sci-paper

A Claude Code plugin for scientific writing, rewriting, source-traced review, figure review, narrative analysis, adversarial critique, and research ideation.

The single normative authority is docs/SCIPAPER_STANDARD.md (v3). It defines scientific integrity blockers, narrow L0 rewrite targets, ranked advisories, explicit measurement states, author dispositions, and scientific-fidelity requirements. It also is the de-AI standard: the layered signal model (L0–L4), the document-scale detection core, the cooperative repair layer, the calibration_unit confidence cap, the ordered de-AI-ization procedure (§5.2), and a disposition for every open item all live in that one file — there is no separate de-AI standard. Corpus profiles and learned models provide evidence; they do not define a separate paper verdict.

What it ships

Skills (8)

Skill	Purpose
paper	Normative scientific-writing framework: accuracy, formula and citation rules, forward narrative, L0 policy, positive voice guidance, measurement states, and stopping semantics.
de-ai	The single de-AI skill: three chained passes — subsystem measurement (L0–L4), the vendored humanizer structural-tell audit (AIScientists-Dev/academic-humanizer Layers 1–5 + blader/humanizer patterns 2.12–2.16 and Pass-2 self-interrogation), and claim-first rewriting under hard fidelity eligibility and the §5.3 length budget. Loads the field corpus dossier and exemplars as descriptive calibration. --audit-only runs passes 1–2 for review integration.
condense	精简: whole-document elimination of all unnecessary and duplicated content. Executes standard §5.3 (delete > condense-in-place > same-length; growth only with recorded justification) with one-canonical-home-per-fact deduplication, loop-until-dry convergence, and the length gate as the closing proof.
paper-review	Source-traced A–R review across mathematics, physics, logic, language, structure with the narrative-spine protocol (purpose record, contribution graph, cold-read questions), citations, data, interfaces, reproducibility, consistency, adversarial verification (three-pass plus 12-framing escalation to CONFIRMED/REFUTED/MARGINAL), staleness, process artifacts, draft language, citation precision, and glossary alignment.
figure-review	Reviews compiled pages at 150 DPI, traces figure/caption/data provenance, measures canvas balance at the pixel level, and separates scientific/build blockers from readability and aesthetic advisories.
brainstorm	Radial research-direction explorer with twelve framing passes, source checks, glossary anchoring, recursive branching, and bounded search convergence.
final-review	Parent orchestrator for isolated paper-review, figure-review, de-ai audit, and modern-physics-review runs. Verifies a stable disposition-complete state rather than zero advisories.
proposal-polish	Funding-proposal editing mode (NSF/NIH). Keeps the vision-and-feasibility register a paper would trim, enforces claim-feasibility matching, edits the score-forming first pages hardest, and applies the shared L0 policy. Never invents preliminary data, partners, or letters.

See CHANGELOG.md for per-version history.

Tools (22)

Tool	Purpose
tools/build_profile.py	Builds the basic field profile: extraction, optional legacy classifier, and exemplar-cache warm-up.
tools/extract_style.py	Extracts lexicon, sentence statistics, transitions, a descriptive dossier, and a section-typed exemplar bank.
tools/retrieve_exemplars.py	Retrieves section- and topic-matched exemplar paragraphs with embedding or explicit fallback retrieval.
tools/ai_ism_lint.py	Unified L0 and advisory CLI with ranked text/JSON output and exit statuses 0/1/2.
tools/length_gate.py	Per-section prose length-budget delta gate (standard §5.3): exit 1 when NET unjustified growth between two document versions exceeds tolerance; growing sections get strong advisories; records --allow justifications.
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/training.
tools/ai_ism_negatives_handcrafted.txt	Seed negative examples for the legacy classifier.
tools/deai_feedback.py	Implements sci-paper.feedback.v1: stable IDs, consequence classes, measurement states, dispositions, ranking, summaries, and rendering.
tools/deai_metrics.py	L1 model-free information-distribution findings with explicit calibration state.
tools/deai_structure.py	L2 sentence/paragraph construction analysis for enumeration, repeated frames, parallel runs, symmetry, and related templates.
tools/deai_docstructure.py	Whole-document rhetorical-shape analysis and complete-document calibration: dispersion band, joint (per-length-stratum) manifold, role coupling, split-conformal operating points.
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_anchoring.py	Section-class conditional claim-anchoring band; a writing-quality axis, explicitly not an AI-discrimination axis.
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 an AI detector; unmeasured without an AI-draft ancestor.
tools/deai_personal.py	Personal dispersion baseline: compares a draft to the author's own prior papers, a confound-free same-author reference; unmeasured below three papers.
tools/deai_oracle.py	Optional surprisal/UID evidence; unavailable assets and compatibility thresholds remain explicit.
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 remain mandatory.
tools/rewrite_reward.py	Applies hard scientific-fidelity eligibility before ranking rewrite candidates.
tools/fetch_arxiv_abstracts.py	Fetches dated abstract corpora for controlled model evaluation/training.

The repository validator, tools/validate_plugin.py, is a development/release tool and is not counted as a shipped product tool. See tools/README.md for the complete registry and failure behavior.

Core feedback model

Every finding uses sci-paper.feedback.v1 and one consequence class:

• integrity_blocker: scientific/source/build contradiction that must be repaired or verified false;
• l0_target: Tier A, em-dash, or Tier B usage above one occurrence per section and word;
• advisory: structural, distributional, learned, rhetorical, clarity, or aesthetic evidence.

Each analysis axis reports measured, degraded, unmeasured, or not_applicable. Strong advisories require a disposition. Ordinary advisories remain visible and do not have to disappear.

The linter exit contract is intentionally narrow:

• 0: no L0 target; advisories may remain;
• 1: at least one L0 target;
• 2: invalid input, configuration failure, or execution failure.

