How contract analysis works — and what is validated today

1. Upload and validation: file type, size and page count are checked against configured limits; admission rules depend on the deployment mode.
2. Deterministic text extraction: existing PDF text layers are used directly; page coverage and readability are verified.
3. Selective local OCR: only image-based pages are read locally (Tesseract, in memory, no files written) with readability and coverage checks and bounded retries. Conflicting readings stay unresolved — nothing is silently truncated. Provider vision recovery is currently not enabled (documented gap).
4. Masking: detected personal identifiers are replaced by request-scoped placeholders before indexing and before any model call; an outbound guard checks every prompt.
5. Index and retrieval: the masked text goes into a per-request bounded Lucene/BM25 index; questions run as bounded batches in parallel through a central rate-aware scheduler (concurrency caps, token admission, 429 backoff, bounded retries).
6. Classification and analysis: contract type and review questions come from versioned frameworks; every question gets an explicit status.
7. Evidence verification and absence: “not found” is only claimed after document-wide absence validation; contradictory or missing evidence stays marked unresolved.
8. Canonical facts and report: facts and issues are deduplicated and reconciled; the report is generated in one resolved report language while quotations stay in the original language. Follow-up questions run only on the validated result.

Measured (deterministic CI run without a live model, 2026-08-28): language detection 1.000; classification EN 0.846 and DE 0.846 (gate ≥ 0.95 → FAIL); retrieval evidence recall EN 1.000, DE 0.818; evidence span overlap 1.000; known PII leakage 0; silent truncation 0. Local OCR (real, 2026-10-28): Unicode word recall at 150/300 DPI 0.78/0.76 (gate ≥ 0.75), character accuracy ≈ 0.92–0.93.

Test corpus: synthetic primary corpus (20 fixture groups, 134 objective facts), independent synthetic corpus (49 groups, 412 facts/questions, 66 expected absences), plus shareable real contracts (CUAD v1, external, test-only, license respected) and a combined multilingual corpus (EN/DE as native datasets; FR/IT as enabled, not yet validated report languages; RO as a configured target).

Benchmarked languages: EN and DE have native benchmark datasets. FR and IT report output was enabled on 2026-11-03 via the versioned per-language prompt templates (language.native.3/language.native.4 in the pinned catalog) and is pending benchmark validation; RO remains a configured target without validation. Report language never overrides the governing law of the contract.

NOT VALIDATED All model-generated metrics (critical-fact accuracy, NOT_FOUND precision/recall, negation handling, evidence grounding, follow-up quality) are NOT_EVALUATED in deterministic CI; live Mistral quality is not certified. End-to-end language certification is UNVALIDATED for all languages including EN and DE. Deterministic benchmarks do not establish live-model accuracy, severity calibration or provider latency.

As results improve, this table is extended with dated entries. Nothing on this page claims a certification, compliance or live accuracy that was not measured.