de-ai

name	de-ai
description	唯一的去 AI 技能。对一个目标文件串联三过：(1) 内部 de-AI 子系统测量（L0 词汇 + L1-L4 结构/学习信号）、(2) 两个 vendor 仓库的 humanizer 结构 tell 审计（AIScientists-Dev/academic-humanizer Layers 1-5 + blader/humanizer patterns 2.12-2.16 与 Pass-2 自审）、(3) claim-first 忠实改写（保真 eligibility + §5.3 长度硬门）。保留每个数字、公式与引用。Use when 用户说 "太 AI 了" / "去 AI 感" / "像人写的" / "de-AI" / "rewrite in voice" / "机器味太重"，或被 paper-review / final-review 作为结构 tell 审计步调用（--audit-only）。
disable-model-invocation	false
argument-hint	<file_path> [--section <name>] [--n N] [--max-iter N] [--audit-only] [--no-apply] [--distill] [--field <name>]

Normative authority: docs/SCIPAPER_STANDARD.md. Where this file and the standard disagree, the standard wins. docs/DEAI_SUBSYSTEM.md explains the design; docs/EVALUATION.md records the empirical evidence. Style profiles and learned models supply evidence and positive anchors, never competing policy.

Provenance (MIT, attribution retained): this skill chains three de-AI sources into one pipeline — the in-house de-AI subsystem (the L0-L4 tools and the standard's feedback contract) and two vendored humanizers: AIScientists-Dev/academic-humanizer v0.3.3 (MIT; the Layer 1-5 catalog) and blader/humanizer (MIT; the structural families 2.12-2.16, the Pass-2 self-interrogation, and the false-positive guards). The upstream word lists are not normative here — corpus evidence is (§7 Corpus overrides).

de-ai — one skill, three passes

0. What this skill does

Three sequential passes over one target, all inside the standard's measure -> type -> rank -> edit -> re-measure -> disposition loop:

1. Pass 1 — measure (de-AI subsystem). The unified linter and the document-scale detectors produce a machine-readable sci-paper.feedback.v1 report: L0 lexical/punctuation, L1 distribution, L2 sentence + document structure, L3 learned similarity, L4 positive voice.
2. Pass 2 — humanizer audit (two vendored repos). The structural-tell catalog (Layers 1-5 plus families 2.12-2.16) runs in audit form; its findings merge into Pass 1's ranked list.
3. Pass 3 — claim-first rewrite. Each selected paragraph is rebuilt from its protected claim graph; only fidelity-eligible candidates within the §5.3 length budget survive; re-measure, then apply with author-visible before/after.

--audit-only stops after Pass 2 (findings only; this is the mode paper-review dimension D and final-review invoke). --no-apply runs Pass 3 to proposals only. For technical prose, neutral and precise IS the human voice — this skill strips machine tells; it never adds personality, humor, or first-person flavor.

Boundary with /sci-paper:condense (canonical statement, mirrored there): de-ai removes the authorship fingerprint (L0-L4 signals, structural tells, voice; its Pass-3 length cap only guards its own rewrites against growth). condense removes redundancy and length (cross-document deduplication under one-canonical-home-per-fact). At the overlap (verbose AI-isms such as rule-of-three padding or connective stacking), de-ai DETECTS the tell; when the right fix is deletion rather than rewrite, the deletion EXECUTES under condense's ranked sweep.

