Article Examine

Article Examine is a local-first Codex Skill for traceable academic-manuscript review. It focuses on four practical questions:

1. Has the scientific data implied by the article been submitted, and are there visible provenance risks?
2. Does the manuscript text accurately describe what the published Figures show?
3. Do the Figures contain obvious labels, axes, units, legends, clipping, duplication, formula, or internal-logic errors?
4. What remains for a qualified human reviewer to decide?

The default workflow does not attempt to recreate every Figure from raw instrument data. Readable numeric anomaly screening is part of Stage 1 when suitable series exist. Native format decoding, source-to-Figure reconstruction, and battery-specific electrochemistry rules remain optional extensions.

Install in Codex

Clone or download the repository, then copy the Skill folder into your Codex skills directory:

cp -R skill/article-examine ~/.codex/skills/

Restart Codex or begin a new task. The installed directory should be ~/.codex/skills/article-examine/ and contain SKILL.md, agents/, assets/, references/, and scripts/.

Runtime and context requirements

The full workflow requires a local agent runtime with filesystem access and Python or shell tools. CC Switch or another generic API router is usable only when the calling client provides that local tool bridge. Do not attach a whole manuscript folder as text or Base64 in one API request.

In Codex desktop or Codex CLI, you may invoke the Skill directly and provide a local directory path:

Use $article-examine to examine /path/to/paper-project and its submitted data.

This is the simplest supported entry when Codex can access that directory. The files remain local and are opened in bounded stages. If the directory is outside the active workspace, Codex may ask for filesystem permission. Supplying a path is different from attaching or encoding the folder.

For Codex CLI, use the pre-request launcher instead of attaching the project in a chat window:

python3 ~/.codex/skills/article-examine/scripts/launch_article_examine.py \
  "/path/to/paper-project" --launch

The launcher runs before the first model request. It starts a fresh local task with the paper directory as the working root, sends only a short instruction with zero attachments, and enforces model_context_window=1000000 plus model_auto_compact_token_limit=900000 for that invocation. It does not modify the global Codex or CC Switch configuration. Omit --launch to inspect the generated plan first under article-examine-audit/launcher/.

This distinction is essential: instructions inside SKILL.md cannot repair an oversized first request because the server must accept that request before the Skill can run. If a client has no local filesystem/tool bridge, the launcher cannot turn it into one; use Codex CLI locally or a client-side orchestration layer that sends bounded batches.

For an already running local task, run the metadata-only preflight before inventorying a project:

python3 skill/article-examine/scripts/preflight_context_budget.py /path/to/project \
  --output /path/to/audit-output/context-preflight.json \
  --report-md /path/to/audit-output/context-preflight.md \
  --output-dir audit-output \
  --transport local-path

The audit then proceeds stage by stage. PDF/DOCX extraction, table profiling, frame sampling, and report rendering happen locally. The self-contained HTML report is a final sharing artifact and must not be sent back to the model because it contains Base64 images.

Example prompts

• Use $article-examine to examine this manuscript and its submitted-data folder.
• Check whether the Results section accurately describes every Figure.
• Screen the Figures for obvious presentation or internal-consistency errors.
• Derive an expected-data checklist from the paper and identify missing submissions.
• Generate a portable Chinese HTML report for human final review.

Four-stage review

1. Paper-derived submission completeness and numeric anomaly screening

Article Examine derives expected scientific units from Figures, captions, claims, Methods, and Supporting Information references before inspecting the submitted directory. File presence, mapping confidence, and readability are reported separately. A candidate mapping is not silently promoted to a confirmed one. Readable numeric series are then screened for periodic patterns, exact or scale/offset reuse, unexpectedly low residual variation under an explicit noise expectation, continuity constraints, and native/export conflicts. Instrument programs, setpoints, simulations, fitted/model outputs, and placeholders are excluded. Signals require human review and never constitute a fabrication or authenticity verdict.

2. Figure-to-text alignment

Manuscript statements are split into reviewable claims and checked for values, conditions, direction, sample ordering, trends, and interpretive strength. A visible association does not by itself establish a mechanism or causal explanation.

3. Obvious Figure-error screen

The complete Figure, panel crops, shared legends, and scale components are considered together. A clean screen is reported as “no obvious error found,” not as proof that the Figure is correct or authentic.

4. Human professional review

AI findings remain review aids. A qualified human records the final scope, unresolved issues, and professional guidance before the audit can be called complete.

Dependencies

Core inventory, contracts, and report generation mainly use the Python standard library. Install optional parsers only when their routes are needed:

python3 -m pip install -r requirements-optional.txt

Proprietary formats may still require vendor software or a validated export. A single unreadable format should limit only that route, not block the remaining article review.

Validate and test

If Codex's Skill Creator is installed:

python3 ~/.codex/skills/.system/skill-creator/scripts/quick_validate.py \
  skill/article-examine

Run the bundled case-independent tests:

python3 -m unittest discover -s tests -p 'test_*.py'

Interpretation and safety limits

• A received-file hash does not prove that a file was unchanged since acquisition.
• Completeness is measured against paper-derived scientific units, not against directory size.
• Missing Supporting Information remains a submission gap when the article cites it.
• An unreadable native file can still be present; readability and presence are separate.
• Automated anomaly signals do not establish fabrication, falsification, intent, or research misconduct.
• “No obvious error found” is not proof of scientific correctness or data authenticity.
• The manuscript and submitted data remain read-only; reports are written separately.
• Unpublished materials are not uploaded without explicit user consent.

License

Article Examine is released under the MIT License.