A Claude Code plugin that writes, reviews, de-AIs, and condenses scientific manuscripts for top-tier journals — under one typed standard, with every claim traced to a source and every unavailable measurement labelled as unavailable.
Built for ApJ / MNRAS / PRD / JCAP-class papers and NSF / NIH proposals. 8 skills, 24 tools, 208 tests, one normative contract.
中文文档 → · Documentation index → · The standard →
- The problem · What makes it different
- How it works · Install · Quick start
- Skills (8) · Tools (24)
- The feedback contract · Field-aware evidence
- Repository layout · Development · Status
Three failure modes ruin technical manuscripts, and none of them is a vocabulary problem:
- Prose still reads as machine-written after keyword cleanup. Swapping "delve" for "examine" leaves the deeper regularities untouched: smoothed sentence-length variation, templated construction, over-regular document shape, and claims with no evidence behind them.
- A single AI-detector score cannot tell an editor what to change — and it is confounded by field, source, section genre, length, jargon, and mathematical density. It answers a question ("who wrote this?") that an author does not need answered.
- Review silently converts missing evidence into good news. An uncalibrated axis reports zero findings, and zero findings read as clean.
sci-paper refuses all three. It emits typed, ranked, source-traced findings instead of a score, it never claims to identify an author, and an axis it could not measure says so.
| Most writing tools | sci-paper | |
|---|---|---|
| Output | One opaque score, or a rewrite | Typed findings with stable IDs, ranked by consequence, each with a source trace and a recommended action |
| Missing calibration | Silently reported as clean | Explicitly unmeasured / degraded, with the reason and what would restore it |
| Rewriting | Optimizes a style score | Hard scientific-fidelity eligibility first: drop or invent a number, unit, citation, equation, acronym, comparison direction, negation, or causal direction and the candidate scores -inf and cannot win |
| Length | Grows with every "fix" | A mechanical length budget: the default direction of every edit is shorter; growth needs a recorded justification |
| Authorship | "87% AI-generated" | Never. Learned scores are field-similarity triage, structurally capped at 0.5 confidence at paragraph scale |
| Style reference | A generic model prior | Your field's own corpus — descriptive statistics and exemplars extracted from papers you supply |
| Negative results | Quietly dropped | Recorded. Three surface formalisations of the "thesis spine" signal were built, measured, and refuted; the rule ships as a writing rule with no detector, and the standard forbids building a threshold on it |
That last row is the design philosophy in miniature: a refuted detector is evidence, and it stays in the record. See docs/architecture/EVALUATION.md.
┌──────────────────────────────────────────────┐
your field │ style-corpus/<field>/tier-{1,2,3}-*/ │ papers you supply
corpus ──▶│ → extract_style.py → style-profile/<field>/ │ (read-only, gitignored)
└──────────────────────────────────────────────┘
│ descriptive statistics, exemplars, calibrated references
▼
draft.tex ──▶ measure ──▶ typed findings ──▶ rank ──▶ edit ──▶ re-measure ──▶ disposition
│ │
│ L0 lexicon + punctuation + domain register │ every strong
│ L1 information distribution, surprisal / UID │ advisory ends
│ L2 sentence templates, salience, document shape │ as acted /
│ L3 learned field similarity (capped, degraded) │ accepted /
│ L4 positive voice + cooperative repair │ false-positive
└────────────────────────────────────────────────────┘
The loop is the whole product: measure → type → rank → edit → re-measure → disposition. It terminates in a disposition-complete state — not in "zero advisories", and never in a paper-level PASS.
git clone https://github.com/skymanbp/sci-paper.gitRegister it with Claude Code:
claude --plugin-dir /path/to/sci-paper # developmentSkills are then namespaced /sci-paper:<name>.
Python ≥ 3.11. The shared schema, the deterministic L0 linter, the model-free L1/L2 axes, the document-structure analysis, and the validator are standard library only. Optional capabilities add dependencies:
pip install -r requirements.txt # all optional extras| Package | Enables |
|---|---|
pymupdf |
PDF corpus extraction, compiled-page inspection |
sentence-transformers |
semantic exemplar retrieval, embedding features |
scikit-learn + joblib |
legacy and learned field-similarity models |
transformers + torch |
token-surprisal / UID measurement |
numpy |
learned feature, cache, and rewrite-score utilities |
Never install an optional dependency merely to turn an unavailable axis into a nominal score. A missing package keeps its axis
unmeasured, by design.
