Methodology v1.0
Exactly how TruthLayer turns a video into an evidence report — published so you can check our work, not just trust it.
Version 1.0 · October 2026 · Changes are versioned and announced 30 days before taking effect.
Our core commitment
TruthLayer is an evidence and media-literacy assistant, not a truth engine. We show what was observed, what was claimed, what was inferred, and what independent sources support or contradict — and we teach you to evaluate it yourself.
1. The evidence pipeline
When you submit a public video URL, it passes through these stages. Every stage produces inspectable artifacts — bounded excerpts, never opaque judgments.
- Download & cache. The video is fetched and cached under a SHA-256 hash of its normalized URL, so repeat analyses are instant and identical.
- Transcription. Speech is transcribed with automatic speech recognition, producing timestamped segments with per-segment confidence.
- Scene detection. Visual scene boundaries are detected; representative frames are sampled at scene changes and claim timestamps.
- Claim extraction. Factual statements made or implied in the transcript are extracted as individual, timestamped claims — separated from opinion, satire, and speculation.
- Source lookup. Each checkable claim is compared against independent sources (see source tiers below).
- Persuasion analysis. Emotional language, framing devices, music-visual congruence, and missing-context signals are detected.
- Report assembly. Everything is composed into the four layers: observed, claimed, verified, persuasion signals — plus an explicit uncertainty note.
2. Source tiers
Not all sources are equal, and we don't pretend they are. Sources are classified into tiers that are shown visually in every report.
| Tier | Definition | Examples |
|---|
| Tier 1 — Primary authority |
Peer-reviewed literature, official government/intergovernmental data, scientific consensus bodies. |
WHO, NIH, IPCC reports, national statistics offices, peer-reviewed journals |
| Tier 2 — Established journalism |
IFCN-certified fact-checkers and established news organizations with disclosed editorial standards and correction policies. |
AP, Reuters, AFP Fact Check, Snopes, PolitiFact, Full Fact, BBC, NPR |
| Tier 3 — Contextual reference |
Encyclopedic and reference sources. Used for background context only — never as sole verification for a claim. |
Wikipedia (definitions only), official organization pages, museum archives |
Explicitly excluded as verification sources: social media posts (including those of public figures) — social media is where claims come from, not what verifies them. A source is never treated as independent merely because it appears in another search result; duplicated reporting is not corroboration.
3. Claim labels
We do not use binary true/false labels. Each claim receives one of these evidence-based assessments:
- Supported — corroborated by independent Tier 1/2 sources.
- Plausible but unproven — consistent with available evidence, but not independently confirmed.
- Unsupported — no supporting evidence found in available sources. Unsupported is not the same as false.
- Misleading — technically accurate elements presented in a way that distorts understanding (missing context, false implication).
- Contradicted — directly contradicted by independent Tier 1/2 sources.
- Could not verify — the claim type or available sources don't permit evaluation (personal claims, future events, private information).
4. Risk scoring
Each analysis receives a 0–100 risk score as a weighted composite, shown with its components — never as a single unexplained number:
| Axis | Weight | What it measures |
|---|
| Harm potential | 40% | Whether claims could cause physical, financial, or psychological harm |
| Factuality risk | 30% | Strength and specificity of claims vs. supporting evidence |
| Persuasion intensity | 20% | Emotional manipulation signals in language, music, and framing |
| Missing context | 10% | Date/location/speaker gaps, repurposed footage, uncited statistics |
Scores map to three bands — Routine (0–33), Worth a closer look (34–66), Be careful (67–100) — always accompanied by the underlying evidence. Teen modes replace numeric scores with descriptive phrases and guiding questions.
5. Uncertainty and limitations
Every report ends with a non-collapsible uncertainty note stating what could and could not be evaluated. Known limitations include:
- Transcription degrades on accented speech, overlapping speakers, heavy background noise, and non-English audio.
- Synthetic media (deepfakes) can produce clean transcripts of things nobody said; visual detection is probabilistic, not proof.
- Edited or spliced clips can present seamless narratives assembled from different contexts.
- Source coverage skews toward English-language, Western outlets; "could not verify" often reflects source gaps, not falsehood.
6. Appeals and audits
- Appeals: any user, educator, or video creator can flag a disagreement on any assessment. Flags receive initial triage within 48 hours; substantiated corrections improve the system and are communicated back. Aggregate appeal-and-overturn rates are published quarterly.
- External audit: an independent third-party audit of labeling behavior, demographic bias, and source diversity is conducted annually, with the full report — including negative findings — published publicly.
- Advisory board: methodology and source-policy changes require approval by an independent advisory board of academics, journalists, and civil-society members. No TruthLayer employees sit on it.
7. Product principles
- Evidence before verdicts.
- Unproven is not the same as false.
- No source is independent merely because it appears in another search result.
- Reactions do not prove causes; timing does not prove causation.
- Context matters as much as the clip.
- High-impact decisions require stronger evidence and human judgment.
- The product should improve user reasoning, not replace it.