Verity Scribe

Evidence-first academic writing for Codex, Claude Code, Cursor, and compatible agents.

简体中文 · Skill · Contributing · Security

Academic writing is not a synonym-replacement problem. Verity Scribe builds and revises scholarly prose around an auditable chain:

flowchart LR
    Q["Research question"] --> E["Evidence ledger"]
    E --> C["Claims and warrants"]
    C --> D["Genre-aware draft"]
    D --> R["Review and rebuttal"]
    R --> S["Natural style pass"]
    S --> G["Final integrity gate"]

Loading

It combines claim-to-source traceability, argument architecture, bilingual style calibration, protected-span checking, peer-review simulation, and rebuttal management. Its anti-slop pass improves specificity and rhythm without weakening scientific hedging, deleting legitimate passive voice, or pretending that polished prose is verified evidence.

What it does

• Maintains a claim ledger with explicit evidence states: verified, user-supplied, inference, unverified, or unsupported.
• Separates citation existence, metadata identity, and claim-level support.
• Routes empirical papers, reviews, theses, proposals, grants, abstracts, translations, reviews, and rebuttals through distinct genre contracts.
• Revises in consequence order: integrity -> argument -> structure -> paragraph -> sentence -> style -> format.
• Calibrates voice from author-owned samples without impersonating a named living writer.
• Preserves numbers, units, equations, citation keys, identifiers, quotations, URLs, code, terminology, modality, and attribution.
• Treats formulaic AI-style signals as contextual clusters rather than banned-word rules.
• Selects and verifies current reporting guidance instead of name-dropping a checklist.
• Ships three dependency-free Python checks and reusable project templates.

What it will not do

• Fabricate citations, data, results, methods, quotations, policies, or journal requirements.
• Hide uncertainty or upgrade association to causation.
• Optimize text to evade AI detectors or misrepresent authorship.
• Upload unpublished material to an external service without explicit authorization.
• Treat a formatted bibliography or fluent paragraph as proof of correctness.

Install

Codex

Run the bundled Codex installer from an existing Codex environment:

python "$env:USERPROFILE\.codex\skills\.system\skill-installer\scripts\install-skill-from-github.py" `
  --repo Florent2025/verity-scribe `
  --path skills/verity-scribe

Restart the task or begin a new task so the skill catalog refreshes.

Manual / other agents

Clone the repository, then copy skills/verity-scribe into the agent's global or project skill directory. Keep the entire skill folder so its references, scripts, and templates remain available.

Use

Invoke the skill explicitly:

Use $verity-scribe to audit this literature review. Build a claim ledger, verify the citations you can, and mark everything else unverified.

Use $verity-scribe to revise this Chinese methods section conservatively. Preserve every number, unit, citation key, statistical symbol, and term; return a protected-span audit.

Use $verity-scribe to turn these reviewer comments and my revised manuscript into a response matrix and rebuttal letter. Do not claim any change that is absent from the manuscript.

The skill supports seven modes: plan, draft, revise, audit, review, rebuttal, and translate.

Deterministic checks

The scripts use only the Python standard library.

python skills/verity-scribe/scripts/audit_manuscript.py manuscript.md --format markdown
python skills/verity-scribe/scripts/compare_protected_spans.py before.md after.md
python skills/verity-scribe/scripts/validate_claim_ledger.py claim-ledger.csv

They flag review targets; they do not make scientific judgments automatically.

Repository layout

skills/verity-scribe/
├── SKILL.md
├── agents/openai.yaml
├── references/
├── scripts/
└── assets/
evals/cases.json
tests/test_scripts.py
tools/validate_repo.py

The installable skill is isolated under skills/verity-scribe/; repository documentation and CI stay outside its runtime context.

Quality model

Every substantial output passes six gates:

1. evidence provenance;
2. claim-source alignment;
3. design-appropriate inference;
4. argument and genre fit;
5. protected-span and cross-artifact consistency;
6. natural, author-calibrated language.

See evals/cases.json for adversarial behavior cases, the dated research-writing landscape for comparative analysis, and NOTICE.md for project influences. Verity Scribe is an original integration; upstream projects are credited rather than silently copied.

Development

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

Contributions should add a reproducible failure case, explain the integrity risk, and avoid giant phrase blacklists. See CONTRIBUTING.md.

License

MIT. See LICENSE.