Everything in the web workspace.
The review workbench, the Chat session, and the exports. One account, one session list. Laid out in the order you use it.
A count on every finding.
1, 2 or 3 of 3, computed from agreement every time the list renders, so a finding moves the moment its evidence changes. Tiers come from the count and severity, not from a stored state, so nothing is stuck.
- Three tiers, derived live. Confirmed, Models disagree, Flagged for review.
- Severity and confidence on every card. Blocker, High, Medium or Low. Confidence as strong, moderate or weak.
- Anchored findings. Select a card and the source panel scrolls to the passage it came from.
- What each analysis said. The inspector shows the three positions side by side, labeled A, B and C.
- Three-panel workbench. Source on the left, cards in the middle, inspector on the right. Drag the dividers.
Two actions, applied instantly.
Reject finding removes it from future rounds. Add context gives the models a note or a file they were missing. No confirmation dialogs. A rejection undoes in one click; added context is permanent audit context. Anything you leave alone carries into the next round.
- Steering prompts from round 2. Frame how the models should treat the previous round's disagreements.
- Round narrative and results summary. Each round ends with a short narrative of what changed and a summary you can read in a minute.
- Multi-document sessions. One subject document and up to nine context files. Add more context in later rounds.
Cost in the state bar, always.
Round number, finding counts and running cost stay visible while you work. No round runs unless you start it.
- Runs in the background. Rounds take minutes and run server-side. Close the tab and come back.
- Resumes anywhere. The session is where you left it, on the web or in Word.
- One sign-in, two surfaces. The same account opens the web workspace and the Word add-in.
- Sessions persist. Every session, round and decision is saved. Reopen one months later and the history is there.
Three propose, one moderates.
For decisions without a document, start a Chat session. Three models answer independently, a moderator writes one synthesis, and the synthesis names the model that moderated.
- Raw answers always visible. The three answers sit beside the synthesis, so you can check a claim without leaving the thread.
- Normal or Reviewed. Normal is faster and cheaper. Reviewed adds a critic pass that flags contested claims before the moderator writes.
Three ways out.
Every export carries the count, so whoever reads it next sees how many analyses agreed.
- Download summary. 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. The findings as a list, for a tracker, a ticket or a colleague.
- Rewrite to Word. A rewritten copy of your Word document with the findings worked into the text. Your original file is untouched. Single-document sessions.