0.1 Non-negotiable rules

1. Scientific fidelity is eligibility, not a score. A candidate that drops or alters a number, unit, citation, mathematical expression, acronym, comparison direction, negation, causal direction, named entity, scope, stance, or qualifier is ineligible regardless of its style score.
2. Never add specificity from imagination. Every number, entity, citation, and quantitative qualifier must already exist in a source read in the same turn. Specificity means retaining supported detail, not inventing it.
3. Rebuild from claims, not surface swaps. Synonym substitution, connector deletion, and punctuation cleanup are not a structural rewrite. Extract the claim graph and compose new sentences.
4. Imitate style, never content. Exemplars guide rhythm, register, information distribution, and transition practice; claims, citations, data, and distinctive wording are never copied from them.
5. Learned scores are advisory field-similarity signals. They do not establish authorship and are not paper gates. Missing calibration is degraded or unmeasured, never a clean result. No universal style verdict exists.
6. L0 is the only axis rewritten to zero. Tier A, em-dash, and Tier B above the section cap are l0_target. L1-L4 findings are ranked advisories with dispositions, never forced to zero at the cost of clarity.
7. Condense, do not accumulate (§5.3). The default direction of every rewrite is shorter; the explanatory patch — answering a finding by appending a clause instead of rewriting the flagged text — is the canonical violation. Forward narrative: never narrate the drafting history.
8. Minimum effective edit + explicit human control. Rewrite only the paragraphs whose findings or the author's instruction justify it; show before/after, fidelity evidence, and residual dispositions.

1. Field resolution and calibration

1. Resolve --field: one available field under style-profile/ selects automatically; multiple require --field; zero means corpus guidance is unavailable — model-free axes still run and the gap is reported.
2. Read docs/SCIPAPER_STANDARD.md and the /sci-paper:paper writing guidance (Anti-AI-isms tiers, forward narrative, formula and citation standards, canonical L0 examples).
3. Read style-profile/<field>/style_dossier.md in full when it exists, and check freshness: if the corpus under style-corpus/<field>/ is newer than the dossier, stop and regenerate with python tools/extract_style.py. A stale profile is unmeasured, never treated as conformity.
4. Retrieve section-typed positive anchors only when the bank exists: python tools/retrieve_exemplars.py --field <field> --section <type> --topic "<verified one-sentence topic>" --k 5. Read every returned exemplar; record provenance; use for rhythm and register only.

2. Pass 1 — measure (de-AI subsystem)

python tools/ai_ism_lint.py <file> --field <field> \
  --structure --distribution --document-structure --oracle --voice \
  --summary --format json --output <scratch>/feedback-before.json

Read the output as sci-paper.feedback.v1; record every axis state (measured / degraded / unmeasured / not_applicable) — silence is never clean. Exit semantics: 0 = no L0 targets (advisories may remain), 1 = L0 targets present, 2 = input/config/execution failure. Never re-parse JSON from printed prose.

Axes and tools:

• L0 lexical + punctuation (ai_ism_lint Tier A/B + em-dash) — the only to-zero axis. The word lists load from style-profile/<field>/lexicon.json; do not re-derive them by hand.
• L1 distribution (deai_metrics, deai_oracle) — burstiness, connective openers, surprisal/UID. Degraded without a field operating point; rank-only.
• L2 sentence structure (deai_structure) — template families (announced enumeration, ordinal runs, setup-list-wrap, repeated modal/anaphoric frames, symmetric closers) plus the auxiliary families (antithesis clusters, short reversal beats).
• L2 document structure (deai_docstructure) — the de-AI centre of gravity: per-stratum cross-paragraph dispersion manifold, role coupling, split-conformal operating points. Word-level rewriting cannot move this signal; the only sanctioned lever is deai_partition (§4).
• L2 claim anchoring (deai_anchoring) — a writing-quality band, not an AI-discrimination axis.
• L3 learned voice (deai_voice) — a decided-degraded offline audit instrument; per-paragraph findings are confidence-capped at 0.5. A low score never proves machine authorship.
• L4 cooperative repair (deai_partition, deai_provenance, deai_personal) — provenance/personal are honestly unmeasured until the author supplies their own draft history or prior papers.

3. Pass 2 — humanizer structural-tell audit (vendored)

Audit against the catalog below, then merge every hit into the Pass-1 ranked list (Layer 1-2 structural hits are style-class advisories, density excess upgrades to strong; Layer-4 claim-evidence failures are integrity-class). Skip LaTeX comments and scaffolding macros. This pass never edits.

