Verification
How the system double-checks its own findings to catch mistakes.
AI analysis is useful but not infallible. The verification step exists to catch mistakes before they reach you.
What gets verified
Only Critical and Major findings go through verification. These are the high-stakes ones - the findings that might lead you to reject a bid or demand revisions. Getting them wrong would be costly.
Minor findings, Notes, and Strengths skip verification. The stakes are lower, and verifying everything would make analysis slower and more expensive with little practical benefit.
How verification works
After the main analysis finishes and all findings are recorded, a separate AI agent reviews each Critical and Major finding. For each one, it:
- Re-reads the evidence - the same quotes from documents that the main agent cited
- Checks the reasoning - does the evidence actually support the claim?
- Considers context - is the finding fair given the full document, or was something taken out of context?
- Makes a judgment - confirm, downgrade, or remove
Possible outcomes
Confirmed - the verifier agrees with the finding. Severity stays as-is. This is the most common outcome.
Downgraded - the verifier thinks it's less severe. A Critical might become Major, or a Major might become Minor. The original severity and the verifier's reasoning are both visible to you.
Removed - the verifier concludes it's a false positive. The finding stays in the results (for transparency) but is marked as such.
Why a separate agent?
Using a different AI model for verification is deliberate. If the same model reviewed its own work, it might make the same mistake twice. A fresh perspective - different model, different pass - catches errors that would otherwise slip through.
It's the same principle as having a colleague review your analysis rather than reviewing it yourself.
What verification doesn't catch
Verification reviews findings the AI already raised. It doesn't find new issues that the main analysis missed. If the main agent never checked a particular requirement, verification can't flag it.
Also, verification works with the uploaded documents. If the real answer to a question is in a document that wasn't uploaded, neither the main analysis nor verification will find it.
You're the final reviewer. Verification reduces errors but doesn't eliminate them. The evidence quotes are there so you can make your own judgment on any finding that matters.