# 1. Verify the checkout.
python tools/validate_plugin.py
python -m unittest discover -s tests -v
# 2. Put your field's papers under style-corpus/<field>/tier-*/ and build the profile.
python tools/build_profile.py --field wgl
# 3. Produce one unified, machine-readable feedback report.
python tools/ai_ism_lint.py draft.tex --field wgl \
--structure --distribution --document-structure --register --salience \
--oracle --voice --format json --output feedback.jsonThen drive it from Claude Code:
/sci-paper:paper # load the writing standard
/sci-paper:de-ai draft.tex --field wgl # measure → audit → faithful rewrite
/sci-paper:condense draft.tex # remove redundancy, prove the shrink
/sci-paper:paper-review draft.tex --field wgl # A–R source-traced review
/sci-paper:figure-review draft.pdf # compiled-page figure evidence
/sci-paper:final-review draft.tex --field wgl # isolated multi-reviewer orchestration
/sci-paper:brainstorm "topic" # radial research exploration
/sci-paper:proposal-polish grant.tex --agency nsf # proposal register
Four jobs: write, revise, review, explore.
| Skill | What it does |
|---|---|
paper |
The writing framework: accuracy rules, formula and citation conventions, forward narrative, the L0 lexical policy with canonical examples, positive-voice guidance, measurement states, and stopping semantics. |
proposal-polish |
Funding-proposal register (NSF Project Summary/Description, NIH Specific Aims, fellowships). Keeps the vision-and-feasibility language a paper would trim, enforces claim–feasibility matching, and edits the score-forming first pages hardest. Never invents preliminary data, partners, or letters. |
| Skill | What it does |
|---|---|
de-ai |
Three chained passes: subsystem measurement (L0–L4), a vendored humanizer structural-tell audit, then claim-first rewriting — prose is rebuilt from the protected claim graph, not polished in place. --audit-only runs passes 1–2 for review integration. |
condense |
Whole-document redundancy elimination under one-canonical-home-per-fact, loop-until-dry convergence, and the mechanical length gate as the closing proof. |
| Skill | What it does |
|---|---|
paper-review |
Source-traced A–R review: mathematics, physics, logic and statistics, language and de-AI, structure and narrative spine, citations, data and figures, interfaces, redundancy, reproducibility, modern-physics checks, consistency, adversarial verification (three passes + twelve-framing escalation), staleness, process artifacts, draft language, citation precision, glossary alignment. |
figure-review |
Reviews compiled pages at 150 DPI — not source. Traces figure/caption/data provenance, measures canvas balance at the pixel level, and separates scientific and build contradictions from readability and aesthetic advisories. |
final-review |
Parent orchestrator. Runs paper-review, figure-review, de-ai --audit-only, and modern-physics-review as isolated worktree agents, merges their typed findings, and verifies a stable disposition-complete state across consecutive rounds. |
| Skill | What it does |
|---|---|
brainstorm |
Radial research-direction explorer: twelve framing passes per node, glossary-anchored terminology, complete derivation per branch, recursive divergence to convergence. Deferred or incomplete leaves are hard-banned. |
Every finding any of these tools emits uses the same sci-paper.feedback.v1
contract; the corpus, training and data-asset entries below produce artifacts
rather than findings. Full per-tool calibration and failure behavior:
tools/README.md.
| Tool | Purpose |
|---|---|
tools/deai_feedback.py |
Implements sci-paper.feedback.v1: stable IDs, consequence classes, measurement states, dispositions, ranking, summaries, rendering. Standard library only. |
tools/ai_ism_lint.py |
The unified CLI. Aggregates L0 and every advisory axis into one ranked text/JSON report. Exit 0 = no L0 target, 1 = L0 target present, 2 = invalid input or execution failure. |
tools/length_gate.py |
Per-section prose length-budget delta gate (standard §5.3). Exit 1 on net unjustified growth between two document versions; --allow records justifications. |
tools/rewrite_reward.py |
Ranks rewrite candidates after hard scientific-fidelity eligibility. Dropping or inventing a protected invariant scores -inf. |
| Tool | Purpose |
|---|---|
tools/deai_register.py |
Domain register: terms the manuscript leans on that the field's own corpus does not carry, judged by corpus document frequency rather than a curated cross-discipline list. Compounds are judged by their rarest part. Advisories only. |
tools/ai_ism_negatives_handcrafted.txt |
Seed negative examples for the legacy classifier (data asset). |
| Tool | Purpose |
|---|---|
tools/deai_metrics.py |
Model-free information-distribution findings — sentence-length variation, connective openers — with explicit calibration state. |
