[cabibbz]

Eighteen files attached to one packet, and the review referenced none of them. Two causes, both real.

The policies never asked for the path

They required the reviewer to check a claim against an attachment, and to cite a path when the two disagreed. Nothing asked it to name the file otherwise — so a finding about attached code could be written with no path at all, which is exactly what happened.

Added to all four policies:

Attached files are the only code you can actually verify, so use them and show that you did. Cite the path whenever a finding concerns a file that is attached: a statement about attached code that names no path cannot be checked by the operator or acted on by the assistant, and reads as an opinion about code nobody can confirm you read. Where an attached file settles a question, quote the line that settles it. Where the attachments as a whole bear on the decision, say which ones you relied on. You need not discuss a file with no bearing on the decision, but never assert anything about a file whose contents you did not read.

The last clause matters as much as the first: this asks for grounding, not decoration. A file with no bearing on the decision still needn't be mentioned.

The measurement was also wrong

References were matched against the released text. A withheld or unselected candidate releases Bob Regress, so that match could find nothing however thoroughly the reviewer had cited its attachments. It now reads what the reviewer actually wrote.

Ungrounded reviews now look like a problem

None are referenced was a neutral grey note. Files sent and never used are files that did no work, so it now reads in amber and says so:

None of the 18 are referenced anywhere in the review, so the review is not grounded in the code you sent. Either those files do not bear on the decision, or the reviewer did not use them.

FileEvidence carries a level — used, partial, unused, none — and the card is styled from it.

Tests

• Every policy, in both research modes, requires the path, the settling line, and forbids a claim about an unread file
• A withheld candidate is matched against the reviewer's text: Bob Regress alone finds 0, plus the reviewer's own words finds 2
• Levels: all cited → used, some → partial, none → unused, no attachments → none

121 unit tests (120 pass / 1 Windows-only skip), mocked e2e, lint, build clean, browser check at 390×844.

What to expect

The policy change only affects packets sent after you deploy. Reviews already in history keep their existing text — the card will now read them correctly, but it can't add citations that were never written.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
sol-review-gate	Building	Preview, Comment	Jul 29, 2026 2:32pm