Layer 1 — general AI tells

Inflated significance; superficial "-ing" tails that fake depth; promotional or figurative language; vague attribution without a citation; copula avoidance ("serves as" -> "is"); negative parallelism ("not just X, but Y"); rule-of-three padding; elegant variation (one referent keeps one name for the whole paper); filler ("it is worth noting that", "in order to"); overlong clause-stacked sentences (2.11); em-dashes (remove entirely; recast with commas, colons per the colon rule, parentheses, or separate sentences). Lexical tells are enforced mechanically by the linter — run it, do not re-derive the lists here.

Layer 2 — academic AI tells

• 2.1 Over-claiming verbs. prove / establish / confirm / guarantee / demonstrate, checked against the evidence (§7 corpus overrides).
• 2.2 Significance hype. paves the way for, a crucial/pivotal step toward, opens new avenues, sheds light on, bridges the gap -> the specific failure mode or result addressed.
• 2.3 Empty intensifiers. extensive/comprehensive/thorough experiments, a wide range of, numerous, various -> enumerate or quantify.
• 2.4 Novelty padding. "novel" more than once per section; "to the best of our knowledge"; "for the first time" -> state the specific gap.
• 2.5 Formulaic openers. "In recent years...", "With the rapid development of...", "Despite recent advances,..." -> open with the structural fact or the problem.
• 2.6 Connective overuse. No consecutive sentences opening with Moreover/Furthermore/Additionally/In particular; let logic carry.
• 2.7 Contribution-list cliches. Each contribution names a specific result with its number, not a restatement of the abstract.
• 2.8 Citation dumping. Evidence-conditional, never a length rule: flag a bracketed list only when entries do not support the sentence or duplicate a role. Never delete a relevant source merely to shorten a list.
• 2.9 Hedging-by-vagueness. somewhat, relatively, fairly, to some extent, quite -> quantify or cut (distinct from calibrated hedging — Layer 3).
• 2.10 Boilerplate emphasis. "It is worth noting that", sentence-initial "Notably,"/"Importantly," -> if it matters, the sentence shows it.
• 2.11 Overlong clause-stacked sentences. Past ~30 words, 3+ subordinate clauses chained by which/that/while/with, or double-nested parentheticals. Split: one idea per sentence. The highest-yield structural check — the mechanical linter does not measure it.

The next five (2.12-2.16, from blader/humanizer) are structural, not lexical — flag the construction, honour each corpus caveat, never the bare word.

• 2.12 False ranges. "from X to Y" with categorical endpoints ("from theory to observation"). Caveat: a genuine quantitative range (redshift 0.1 to 0.5) is correct and required; flag only the rhetorical range whose endpoints share no scale.
• 2.13 Aphorism formulas. "X is the Y of Z" / "X becomes a Y" epigrams for resonance -> state the literal fact. Caveat: a formal definition of the same shape ("the aperture mass is the convolution of the shear with the filter") is a definition; keep it.
• 2.14 Persuasive authority tropes. "at its core", "fundamentally", "the real question is", "it is important to realize" -> cut, or give the specific reason. Caveat: "fundamentally different" naming a real physical distinction (regime, symmetry, scaling) is substantive; keep it.
• 2.15 Manufactured staccato drama. A run of terse fragments strung for impact -> measured sentences; academic emphasis comes from content. A single emphatic short sentence is fine (Layer 3); only a run is a tell.
• 2.16 Hyphenated-pair predicate overuse. Compound modifiers piled in the predicate ("the result is model-dependent") -> prefer the plain relation ("the result depends on the model") where it reads naturally. Fine as an attributive modifier; flag density, not any single use.

