RossRecall

How to use it

  1. 1

    Choose a case

    The prepared one, or attach your own filings.

  2. 2

    Draft the motion

    One button. About half a minute.

  3. 3

    Click a sentence

    The filing opens at the words it quotes.

  4. 4

    Decide the amber ones

    Accept or strike, with a reason.

  5. 5

    Sign

    Only once nothing is open.

How it works

Jev is by Typesafe AI. Gemini 3.8 Flash is by Google. One chooses, one writes. That is the point.

  1. 1

    It quotes, it does not paraphrase.

    The judge, Jev, reads every page and picks the passages that prove each point. They are quoted word for word. It cannot write, so it cannot make a fact up.

  2. 2

    The writer writes only the sentences that apply the law to the facts.

    Gemini 3.8 Flash. Those sentences always come to you.

  3. 3

    Every sentence is checked before you see it.

    Is the quote in the filing? Is the case real? Does the source support the sentence?

  4. 4

    Same steps, same order, every time. Nothing signs while a sentence is open.

    It never chooses what to do next, and the lock is not a rule it follows but a check as you sign.

The marks in the margin

Verified

The quote is in the source, and the source supports the sentence.

Blocked

A check failed. Cannot be filed as written.

Needs your review

Nothing failed, but the check was not sure. Your call.

Argument

States no new fact or rule. Nothing to check.

Accepted by you

You kept it. Your reason is recorded.

Struck by you

You struck it. Your reason is recorded.

How often it is wrong, and what it costs

Bad sentences that got through
0 of 19
Could be as high as 17% with this few sentences.
Sound sentences held back
0 of 17
Could be as high as 18%.
Cost of a draft
$0.017
240,000 tokens. $0.07 per million.
The test
36 sentences it had never seen. 17 were sound; 19 each carried one planted error, the kind an AI drafter makes.
The result
Every planted error was caught and no sound sentence was held back. With this few sentences the true miss rate could still be as high as 17%, so the honest reading is "low", not "zero".
The cost
$0.017 a draft: the judge reads about 236,000 tokens at $0.042 per million; the writer reads 2.3,000 and writes 1.5,000 at $0.75 and $3.75. A token is about three quarters of a word.
Why it drafts this way
An assistant that plans its own steps was built too and measured the same way. For a draft of similar size it cost about 18 times more, took about 9 times longer, and came out a different length every run.
What this does not show
Most test sentences were written by the engineer, not a practising lawyer. One case, one court. Middle of eight runs at list prices.

The prepared case is real, from public filings on CourtListener. Not affiliated with anyone in it; not legal advice.