repetition).
4. Rewrite surgically. Edit only flagged spans. Replace each tell with a natural rendering in the same register. Do not touch unflagged text.
5. Apply the mode's guardrail. Prose mode: estimate the change rate and apply the change-rate guard (WARN >30%, HALT >50%, conservative near the thresholds). Copy mode: verify every fact anchor and the core promise/benefit instead.
6. Self-verify (Fast) or audit + review (Strict) — run the Delta Audit. Compare the output against the Invariant Ledger for claim & intent parity, identifier/number/condition/limitation/risk survival, and audience/tone/purpose fit (see Invariant Ledger and Delta Audit above). A supplied-item violation rolls back that edit and raises a meaning-distortion flag (forcing Grade D); an inferred-item removal is reported but not rolled back. Re-run the meaning-preservation checklist. In Strict mode, run the content-fidelity audit and naturalness review as separate stages.
7. Grade. Count residual S1/S2 and improvement %; assign A/B/C/D. Second pass on C; human review on D.
8. Emit the humanized text + change report.

Detection note (shared across languages)

Automated AI-text detectors are unreliable across these four languages (notably weak on CJK polite registers, where they false-positive on correct formal writing). This skill is a pattern-based editing tool, not a detection oracle: rely on the catalogued tell categories and the clustering-based severity gates, not on a detector's verdict.

Common pitfalls

• Re-injecting AI-ness. Rewriting AI text with a fresh full regeneration tends to add new tells. Favor surgical edits to flagged spans over wholesale rewriting.
• Fabricating specifics. When a module calls for concrete detail to replace vague filler and no real specifics exist, simplify or flag for the author — never invent.
• Style drift. "Cleaning up" beyond the flagged tells violates principle 2 and inflates the change rate toward the HALT threshold.
• Mixing registers mid-document. Keep one consistent register (and, for Japanese/Chinese, one consistent politeness/sentence-ending style) across the whole output.

Works Well With

• sync-auditor: independent skeptical review. Use it to score whether the humanized output preserved meaning against the original and met the target grade.

Category-catalogue structure inspired by the im-not-ai (Humanize KR) project.

Version: 1.3.0