One review session, start to finish.
What happens between upload and export, in the order it happens. About four minutes to read.
- 01
Create a session
Name the session and drop in your files. One document is the subject under review. Everything else is context: a prior draft, a policy, a term sheet. The models may cite context but never raise findings against it. Up to 10 files. PDF, DOCX, TXT, MD or RTF.
What you seeThe New review session dialog with a file drop zone and the objective field.
- 02
Set the objective and depth
Write the review objective in one sentence so all three models review the same task. For example: "Flag indemnity and liability terms that depart from a mid-market SaaS baseline." Pick a depth: Quick, Standard or Deep. Longer material and deeper passes take more time and cost more.
What you seeThe objective field filled in, the depth selector, and the helper text "Use a specific objective so all three models review the same task."
- 03
Run round 1
Claude, GPT and Gemini receive the same material and the same objective at the same moment. Each works alone. If fewer than two finish, the round stops rather than presenting one model's output as a comparison.
What you seeA progress state with three model chips and the state bar reading Round 1, Dispatching, then Analyzing.
- 04
The moderator merges
A moderator model groups overlapping findings into one list and writes a short round narrative. It works from the analyses labeled A, B and C in place of provider names. Each finding keeps the position of every analysis that raised it.
What you seeThe round narrative panel, two or three sentences long.
- 05
Read the count
Every finding shows how many of the 3 analyses found it. 1 of 3 says pause. 2 of 3 says think. 3 of 3 says move. Confirmed holds findings all three raised. Models disagree holds findings where at least one took a different position. Flagged for review holds high-severity or weak-signal findings that deserve a look regardless of agreement. Every card also shows severity (Blocker, High, Medium or Low), confidence, and the passage it came from.
What you seeThe card list with three tier headers, the agreement meter on every card, the source panel scrolled to the selected finding's passage.
- 06
Decide, with two actions
Open a finding to read what each analysis said. Reject it if it is wrong or irrelevant; it drops out of future rounds. Add context if the models are missing something: a sentence, or a supporting file. Both apply instantly. A rejection undoes in one click; added context stays on the record as audit context. Anything you leave alone carries forward. Not acting is deferring.
What you seeThe inspector drawer with Analysis A, B and C side by side and the two action buttons.
- 07
Run another round
Round 2 sees round 1: the surviving findings, your added context, and the previous counts. You can add a steering prompt that frames how the models should treat the contested set. Repeat until the contested set is small enough to decide.
What you seeThe state bar reading Round 2 with the finding count and the running cost.
On cost
Every round shows its cost in the state bar. Cost depends on document length, depth and the model tier in use. There is no hidden spend and no round runs without you starting it.
- 08
Export
Download summary gives you a decision-first HTML report that prints to PDF, grouped by severity with agreement counts, in a fast-scan section and a working-review section. Export findings gives you the list for a tracker or a colleague. Rewrite returns a rewritten copy of your Word document with the findings worked into the text, for single-document sessions; the original is untouched.
What you seeThe Export menu with its three items and a page of the printed summary.