How Mross handles your material
Written for the diligence a firm owes under ABA Model Rule 1.6(c) and 5.3 before letting a vendor near privileged material. Every figure below is read from this installation's running configuration, not from marketing copy — if the operator changes a setting, this page changes with it.
Where the material goes| Model provider | OpenRouter — a ROUTER, so the text also reaches an upstream provider under its own terms — this is the subprocessor that sees the text of documents submitted for extraction, and nothing else does. |
|---|---|
| Model | anthropic/claude-haiku-4.5. Your competence duty includes reassessing when this changes, so it is stated here rather than described. |
| OCR | Runs on this server, with ocrmypdf and tesseract, or not at all — a host without them refuses OCR rather than sending the scan elsewhere. Either way a scanned document leaves this machine to be read by nobody. |
| Transcription | No speech-to-text is bundled. A recording is stored and reported as a hole in the Record until a provider is configured, and none is configured here. |
| Research sources | When counsel runs subject research, the subject's name is sent to the public-records source being searched — CourtListener, the Bureau of Prisons inmate locator and SEC EDGAR full-text search are queried directly; the other ten sources raise a task for a human and Mross sends them nothing. Those three see a name and a query, never a document. This is a search, not a disclosure of the file, but it is a third party receiving a name and it belongs on this list. |
| Analytics, error reporting, fonts | None. No analytics, no error-reporting service, no CDN and no third-party font host — the fonts are served from this origin precisely because a webfont request would tell a third party which attorney is reading which matter, at what time. The Content-Security-Policy on every page is default-src 'none' with no third-party origin, and the application has zero third-party runtime dependencies. |
| The Record and the work product | Until an attorney destroys the matter. Discovery, findings, positions, work product and the deadline register are the case; nothing about them expires on a timer. |
|---|---|
| The colloquy | 90 days, then removed automatically. Deliberately short: a durable archive of the questions counsel asks about their client is subpoena surface, and in criminal defence the adversary is the government, which has subpoena power over vendors. The matter export carries it so counsel can keep their own copy. |
| The audit trail | Kept. It is the accountability record — who reviewed what, who signed, who lifted a "no AI" instruction — and it is what answers a bar complaint or a malpractice claim years later. Deleting it to reduce our subpoena surface would trade your defence for ours. |
| Destruction on demand | An attorney can destroy a matter and everything behind it — rows, stored bytes and that matter's own audit trail — for a protective order's return-or-destroy term. It exports first and it cannot be undone. |
| A client who says no | Model processing can be prohibited per matter. With it in force no path sends that defendant's material to a model, and every refusal is recorded, so a firm can show what was not sent. |
- We do not train on your material. Every request to a router carries
provider: { data_collection: 'deny' }, which refuses any upstream that retains or trains on prompts. - Mross never signs, files, appears, or communicates with your client. Recording a review inside Mross produces no signature, and the .docx says so on its face.
- Mross does not learn law from users. A correction to the law is escalated for the jurisdiction pack to be fixed at the source; it never becomes a preference that quietly changes later answers.
There is no signed zero-retention term with a model provider. The code path is built and the request refuses training, but a contract is not something code can produce, and a firm doing this diligence should ask for one before putting real discovery in. There is also no bundled speech-to-text, and backups are on-host only unless the operator has supplied an off-host destination. Payments run through Stripe Checkout, on Stripe's own page — card details are entered there and reach this server never — and are inactive unless the operator has connected a merchant account. These are stated here for the same reason the rest of this page exists.