System card
What the system is, what it was measured on, and where it fails.
What is behind a dossier: the rules, the models, the data they were built on, how they were evaluated, what they cannot do, and what happens to your document. Written from the same data the product ships, so it cannot say something the product does not.
What decides what
Rules decide wherever a rule exists — legal eligibility of a predicate, the Refuse-to-Accept checklist, exemption screens, the cyber screen.
Models rank and extract. They never issue a verdict on their own.
Where neither a rule nor the record can decide, the dossier asks one question or sends the case to a person, and counts that as a miss on our side.
Models, pinned
- r2 classifier
- FUSION_SERVING_ARGMAX · argmax over the fused channel scores — a rule, not a trained artifact
- r2 encoder
- bge-m3-r2-ft-v2 · 7edf5fcb · 2026-07-11
- r2 meta gate
- r2_metagate_serving.joblib · 05f2cabc · 2026-08-19
- r3 ranker
- lambdarank · f94bd125 · 2026-08-17
- analysis build
- spine-2026-08-10
- rules version
- v1.0-rc3
Data each answer is built on
- Classification
- 175,885 FDA clearances (1976-2026) + the CFR product-code definitions. Structured summaries exist mainly 1996+ (SMDA); accuracy is stated on the obtainable era.
- Predicate
- the set of legally-eligible predicates (pre-dates your device, never recalled, not barred under 513(i)(2)) — not just the most similar device.
- SE matrix
- your intake fields vs the predicate's public summary (a silent cell is declared, never invented).
- Review time
- FDA decision dates, 2019+ only (older data drifts 29d->134d and would mislead).
- Risk
- FDA recalls 2002+ (openFDA) — only 39.7% of product codes have any recall ever.
Evaluation
Each answer’s accuracy is measured on past clearances kept out of training or from after the training period, and published with its count and date on the error-rates page. Product-code confidence is calibrated out-of-fold (1,528 clearances, expected calibration error 0.018, measured 2026-08-11), and the confidence shown is floored to the measured error of its band so it is never better than the truth.
Limits
- It only knows what FDA has published. FDA’s letters to sponsors are confidential, so the likely-questions answer is a proxy from public summaries.
- It does not write test reports, protocols or labeling, and it does not sign.
- Where your document leaves a fact open, it asks rather than guesses. Where the predicate’s record is silent, it says so rather than compares.
- Review times are ranges measured on this device type, not promises.
- Combination products, software-as-a-device, wellness products and banned devices are not yet routed — a dossier for one of these must be taken to a person.
What it learns, and from whom
A correction you make changes this dossier, is kept on its audit record, and — on this browser only — is offered back to you as a suggestion on your later dossiers, marked as yours and never applied on its own.
Nothing learned from one sponsor’s dossier is ever used in another’s: the existence of a submission is confidential (21 CFR 807.95), and so is everything in it.
Your document
- OpenRouter receives the full text to pull out the structured fields — and, if the document is not in English, to translate it. Zero data retention is set on every request.
- OpenAI receives the joined device fields, to find similar predicates. That route has no per-request retention setting, so it rests on the account terms.
- This product keeps only the extracted fields on the audit record. This browser may hold an unfinished draft so a failed run is recoverable; a successful analysis deletes that draft and does not save the source document in the caseload. Predicate ranking runs on our own machine; the FDA, eCFR and Federal Register queries carry no text of yours.