Publishing DocJev

The destination repository is jerryjliu/docjev. The initial release commit has been pushed and its GitHub checks passed. The accuracy corpus, runner, results, and visual report are included in the follow-up repository release.

Repository

The current source bundle is output/release/docjev-0.1.0.zip, with a per-file release manifest and SHA-256 sidecar. The follow-up commits include the complete benchmark evidence and interactive visual report.

The package name and primary CLI are docjev; availability on PyPI has not been checked or reserved. Python imports remain jev_docs, and the older jev-docs CLI continues to work. Existing recordings and measured run artifacts retain their original names and evidence.

Visual report

The visual benchmark report includes a self-contained HTML explorer and a shareable summary image. Download the HTML to open it in a browser; GitHub renders HTML as source. The report makes no model calls. Keep its small-sample scope and exact-packet accuracy visible when sharing speed figures.

Demo assets

• Short social videos: Classification and Splitting, with thumbnails, captions, and run evidence.
• Longer walkthroughs: Classification and Splitting.
• Actual split PDFs and JSON.
• Source provenance and measured timing report.
• Separate 40-document accuracy report, error analysis, and dataset notices.

The videos are silent, captioned 1080p recordings. They can be uploaded directly as release assets or shared with the repository link. Every retained processing sequence plays at normal speed. The short clips show individual concurrent comparisons; the longer walkthroughs show a separate saved pilot with three timed repeats on one source per task. Neither establishes general accuracy. Keep that scope with any benchmark numbers reused in a post.

The repository includes source code, original public demo documents, separate synthetic evaluation fixtures, and all their notices. Local credentials, document caches, virtual environments, and raw screen recordings are excluded from the clean release archive.

Local verification

uv sync --all-extras --locked
uv run ruff check .
uv run mypy src/jev_docs
uv run pytest
uv build

These checks passed locally on macOS with Python 3.12: 177 offline tests passed and three opt-in live OCR tests were excluded. The initial release passed GitHub Actions; subsequent benchmark changes have local verification. All corpus pages passed rendered equality checks; source archives include the evaluation corpus while the wheel excludes it. The optional full synthetic benchmark has not been executed.