THE
TRUTH
MACHINE.

Mic streaming, Word and PDF uploads, or pasted text — all flow through a four-agent LangGraph pipeline that extracts every verifiable claim, runs a ReAct loop over Tavily, Wikipedia, and PolitiFact, and streams per-speaker verdicts back over SSE. Gemini with Groq auto-fallback. Built for the open internet, $0 infrastructure.

Mic streaming, Word (.docx) and PDF upload, or pasted text — all funnel into the same agent graph. Switch single-speaker vs multi-speaker mic mode at click time.

AssemblyAI diarization labels live audio. Documents and pasted text honour Speaker A: / Speaker B: prefixes. Every claim, verdict, and accuracy score is bound to the person who said it.

Live internet retrieval, never stale — three sources fired in parallel for every search query, then deduplicated, ranked, and compressed before reaching the LLM.

A StateGraph orchestrates four specialised agents — claim extraction, ReAct verification, verdict synthesis, per-speaker report — with conditional routing between them.

Each claim is verified by reasoning over evidence, issuing follow-up search queries when the LLM judges the evidence insufficient, then synthesising a labelled verdict with confidence.

Gemini 2.0 Flash by default. On quota or rate-limit errors the pipeline transparently fails over to Groq Llama 3.3 70B mid-run — no dropped claims, no user-visible error.

Verdicts in the borderline 40–70% confidence band are flagged for human review. Approve or reject inline; decisions persist to the session for the exported report.

A single Server-Sent Events channel streams every pipeline event — stage changes, transcript lines, detected claims, search activity, verdicts, speaker stats — straight to the browser.

One click downloads a self-contained HTML report grouped by speaker, with every verdict sourced and every claim timestamped. Browser-printable to PDF; persistent share link when Supabase is configured.

Built end-to-end on free tiers — AssemblyAI, Gemini, Groq, Tavily, Supabase, Vercel. No paid APIs, no enterprise contracts, no credit card required.

A force-directed graph of every claim made across the session, linked by shared named entities. See which claims hang together — which support the same world-view, which contradict each other — instead of reading them as an undifferentiated list.

Eleven classical patterns flagged inline on each claim. Appeal to authority, false dichotomy, slippery slope, ad hominem, straw man, appeal to fear, cherry-picking, Gish gallop, moving goalposts, appeal to nature, bandwagon. LLM output is clamped to the enum — adversarial input cannot invent a new pattern.

Statistical claims age fast — last month’s unemployment number means something different from 2019’s. Every verdict’s confidence is multiplied by a freshness factor whose half-life is tuned to the claim type. Old statistics fade; historical facts never decay.

Evidence from .gov, .edu, IPCC, WHO, NIH and major newswires is anchored to a fixed credibility tier the LLM is not allowed to override. Removes one full LLM judgement per evidence item and stops a random blog from getting equal weight as primary government data.

Every speaker carries a rolling credibility score that updates across sessions — the more of their claims that hold up, the higher it climbs. Persists in Supabase, so the same speaker keeps their reputation between recordings. The system has memory.