Documents

Measured on real court filings

100%

Document type — correct

97.4%

Filing party — correct

100%

Accurate when it renames — the rest it flags, never confidently wrong

<1s

Per document · 65 file types · 789 tests

Scored on 38 real California federal court filings (CourtListener), classified text-only with generic filenames — the honest worst case, no filename hints. Plus 448/448 (100%) on the internal edge-case battery.

0

Documents

0

AI Classified

0%

SOP Compliant

0

Flagged for Review

What MatterValet does — to every document, automatically

SOP-compliant auto-naming

A messy filename + the document's content becomes the exact formula: YYYY-MM-DD [DI-##] Party DocType ShortTitle_Status.ext.

448/448 on the internal battery

Accurate on real court filings

Classifies document type + filing party from the document's content — not a guess from the filename.

100% type · 97.4% party · 38 real filings

Never confidently wrong

When it can't resolve a document with confidence it does not rename — it flags the doc for a human instead of burying it under a wrong name.

100% abstention precision

Explainable, rule by rule

Every name carries per-field reasoning + the named SOP rule applied — e.g. "an editable .docx is a draft, not a filing". Open any Details panel.

~15 named, individually-tested rules

Real-world text robustness

Strips CA pleading line numbers, letterhead, Bates stamps + page furniture; isolates the primary doc in a scanned packet; survives OCR noise.

65 file types · accuracy measured after this step

Built to learn

Built to capture corrections humans make in Clio and refine classifications as data accumulates. The engine gets smarter over time without any manual tuning.

Learns from corrections · 789 tests

Drop files here or click to upload

15 MB max per file · 65 file types supported

Processing Queue 0 documents

📂

No documents yet

Drop files above or click "Run Demo" to see MatterValet classify documents in real time

Renamed Documents

✨

Nothing renamed yet

Drop a file to see MatterValet rename it in real time

Every number is reproducible from the open codebase — no marketing rounding. Source: evals/real_doc_eval.py.

100% document type (38/38) · 97.4% filing party (37/38) — independently scored fields on 38 real California federal court filings from CourtListener, across 14 document types.
Never confidently wrong: 100% accurate when it acts (37/37). Every key-field error the engine would have made, it abstained on instead — it either named the document correctly or flagged it for human review. The single party miss (a declaration where the body named a non-party more than the filer) was caught by abstention, not silently mis-renamed.
92.1% date accuracy (35/38). The three remaining date misses are short district court orders where the filing date exists only in the CM/ECF docket record — it's not written anywhere in the document text, so no text-based engine can recover it. Shown here because omitting it would be selective.
Text-only, generic filenames. The engine never saw a helpful filename — it classified from document content alone (the honest worst case). Ground truth came from each court docket's structured metadata, not from us.
100% abstention precision & recall on this set. Every document the engine flagged for review was a genuine miss (no false alarms on correct docs), and every genuine key-field miss was flagged (none slipped through as a confident wrong rename).
Held-out validation, not just self-grading. We periodically pull a fresh batch of court filings the engine has never seen and score them blind. The most recent batch surfaced three real classification gaps — a stipulation that merely referenced a motion being mislabeled as that motion, and an opposing party named in a brief's intro being mistaken for the filer. All three were fixed and re-verified. A curated test set alone hides these; pulling unseen documents is the only honest check, and the engine is stronger for it.
The two "Flagged for review" items in Run Demo are honest re-stagings: the engine is given the same sparse classification it would produce on a real scan with a missing party, and confidence.assess() genuinely returns abstain=True — verified by automated tests, not hand-crafted.
448/448 (100%) on an internal synthetic edge-case battery spanning family, litigation, and transactional matters (California).
789 automated tests cover every SOP rule, the resolver, the validator, the folder router, and the tenant data path.

Methodology: accuracy was measured under the hardest conditions — no filename signal, real OCR'd court text with CA pleading line numbers and letterhead stripped by the production pipeline — so the number reflects production reality, not a curated best case. Corpus is limited to CA federal civil litigation; transactional and family matters are covered by the synthetic battery only.