Pass-2 self-interrogation (blader/humanizer two-pass audit)

After the first draft of any Pass-3 rewrite, ask of it: "what still reads as machine-written here?" Answer in 2-4 concrete bullets (a specific phrase, a rhythm, a residual tell), then apply one further targeted rewrite that clears them and re-checks the fidelity and length gates. Do not manufacture voice to pass this step.

Layer 3 — preserve these (do NOT over-correct)

Evidence-tied hedging ("suggests", "is consistent with", "we hypothesize", "may indicate", "appears to") is correct and required when the claim is genuinely uncertain; strengthening a calibrated verb is a claim-evidence defect. Passive voice is fine when the actor is irrelevant; first-person plural "we" is standard; semicolons and an occasional triple are fine; formal definitions, named methods, symbols, and equations stay verbatim. False-positive guards: do not flag formal vocabulary on its own; a single mixed register; straight-vs-curly quotes in isolation; a lone emphatic short sentence; text inside quotations, a caption's verbatim labels, or a worked example; or field-common knowledge. Specific hard-to-fabricate detail, a genuine mid-sentence self-correction, and calibrated hedging are human signals, not tells.

Layer 4 — claim-evidence discipline

For every empirical claim: (a) is it backed by a number, figure, table, or citation in the text, and (b) does the verb match that evidence's strength? Unbacked -> add the pointer or soften. Verb too strong -> downgrade to what the data show. Vague magnitude -> a number or an attributed range; lead comparisons with the strongest baseline. Layer-4 failures are integrity-class, not style-class.

Layer 5 — voice and venue

If the author supplies prior papers, sample them first and match sentence rhythm, connective habits, hedging placement, and section openers. Otherwise default to clean, precise, venue-appropriate prose; the corpus dossier (§1) is the measured voice baseline. Funding proposals route to /sci-paper:proposal-polish, never here.

4. Pass 3 — claim-first rewrite

Select work in ranked order: L0 targets in the requested span first, then strong L1/L2/L3 advisories with concrete actions, then document-shape findings. For each selected paragraph:

1. Protect. Write the protected claim record (<scratch>/claim.txt): every claim and its evidence relation, all numbers/units, citations, named entities, inline math and acronyms, comparison/negation/causal direction, scope, stance, qualifiers. Re-read every numeric and citation source in the same turn. Snapshot the original verbatim (<scratch>/original.txt) as the §5.3 length baseline.
2. Generate. Produce N candidates (default 5) from the claim record, not by editing sentence-by-sentence. Each preserves every protected item, states the claim at the same strength and scope, adds no new fact, avoids Tier A and em-dashes, keeps Tier B within cap, breaks the specific L2 pattern that motivated the rewrite, keeps technically necessary lists, and is no longer than the original unless a protected invariant forces it.
3. Gate + rank.

   python tools/rewrite_reward.py --field <field> \
     --reference <scratch>/claim.txt --original <scratch>/original.txt \
     --candidates <scratch>/cand_0.txt <scratch>/cand_1.txt ...
   

   eligible=False (fidelity failure, or longer than the original → -inf) cannot be selected; --allow-growth "<reason>" only with an author-approved justification, printed into the run record. Among eligible candidates rank by L0 reduction, semantic fidelity, voice score, and condensation. If none is eligible, preserve the original and regenerate tighter — never pick the least-bad ineligible candidate. Then verify entities, scope, stance, and logical dependencies against the source by hand: the deterministic checker is necessary, not sufficient.
4. Re-measure. Re-run the Pass-1 linter on the candidate in enough section context for section caps. A good rewrite introduces no new integrity blocker or L0 target, stays eligible, acts on the selected advisory, worsens no higher-priority finding, does not grow the passage without a stated reason (report the length delta), and reports residual advisories.
5. Apply. Show before/after, protected invariants, eligibility, affected findings, and dispositions; apply with a minimal Edit and re-read the changed region in context. Run the Pass-2 self-interrogation on the applied text.

