Claude, GPT and Gemini read the same document, blind to each other. Every finding shows how many of the 3 agreed. Agreement lets you move faster; disagreement tells you where to slow down.
Governing law clause conflicts with the venue clause
Mediumweak
1 of 3
Flagged for reviewAny count
Data processing addendum referenced but not attached
Highweak
3 of 3
Why three
One model gives you one confident answer. Three give you a count.
Ask one model to review a contract and you get one confident answer. It sounds just as sure when it is wrong as when it is right. So you either trust all of it or check all of it yourself, and neither is review.
Ask three independently and you learn what one answer cannot tell you: how many found the same thing. That count is what lets you spend your attention where it matters. Crossfire is built around it.
ABC
How it works
A review session in three moves.
01
Upload and set the objective
One subject document, any supporting context, and a one-sentence brief. PDF, DOCX, TXT, MD or RTF.
02
Run a round
Claude, GPT and Gemini review independently and in parallel. A moderator merges overlapping findings, working from the analyses labeled A, B and C in place of provider names.
03
Move on what agrees. Decide what does not.
Findings arrive in three tiers: Confirmed, Models disagree, Flagged for review. Reject or add context. Anything untouched rides into the next round.
PauseOne analysis raised it and two did not. Read it before you believe it.
ThinkTwo agree, one took a different position. Open it and decide for yourself.
MoveAll three found it. The strongest signal a round can give you. Skim it and keep moving.
Confirmed3 of 3
All 3 analyses found it. Skim it and keep moving.
Models disagree1 or 2 of 3
Slow down here: open it, read what each analysis said, and decide for yourself.
What you do with a finding
Do nothingDefault
The finding carries into the next round unchanged. Not acting is deferring, not accepting.
Reject finding1 click
It is wrong or does not matter. It drops out of every later round. Undo in 1 click.
Add contextA note or a file
The models missed something. Give them a sentence or a document, and the next round sees it. It stays on the record as audit context.
Flagged for review
Agreement is not proof.
Three models can agree and still be wrong. They learn from much of the same material, so they can share a blind spot. They can also agree on a finding that rests on thin evidence in your document.
Crossfire treats agreement as a reason to move faster, not a reason to stop looking. A finding that is high severity gets a person's eyes even at 3 of 3. So does a finding with weak signal, where the models agreed but the evidence is thin.
That is the Flagged for review tier. It is small by design and it is never hidden. The count tells you how far to trust a finding. It never makes the call for you.
Built for people who are accountable for the answer
Blind to brand
The moderator works from Analysis A, B and C in place of provider names.
Cost on every round
The workspace and the add-in show what each round cost before you run the next one.
Your document stays yours
Nothing is edited without you. Exports are an HTML summary, a findings list, or a rewritten copy of your Word file. The original stays as you uploaded it.
Analysis, not advice
Crossfire organizes what three models found. The call is still yours.
Microsoft Word
Findings land where you already work, as Word comments
Add the Crossfire tab to Word. Run a round on the open document and read each finding as a native comment anchored to the exact text. Depth control: Quick, Standard, Deep.
Ask a question. Get three answers and one synthesis.
For decisions without a document, start a Chat session. Three models propose, one moderates and writes a synthesis, and the three raw answers stay visible beside it. Choose Reviewed to add a critic pass that flags contested claims before the synthesis is written.
Review the next document with three models instead of one.
Crossfire is in early access. Tell us what you review and we will set you up.