tools/deai_oracle.py |
Optional token-surprisal and Uniform Information Density evidence. Unavailable assets and compatibility thresholds stay explicit. |
| Tool | Purpose |
|---|---|
tools/deai_structure.py |
Sentence and paragraph construction: enumeration, repeated frames, parallel runs, symmetry, and related template families. |
tools/deai_salience.py |
Salience hierarchy: how far a passage's measured quantities run without an interpreting sentence between them, against a per-section human reference. Sole consumer of the numeral-preserving LaTeX projection. |
tools/deai_docstructure.py |
Whole-document rhetorical shape and complete-document calibration: dispersion band, per-length-stratum joint manifold, role coupling, split-conformal operating points. |
tools/deai_anchoring.py |
Section-class conditional claim-anchoring band — a writing-quality axis, explicitly not an AI-discrimination axis. |
| Tool | Purpose |
|---|---|
tools/deai_features.py |
Reusable distributional, UID, punctuation, embedding, and structural features. |
tools/deai_voice.py |
Optional learned field-similarity triage. A bundle without an operating point is degraded and never an authorship verdict. |
tools/train_voice_model.py |
Trains the optional field-similarity model with source-paper grouping. Confound audits are mandatory. |
| Tool | Purpose |
|---|---|
tools/deai_partition.py |
Fidelity-free merge/split suggestions that move a document toward the human dispersion band. Suggest-only, zero-token operations. |
tools/deai_provenance.py |
Editing-provenance ledger over the author's own draft history; labels each span AI-untouched → author-original by token edit ratio. Not a detector; unmeasured without an AI-draft ancestor. |
tools/deai_personal.py |
Personal dispersion baseline against the author's own prior papers — a confound-free same-author reference. unmeasured below three papers. |
| Tool | Purpose |
|---|---|
tools/build_profile.py |
Builds the basic field profile: extraction, optional legacy classifier, exemplar-cache warm-up. |
tools/extract_style.py |
Extracts lexicon, sentence statistics, transitions, a descriptive dossier, and a section-typed exemplar bank. Owns the two named LaTeX projections and the section-bucket vocabulary. |
tools/retrieve_exemplars.py |
Retrieves section- and topic-matched exemplar paragraphs, with embedding or explicit fallback retrieval. |
tools/fetch_arxiv_abstracts.py |
Fetches dated abstract corpora for controlled evaluation and training, optionally restricted to a subfield query set and named refereed journals. Rate limiting stops the sweep and exits 2 rather than writing a truncated corpus as if it were complete. |
| Tool | Purpose |
|---|---|
tools/train_ai_ism_classifier.py |
Trains the legacy word-ngram classifier, used only as degraded advisory evidence. |
tools/extract_md_negatives.py |
Harvests candidate generated paragraphs for controlled evaluation and training. |
tools/validate_plugin.pyis a development and release tool, not a shipped product tool, and is excluded from the count above.
Every finding carries exactly one consequence class:
| Class | Meaning | Required consequence |
|---|---|---|
integrity_blocker |
The scientific record may be wrong, unsupported, inconsistent, unreproducible, or unusable | Must be resolved from sources. Cannot be waived as a style preference. |
l0_target |
A Tier A word, an em-dash, or a Tier B word beyond one occurrence per section | Rewrite to zero. Not a claim that the paper is scientifically invalid. |
advisory |
Structural, distributional, learned, rhetorical, clarity, or aesthetic evidence | Rank, act on the strongest, then record a disposition for the rest. |
Every axis reports one measurement state — measured, degraded,
unmeasured, or not_applicable — and a final report lists all four.
Silence is never read as clean.
Every strong advisory ends at one disposition: acted, accepted,
rejected_as_false_positive, or pending with a stated reason. Ordinary
advisories stay visible and do not have to disappear.
A field is one subdirectory under style-corpus/ with a matching directory
under style-profile/. With exactly one field present, tools auto-detect it;
with several, pass --field <name> explicitly. Nothing assumes a particular
field exists.
style-corpus/<field>/tier-1-top/ top-journal exemplars
tier-2-mentor/ mentor or target-author exemplars
tier-3-reference/ other relevant field papers
| python tools/extract_style.py --field <field>
v
style-profile/<field>/ generated evidence (gitignored)
Corpus contents are read-only, copyright-sensitive inputs and are never committed. Generated dossiers and exemplars may quote source prose and must not be published unless their rights permit it. A corpus dossier is descriptive evidence — not a normative standard, and not proof of human or machine authorship.
