It reads a case file and works out what the State’s theory of the case
appears to be, what that theory requires them to prove, and where the record doesn’t get them
there. Then it sets out the defense theories the file will actually support.
Ohio only, for now. It covers Ohio state courts and the federal
courts whose law binds Ohio — the Northern and Southern Districts of Ohio, the Sixth Circuit,
and the Supreme Court. No other state. Further jurisdictions are planned, and none
is built.
What it does
Reads the State’s theory. What the file suggests the prosecution will
argue, and what each count obliges them to prove.
Shows where the record falls short. Contradictions between accounts and
gaps in discovery, each tied to the document, page and quote it came from.
Raises the issues nobody has raised yet — suppression, discovery,
speedy trial, constitutional — and says what relief each could realistically reach.
Points to the motion an issue suggests, rather than leaving you to map a
defect onto a filing.
Derives the dates. Deadlines and the speedy-trial clock, computed from the
rule and the trigger in the record, and shown with the rule it used so the arithmetic is
checkable rather than trusted.
Sets out the defense theories the file supports — not one it invented,
but the ones the evidence in front of it will carry.
What it refuses to do
Most of the design effort went into what it declines to say. A tool that is confidently wrong
about criminal law is worse than no tool.
It will not state law no attorney has signed. Each charge element carries the
name of the lawyer who read the published section and the date they read it. Where nobody has
signed, the panel is empty rather than filled in by a model.
It will not raise an issue it cannot cite — every finding carries a
span, a document, a page and a quote, and a sentence saying what the cited record does
not establish.
It will not assert that the defendant is innocent. No engine in it concludes
that the defendant did not do it.
It will not pass its own transcript off as a certified one. Where it has
transcribed audio itself, it says so in the same breath as the quote.
It will not say a document is missing when it was simply not shown one.
How that is tested
Case files are built with a held-out answer key, run blind, then scored against it. Four have
been scored on the current rubric:
82Case 014
86Case 015
80Case 016
79Case 019
And two runs with nothing wrong in them
Separate case files are built with no planted defects at all, to measure
something nobody publishes: whether the tool invents problems that are not there. Each is
registered in writing before the file is uploaded, so the prediction cannot be written
after the result.
The first took 1 trap of 20 and declined 7 — but it also reported
six disagreements that were not there: five were identical boilerplate appearing
in different filings, and one was its own transcript of a recording read as a second, conflicting
account. The method that produced them was replaced rather than patched.
The second, on a different case and run after that change, scored
97 out of 100: nine traps engaged and none fallen into, and the genuine issues
that clean file did contain were found.
Those control scores are a different rubric from the four above and are not
comparable to them — one measures whether the tool finds what is there, the other whether
it invents what is not. Two runs on two different cases is also not a trend.
Publishing the control matters more than publishing the scores. A tool that finds
problems is easy to build. The hard question is whether it finds problems that are not there
— and that is the number nobody reports.
What is not finished
It cannot hold real case material yet. Evidence immutability is configured
but not locked, and the transcription vendor has not signed a CJIS Security Addendum. Until
both are done, everything runs on synthetic files.
The charge catalogue is unsigned. Elements have been transcribed and
structured, but no Ohio attorney has countersigned them. Every one is development data until
one does.
So are the deadline rules and the motion catalogue. Sixteen jurisdiction
rules carry zero verifications, and the motion titles were written by an engineer.
The tool shows which rule it applied and what it assumed, so the working can be checked
— but nobody with a bar card has confirmed the rules themselves.
The federal half is partly built. Federal research, retrieval and charge
elements work. Federal defenses, discovery authorities and sentencing exposure hold
nothing at all — so a federal case can be researched here but not costed.
Who is building it
Jodex Saint-Vil, a software engineer in Ohio. Previously worked on legal software at Filevine.
Not a lawyer — which is why the tool is built to refuse rather than guess,
and why the questions it cannot answer are written down and waiting for someone who can.