Methodology and quality evidence

We distinguish software reliability, quote grounding, and human-reviewed model accuracy. Passing tests do not automatically mean legal or factual accuracy.

1. Extract

Parse the uploaded file, use OCR only when needed, and surface extraction warnings.

2. Ground

Extract contract facts separately and require exact source quotes for verified findings.

3. Analyze

Generate scores, risks, questions, and negotiation actions with a versioned prompt and model identifier.

4. Guard

Downgrade unsupported claims to questions and block strong legal recommendations without verified severe evidence.

Published product evidence

Automated test files: 11
Quality gate dimensions: 5
Annotated golden cases: 60

Critical-risk recall gate: ≥98%

High-risk false-positive gate: ≤2%

Unsupported-quote gate: ≤2%

Recommendation accuracy gate: ≥95%

Repeatability gate: 100%

These are release thresholds on synthetic, annotated contract clauses and deterministic normalization—not a claim of 95% accuracy on every real contract or jurisdiction. A separate, independently reviewable real-document benchmark is still required before publishing a general model-accuracy headline.