It prepares the work.
A licensed attorney signs it.
Mross reads the Record, computes the Guidelines by rule, and drafts what gets filed — then hands all of it to counsel, who signs and sends. It is not an AI lawyer, and that is a design decision before it is a legal one: because a licensed human is accountable for every output, Mross can be blunt and unhedged.
Open an account as a law firm, or as someone representing themselves — the two behave differently, and Mross asks which at the door. Creating one takes no payment and starts no subscription. What it costs.
A language model doing arithmetic in prose is the most reliable known way to produce a confidently wrong number that reads beautifully — and an attorney signs their name under these figures. So the Guidelines are encoded as data generated from the official Sentencing Commission PDF, a deterministic engine computes every mechanical step, and the model only supplies inputs and reasons about genuinely contestable questions.
Every step is tagged DERIVED or ASSERTED, and the interface never renders them alike. An attorney's review budget belongs on the second kind.
Mross drafts filing-ready motions, computes exposure, and tells counsel when their theory is wrong. It never signs, files, appears, or holds a relationship with a client. A document it produces says so on its face — in the filename, in the document properties, in a banner above the caption, and in a footer on every page.
Signing inside Mross still produces no signature. That is a recorded review, not counsel's signature on a filed paper.
A real §371 conspiracy count, statutory maximum 60 months. The government's position computes to offence level 31 — a Sentencing Table range of 108–135 months — which §5G1.1(a) then caps at the statutory maximum:
| Total offence level | 31 | Criminal History Category I |
| Sentencing Table | 108–135 | months |
| Guideline sentence | 60 | months — §5G1.1(a) |
At offence level 26 and above, the guideline sentence is always exactly 60 months. Eighteen offence levels are indistinguishable.
So Mross ranks each available objection by the months it actually saves, and says plainly when the answer is none. On this matter all four objections win levels and not one wins a day. That is the difference between a worksheet and advice — and it is arithmetic, not judgement, which is why code does it.
real decided federal sentencings the engine is validated against, from the Commission's own datafiles. 99.98% are either exact on the Sentencing Table or reproduced from a named provision; the 27 it cannot explain are listed rather than hidden.
of federal sentencings covered by the thirteen Chapter Two guidelines it computes. Anything outside them is refused rather than approximated.
third-party runtime dependencies. For a product holding privileged material a dependency is a supply-chain decision rather than a convenience.
This is the shortest way to understand the product, and the list is deliberate rather than aspirational.
- It does not decide a contestable legal question. Whether a prior is a crime of violence, whether a protective order permits disclosure, whether a killing is murder under §1111 — each is put to counsel and blocks the figure until answered. An unresolved element is never read as "does not apply".
- It does not state a figure it did not compute. Where a fact is missing the calculation is marked as a floor, and every surface that would quote it says so instead.
- It does not assert what the Record does not say. Every fact a model reports is checked, word for word, against the page it claims to come from. What cannot be grounded is dropped, counted, and given a reason.
- It does not guess when it cannot read. An unreadable scan is reported as a hole in the Record, never as an empty document.
Named after Mike Ross, who did flawless legal work without a licence — supervised by someone who had one.