KUL
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote
**EvidenceTrail — AI That Shows Its Evidence** AI can extract the right quotation from a document and still reach the wrong conclusion. In hiring, that can mean attributing someone else's experience to an applicant or treating unsupported claims as verified facts. EvidenceTrail is an inspectable evidence-verification layer for hiring documents, powered by Meta Muse Spark 1.3. Instead of scoring or ranking candidates, EvidenceTrail helps recruiters understand what a CV actually establishes. **How it works:** 1. Meta Muse interprets a CV against job requirements and identifies supporting passages. 2. EvidenceTrail independently verifies quotations against source pages, checks who actually performed the work, and identifies unsupported or conflicting evidence. 3. Recruiters inspect highlighted source documents, request clarification, record notes, and export an auditable evidence report. **What makes EvidenceTrail different?** A quotation can be genuine and still support the wrong conclusion. In one recorded test, a CV stated that the applicant's procurement director negotiated supplier agreements. The AI initially classified the passage as conflicting evidence. EvidenceTrail corrected the interpretation because the document did not establish that the applicant personally performed those negotiations. In a controlled evaluation using 12 fictional document pairs and 24 recorded Meta Muse requests, raw model labels matched 19 of 24 author-written reference labels. EvidenceTrail's verified results matched all 24, with 25 of 25 proposed quotations passing source-page verification. These are limited synthetic-test results, not real-world accuracy estimates. EvidenceTrail includes a public, no-login interactive demonstration, source-document inspection, verification receipts, scanned-document support, and human-controlled review workflows. **Our mission:** Make AI-assisted document review transparent, verifiable, and accountable without replacing human judgment. **Meta Muse interprets. EvidenceTrail verifies. Humans decide.**