⚡ FME Version 19.1 · May 2026

Forensic Media Evaluation
Methodology

Every Rhetoric Audit scan is grounded in a four-stage forensic pipeline — from span-level propaganda detection to external claim grounding. This page explains what we measure, how we measure it, and where the limits are.

Principles

Four commitments that shape every scan

Forensic, not editorial

We score what is structurally present — framing, evidence quality, omissions, emotional load — not whether we agree with the conclusion.

Multi-axis, not left/right

A single ideology label flattens complex argumentation. FME decomposes text into independent rhetorical dimensions that can be reasoned about separately.

Span-anchored evidence

Every score traces back to the exact words that produced it. "Why this 72?" is always answerable with a quoted passage and a scholarly technique name.

Reproducible & auditable

Every scan returns a prompt_hash and fme_version. Same inputs produce equivalent outputs. Bench F1 is published per release.

Capabilities

What FME V19 measures

V19 ships seven capabilities in P1, with argument mining and cognitive bias schema-reserved for P2. The table below shows the full traceability from scholarly literature to current status.

#	Capability	Scholarly Source	Status
1	Span-level propaganda detection	Da San Martino et al., SemEval-2020 Task 11	P1 ✓
2	Aristotelian appeals	Aristotle's Rhetoric; LEPAn	P1 ✓
3	Full fallacy enumeration	SemEval persuasion + classical logic	P1 ✓
4	Emotion model + arc	Plutchik-8; VAD model	P1 ✓
5	Full-document analysis	Eliminates V18.6 truncation bias	P1 ✓
6	External claim grounding	Google Fact Check Tools, Wikidata	P1 ✓
6b	Cross-platform corroboration (Stage 1.6)	Real-time X, News, Web signals — blended into FGI	P1 ✓ NEW
7	Ground-truth validation pipeline	SemEval-2020 public set	P1 ✓
8	Argument mining	Stab & Gurevych (2017)	P2 reserved
9	Cognitive bias / nudge detection	Kahneman, Thaler	P2 reserved
10	Cross-source framing	Media Frames Corpus	P3 — CSCE
11	Strategic silence	Agenda-setting theory	P3 — corpus
12	Discourse structure	Rhetorical Structure Theory	P3 — parser

Pipeline

The four-stage forensic pipeline

V19 replaces the V18.6 monolithic single-pass LLM prompt with a deterministic five-stage architecture. The LLM observes; code scores.

Stage 0

Preprocessing

Paragraph-aware chunking into 3-paragraph windows with 1-paragraph overlap. Eliminates V18.6 truncation bias.

Deterministic · No LLM

Stage 1

Span Annotation (batched, parallel)

Each chunk is sent as one LLM call. Output per span: char offsets, technique, appeal, emotion, confidence, and rationale.

LLM · gpt-4o-mini

Stage 1.5

Claim Grounding (parallel with 1.6)

Factual claim spans are canonicalized and queried against Google Fact Check Tools and Wikidata. Stage 2 is never blocked.

External APIs · 30-day cache

Stage 1.6

Cross-Platform Corroboration (parallel with 1.5)

Article title triggers real-time signal retrieval from X, News, and Web via the Intelligence Brief edge function. Relevant signals are scored for entity overlap and recency, then blended into the Factual Grounding Index (FGI = Stage 1.5 × 70% + cross-platform × 30% when ≥2 signals found). Fail-open — never blocks Stage 2.

Real-time · X · News · Web

Stage 2

Aggregation (deterministic)

Document scores rolled up from paragraph scores using a prevalence-weighted, paragraph-length-normalized formula. FGI incorporates Stage 1.6 cross-platform blend when coverage is sufficient.

Deterministic · No LLM

Stage 3

Validation

Zod schema check, span offset sanity, and bench-score CI gate (merge blocked on >2pp macro-F1 regression).

Schema · CI gate

See FME V19 applied to a real article

Install the Chrome extension or sign up to run your first forensic analysis. Free credits included on signup.

Forensic Media EvaluationMethodology