Quick start

# Install optional full-pipeline dependencies.
pip install -r requirements.txt

# Validate the repository contract and run tests.
python tools/validate_plugin.py
python -m unittest discover -s tests -v

# Put field papers under style-corpus/<field>/tier-*/ and build the basic profile.
python tools/build_profile.py --field wgl

# Register the plugin for development.
claude --plugin-dir <path-to-this-repo>

# Produce unified feedback.
python tools/ai_ism_lint.py draft.tex --field wgl \
  --structure --distribution --document-structure --oracle --voice \
  --format json --output feedback.json

Example skill invocations:

/sci-paper:paper
/sci-paper:de-ai draft.tex --field wgl
/sci-paper:condense draft.tex
/sci-paper:paper-review draft.tex --field wgl
/sci-paper:final-review draft.tex --field wgl

Why explicit profiles and typed feedback

A single learned score cannot tell an editor what to change and is vulnerable to field, source, section, length, jargon, and mathematical-density confounds. sci-paper instead combines:

1. deterministic scientific and L0 rules;
2. corpus-derived descriptive statistics and exemplars;
3. sentence and whole-document structural evidence;
4. optional UID and learned field-similarity evidence;
5. claim-first rewriting with hard protected-invariant eligibility;
6. source-traced review and explicit author dispositions.

This design keeps measurements inspectable and replaceable. Missing calibration remains visible instead of becoming a nominal score. Current performance and gaps are recorded in docs/EVALUATION.md; implementation details are in docs/DEAI_SUBSYSTEM.md.

Field-aware evidence

A field is one subdirectory under style-corpus/ with a corresponding directory under style-profile/. With one field, most tools can auto-detect it; with several fields, pass --field <name> explicitly.

Corpus contents are read-only, copyright-sensitive inputs. 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.

Project structure

.
├── .claude-plugin/
│   ├── plugin.json
│   └── marketplace.json
├── docs/
│   ├── SCIPAPER_STANDARD.md       # sole normative writing/review contract
│   ├── DEAI_SUBSYSTEM.md          # implementation architecture
│   ├── EVALUATION.md              # current metrics, gaps, and confounds
│   ├── DEAI_ARCHITECTURE_ROADMAP.md  # design note (2026-07-12)
│   └── DEAI_FRONTIER.md           # design note (2026-07-13)
├── skills/
│   ├── paper/SKILL.md
│   ├── de-ai/SKILL.md
│   ├── condense/SKILL.md
│   ├── paper-review/SKILL.md
│   ├── figure-review/SKILL.md
│   ├── brainstorm/SKILL.md
│   ├── final-review/SKILL.md
│   └── proposal-polish/SKILL.md
├── style-corpus/
│   └── <field>/tier-{1,2,3}-*/    # user-supplied read-only corpus
├── style-profile/
│   └── <field>/                   # generated/calibrated evidence
├── tests/                         # schema, linter, structure, and fidelity tests
├── tools/                         # product tools plus repository validator
└── CHANGELOG.md

Build and calibration boundaries

build_profile.py builds the basic descriptive profile and legacy assets. Optional UID, learned field-similarity, sentence-structure, and whole-document calibration have separate tools and evidence requirements. Do not claim a measured axis merely because a model file exists.

A field policy asset should document:

• the independent sample unit;
• corpus selection and provenance;
• sample size;
• uncertainty method;
• operating point and applicability;
• leave-source/document-out human flag behavior;
• known confounds.

Without that record, the corresponding axis remains degraded or unmeasured.

Development and release verification

python tools/validate_plugin.py
python -m unittest discover -s tests -v

The validator checks release metadata, skill frontmatter, standard references, stale contract markers, product registries, Python syntax, runtime imports, CLI entry points, schema fields, linter exits, Tier B behavior, tests, and CI wiring.

A release additionally requires independent code review and clean-checkout verification.

Status

Current: v0.23.0. Full per-version history is in CHANGELOG.md.

• Normative core: docs/SCIPAPER_STANDARD.md v3.4 — the complete de-AI standard (layered model, document-scale core, cooperative layer, calibration_unit cap, the §5.2 de-AI-ization procedure, the §5.3 condense-not-accumulate rule with mechanical enforcement, auxiliary L2 template families with the blind perceptual panel as an L2 validation instrument, and a disposition for every open item). There is no separate de-AI standard.
• Skills (8): paper, de-ai, condense, paper-review, figure-review, brainstorm, final-review, proposal-polish.
• Tools (22): exact product registry above.
• Current calibrated gaps: no learned-model operating point (the L3 document-level surprisal path is now measured not to provide one, EVALUATION.md §9.8), no completed author hard-set labels, and the cooperative-layer tools (deai_provenance, deai_personal) are honestly unmeasured until the author supplies their own draft history / prior papers. These remain explicit in docs/EVALUATION.md.
• Field-specific guidance: WGL-specific scientific anchors remain marked where applicable. Shared writing/review policy is field-agnostic.

Acknowledgments

• AIScientists-Dev/academic-humanizer (MIT). The 2026-07-16 lexicon extensions (underscore*, pivotal, tapestry, testament, realm*, intricate, foster*), the serves as, ing-tail, and colon-elaboration linter rules, the Claim–Evidence Discipline and Preserve List sections in skills/paper/SKILL.md, the proposal-polish skill, and the de-ai skill'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, blanket demonstrate/significantly bans) were deliberately not adopted. academic-humanizer itself builds on blader/humanizer (MIT).
• blader/humanizer (MIT). The de-ai skill'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/chat-specific patterns (emoji, title-case headings, chatbot artifacts, curly-quote flags) and its landscape-flagging word list were deliberately not adopted, since corpus evidence governs here (see the skill's Corpus Overrides).

License

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.