Whole-document calibration requires complete papers as independent observations. Paragraph exemplars cannot be relabelled as independent documents.
sci-paper/
├── .claude-plugin/ plugin.json · marketplace.json
├── .github/workflows/ ci.yml — validator + test suite on every push and PR
├── docs/ ← documentation index at docs/README.md
│ ├── SCIPAPER_STANDARD.md the single normative contract (v3.6)
│ ├── architecture/
│ │ ├── DEAI_SUBSYSTEM.md implementation architecture
│ │ └── EVALUATION.md current metrics, gaps, confounds, refuted signals
│ └── design-notes/ frozen, dated reasoning records (not status)
├── skills/<name>/SKILL.md 8 skills
├── tools/ 24 product tools + the repository validator
├── tests/ 15 test files, 208 tests
├── style-corpus/<field>/ user-supplied read-only corpus (gitignored)
├── style-profile/<field>/ generated and calibrated evidence (gitignored)
├── CHANGELOG.md per-version history
└── CLAUDE.md working rules for this repository
| Read this | For |
|---|---|
| docs/README.md | Documentation index and authority order |
| docs/SCIPAPER_STANDARD.md | The normative contract. If anything disagrees with it, it wins. |
| docs/architecture/DEAI_SUBSYSTEM.md | How the subsystem is built |
| docs/architecture/EVALUATION.md | What is measured, what is not, and every known confound |
| tools/README.md | Per-tool registry, calibration, failure behavior |
| style-corpus/README.md | Supplying a field corpus |
| style-profile/README.md | Generated assets and build boundaries |
python tools/validate_plugin.py # 9 contract checks
python -m unittest discover -s tests -v # 208 testsThe validator checks release metadata, skill frontmatter, standard references,
documentation authority boundaries and index completeness, recorded suite sizes
against real discovery, stale contract markers, product registries, Python
syntax, runtime imports, CLI entry points, schema fields, linter exit semantics,
Tier B behavior, tests, and CI wiring. The authoritative list is
tools/validate_plugin.py itself.
A release additionally requires independent code review, clean-checkout verification, and a green hosted CI run on the release commit.
Current: v0.27.1. Full per-version history in CHANGELOG.md.
- Normative core:
docs/SCIPAPER_STANDARD.mdv3.6 — the complete de-AI standard in one file (layered model, document-scale detection core, cooperative layer, thecalibration_unitconfidence cap, the §5.2 de-AI-ization procedure, the §5.3 condense-not-accumulate rule with mechanical enforcement, and the §5.4 thesis spine shipped deliberately without a detector). There is no separate de-AI standard. - Skills (8):
paper,de-ai,condense,paper-review,figure-review,brainstorm,final-review,proposal-polish. - Tools (24): the registry above is exact.
- Calibrated gaps, stated plainly: no learned-model operating point (the L3
document-level surprisal path is measured not to provide one); no completed
author hard-set labels; and the cooperative-layer tools (
deai_provenance,deai_personal) are honestlyunmeasureduntil the author supplies their own draft history or prior papers. - Field-specific guidance: weak-lensing scientific anchors stay marked
[WGL]where they apply. Shared writing and review policy is field-agnostic.
- AIScientists-Dev/academic-humanizer
(MIT) — the 2026-07-16 lexicon extensions (
underscore*,pivotal,tapestry,testament,realm*,intricate,foster*), theserves as,ing-tail, andcolon-elaborationlinter rules, the Claim–Evidence Discipline and Preserve List sections ofskills/paper/SKILL.md, theproposal-polishskill, and thede-aiskill's Layer 1–5 audit catalog adapt its material. Every lexical adoption was re-verified against the curated field corpora before tier assignment; venue-specific rules that conflict with astro usage (landscape, blanketdemonstrate/significantlybans) were deliberately not adopted. academic-humanizer itself builds on blader/humanizer (MIT). - blader/humanizer (MIT) — the
de-aiskill's structural patterns 2.12–2.16 (false ranges, aphorism formulas, persuasive-authority tropes, manufactured staccato drama, hyphenated-pair predicates), its Pass-2 self-interrogation step, and its false-positive guards adapt this skill. Only its academically relevant structural tells were absorbed; its blog and chat-specific patterns (emoji, title-case headings, chatbot artifacts, curly-quote flags) and itslandscape-flagging word list were deliberately not adopted, because corpus evidence governs here.
MIT covers code, skills, documentation, and tooling authored in this repository. User-supplied corpus contents and generated excerpts retain their source rights and are not covered by this repository license.
Keywords: Claude Code plugin · agent skills · scientific writing · academic writing · paper review · peer review · manuscript preparation · AI text detection · AI-generated text · humanizer · de-AI · LaTeX · arXiv · astrophysics · weak gravitational lensing · cosmology · ApJ · MNRAS · PRD · JCAP · NSF proposal · NIH Specific Aims · research writing assistant · corpus-driven style · reproducibility · scientific integrity · LLM tooling.