Document-shape findings (dispersion manifold, role coupling) cannot be moved by word-level rewriting. Run python tools/deai_partition.py <file> --field <field> and surface its fidelity-free merge/split suggestions to the author verbatim; they change zero tokens and are applied by hand, never automatically.

Optional self-distillation (--distill)

Only paragraphs explicitly accepted by the author may enter style-profile/<field>/exemplar_paragraphs.jsonl as positive voice anchors, with provenance and section type recorded. Never feed automatically generated or merely detector-favored text back as human positive data. Retraining on the grown bank requires held-out and confound-aware evaluation; a multi-minute heavy run executes on the authorized compute environment, not locally.

5. Stopping rule and report

Stop only when: all integrity blockers are resolved or verified false positives; applicable L0 targets are zero; every strong advisory is acted, accepted, rejected_as_false_positive, or pending with a stated reason; ordinary advisories and unavailable axes are reported; and every changed paragraph passes scientific-fidelity verification. This is a disposition-complete feedback state, not a universal prose verdict. If the iteration budget is exhausted, leave the original text for unresolved cases and return the pending findings.

# de-ai — Feedback Report
Target: <file> | Field: <field> | Candidates/paragraph: N
Measurement: <axis=status list, all four states shown>
Passes run: measure / humanizer-audit / rewrite (or audit-only)
Paragraphs considered: K | rewritten: R | unchanged: U
L0 targets: before B -> after A   (Tier A, em-dash, Tier B excess)
Strong advisories: acted X | accepted Y | false-positive Z | pending P
Length delta (rendered prose): <per-paragraph words, §5.3>

## Paragraph changes
- <line/section>: <finding ids and rules>
  - eligibility, missing invariants, before/after, disposition

## Residual feedback
<ranked ordinary advisories, document-shape suggestions, unavailable axes>
No number, equation, or citation changed.

6. Anti-patterns

• Treating synonym swaps or connector deletion as a structural rewrite.
• Quoting a manuscript number or citation from the dossier or memory.
• Forcing every advisory to zero, or every sentence-length outlier to the mean.
• Manufacturing raggedness, voice, or personality to lower a detector score.
• Calling a compatibility threshold calibrated, or a stale profile fresh.
• Declaring a paper human- or machine-authored from a learned score.
• Answering a finding by appending an explanatory clause (the §5.3 patch).
• Deleting a relevant citation merely to shorten a list.
• Copying or lightly paraphrasing an exemplar.
• Doing condense's job: bulk redundancy deletion belongs to /sci-paper:condense; de-ai only rewrites what a tell finding selects.

7. Corpus overrides and interfaces

Corpus overrides (measured evidence beats the upstream word lists): landscape is a legitimate astrophysics term — never flag it lexically; demonstrate* and significantly are normal astro usage — flag only on a Layer-4 failure, never as bare words. Tier A/B lists live in style-profile/<field>/lexicon.json and tools/ai_ism_lint.py; this skill defers to them. When the corpus changes, re-run tools/extract_style.py and tools/build_profile.py.

Interfaces:

• docs/SCIPAPER_STANDARD.md — consequence classes, ranking, disposition, stopping, §5.2 de-AI-ization order, §5.3 length budget, §6 rewrite eligibility.
• /sci-paper:paper — writing guidance and canonical L0 prose examples.
• /sci-paper:condense — the redundancy/length action surface (boundary statement in §0).
• /sci-paper:paper-review — invokes this skill --audit-only as its dimension-D structural-tell step.
• /sci-paper:final-review — runs this skill --audit-only as an isolated reviewer.
• Tools: ai_ism_lint (L0 hub + aggregation), deai_metrics, deai_oracle, deai_structure, deai_docstructure, deai_anchoring, deai_voice, deai_partition, deai_provenance, deai_personal, rewrite_reward (fidelity + length gate), retrieve_exemplars, extract_style, build_profile, deai_feedback (schema + ranking).