TraceClause v0.2.0 — Traceable corrections and recoverable reviews

TraceClause v0.2.0 — Traceable corrections and recoverable reviews

This preview adds manual requirement correction, multiple evidence passages, browser-local drafts and English/Chinese interfaces to the local document review workbench.

What changed

• Add omitted requirements from exact source spans, edit wording, and split compound requirements into 2–20 children.
• Preserve original quotations, stable identifiers and previous decisions. Revised requirements return to pending; split children expose parent history.
• Select multiple evidence passages and include them in Markdown, CSV and JSON reports.
• Recover review/editor drafts after switching tasks or refreshing. Competing saves preserve drafts and require explicit comparison with current content.
• Choose English or Chinese interface/report labels while leaving quotations and human rationales unchanged.
• Publish a separate 14-document synthetic set, including extraction and paraphrase failures. The BM25 engine remains unchanged.

Upgrade and try

Back up data/traceclause.sqlite3 with the service stopped. v0.1 snapshots are read additively; the next successful mutation stores the v0.2 schema. Original files and old events remain intact. Restore the pre-upgrade backup if downgrading.

Browser drafts are separate from the database, tied to this browser and site address. They are excluded from backups and exported reports. Save intended reviews before backup or sharing.

Follow the quick start, choose English in the top bar and load a demo. No model API key or Node runtime is required to use the application.

User guide · API · Evaluation

Validation

• 47 Python tests cover the existing baseline plus snapshot compatibility, exact source spans, multiple evidence, edit/split invalidation, stable IDs, capacity limits, stale updates and bilingual exports.
• Five Node tests cover draft persistence, window isolation, recovery, malformed entries and storage failures.
• Browser checks cover refresh/task-switch recovery, adding/editing/splitting, multiple evidence, inherited history, language/report selection and two-window stale-save handling.
• The new synthetic set reports exact extraction precision 92.3%, recall 80.0%, and retrieval Hit@1 / macro Recall@3 of 84.6%. These are small authored-set metrics, not independent or real-world accuracy claims.
• CI is configured for Windows/Linux with Python 3.11/3.13. See the release commit's Actions run for actual outcomes.

Scope

Local, single-user preview. No accounts, parser sandbox, OCR, semantic retrieval, document replacement or cross-version impact analysis. Bind to 127.0.0.1 and use trusted documents. Docker build/runtime and macOS runtime are not validated.

TraceClause v0.1.0 — First public preview

TraceClause v0.1.0 — First public preview

TraceClause is a local document evidence review workbench that connects requirements, response passages and human decisions with source references and rationales.

Available in this release

• PDF, DOCX, TXT and Markdown import, extraction and source locations.
• Explainable BM25 retrieval with candidate, possible-conflict and missing-evidence hints.
• Manual evidence selection, saved decisions, history and stale-update protection.
• Markdown reports, CSV checklists, JSON snapshots and original-file downloads.
• Chinese application interface, synthetic examples and English documentation.

Try it

Use Python 3.11+ and follow the quick start. Open the local page and choose the demo action. No model API key is required.

User guide · Example review · API examples · Roadmap

Validation and limitations

The application's 25 tests passed across Windows/Linux and Python 3.11/3.13. On ten development examples, first evidence hit rate is 7/9 and hint agreement is 8/10. Two paraphrase failures are retained. This is not an independent dataset and does not establish real-world accuracy.

OCR, semantic retrieval, version-impact analysis and multi-user authentication are not implemented. Use a trusted local, single-user environment. Docker build/runtime has not been validated. Automatic matches provide candidates; humans make decisions.

Licensed under MIT. Submit synthetic failures and suggestions through Issues, or follow the contribution guide.