01
Preserve the claim
Split the answer into atomic statements, retaining the conditions and exceptions that affect meaning.
Following the evidence…
Following the evidence…
Second Opinion brings claims, policy passages, and reviewer judgment into one place. Here’s what the system checks—and what it leaves to you.
Split the answer into atomic statements, retaining the conditions and exceptions that affect meaning.
Validate source identities, selected documents, page locations, and exact quoted text in code.
Hold the answer and claims fixed. Reassess them against the remaining selected policies.
Inspect the passage, record a note, and mark what you have reviewed. Interpretation still needs care.
A matching quote proves that the words occur in a source. It does not prove that the source applies, that the interpretation is correct, or that the packet is complete.
Try the distinction yourselfThese papers inform the product. We do not claim to invent a new factuality algorithm or reproduce their published scores.
Decompose a long answer so each statement can be evaluated separately.
EMNLP 202302Human annotations expose hallucinations in retrieval-augmented responses.
ACL 202403Treat answer quality and citation quality as separate questions.
EMNLP 202304Study how changing retrieved evidence affects answers and attribution.
Research paperYour draft, documents, notes, and reports remain in application memory as you navigate. Refreshing or clearing removes them. There is no saved document history.
English selectable-text PDF, UTF-8 TXT, or pasted policy text. Up to 5 documents, 20 total pages, 5 MB per file, and 60,000 source characters. Scanned PDFs need OCR before upload.
Use public or synthetic, non-sensitive policies in this pilot. Live audits transmit extracted text to Gemini. Google’s unpaid tier may use inputs and outputs for improvement and human review.
Read Gemini’s service termsNo. Sample findings are authored for the exact fictional packet. Editing those inputs requires a configured live audit. Samples are labelled throughout the app and exports.
No. “Reviewed” records a human workflow action. It does not change the model’s judgment or certify correctness. Reviewer notes stay separate from audit findings.
Live quality has not been established by the authored demo. The project includes evaluation fixtures and a runner; real results must be measured using configured model access. No accuracy percentage is claimed.
Start with a ready-made example, then follow the evidence.