
AI / Political Technology · 2026
A multi-agent fact-checking pipeline for political speech.
The Problem
Political speech is designed to persuade, not to inform. Javier Milei's economic speeches are a case study in this — dense with statistics, historical claims, and ideological assertions, all delivered with rhetorical force. Fact-checking them manually is slow, inconsistent, and often conflates two different problems: whether something is factually true and whether the argument is logically valid.
A claim can be factually accurate but logically misleading. Standard fact-checkers collapse these into a single verdict, which obscures more than it reveals.
The Solution
Logos is a multi-agent pipeline that separates fact from logic — treating them as two distinct tracks that converge into a final verdict.
① Orchestrator
Parses each speech segment and extracts up to 3 logical claims — labeling each by role (premise, conclusion, rhetorical filler) and domain (economy, inflation, debt, politics). The goal isn't to list sentences, but to map the argument's structure.
② Workers — in parallel
Each claim is fact-checked independently via asyncio.gather. No claim waits for another — latency scales with the longest single call, not the total count. Each worker returns a verdict: TRUE, FALSE, MISLEADING, or UNVERIFIABLE.
②½ Formalizer — in parallel
Simultaneously, each claim is formalized in predicate logic and assessed for validity — independent of whether it's factually true. A claim can be logically valid but empirically false. Keeping these tracks separate makes that distinction explicit.
③ Aggregator
Receives the original segment, all claims with roles, all fact-check verdicts, and all logic formalizations. Synthesizes a final report — flagging fallacies and rhetorical devices that only emerge when claims are read together (e.g. false dichotomy, non-sequitur).
All LLM calls use temperature=0 — same segment in, same analysis out. Logos has no political position. It only analyzes logic and facts.
Slides