MediFact — Evidence-Based AI Argument Mediator

Arguments are asymmetric: both sides walk away remembering that they won. "You said you'd pay." — "I never said that." Nobody remembers — but the receipts, logs, and message threads do. MediFact turns that evidence into an impartial verdict. MediFact is a live AI mediator. It listens to a conversation in real time through AssemblyAI's streaming speech-to-text with speaker diarization, attributing every word to the right person. As the argument unfolds, it extracts each debatable statement and classifies it — fact, claim, assumption, evidence reference, or opinion — updating a live Claim Board while the parties are still talking. When the session ends, the heavier analysis runs. The recorded audio is re-diarized for a canonical transcript, claims are re-extracted with pronouns resolved into standalone assertions, and every uploaded document — expense logs, spreadsheets, chat threads — is "verbalized" into explicit sentences (a CSV row becomes "Alex paid the March rent of $1800 on March 3, 2024"), so who, when, and how much sits on the surface of the text. A single batched LLM call then matches every claim against the full evidence corpus. The core principle: MediFact never declares a winner without evidence. When one party's claims are consistently supported and the other's are contradicted, the report says so, by name. When the evidence doesn't exist, it says that too — and states what would resolve the dispute. Trust is enforced by construction: every quoted passage in the report is mechanically verified as a verbatim substring of the evidence corpus. If the LLM alters a name, date, or amount, the verdict is automatically downgraded. The LLM proposes; code verifies. It works across languages — people argue in Urdu and Hindi while the evidence is in English — and exports a clean PDF report. ~9 LLM calls per session, report in ~15 seconds.