IFCN PRINCIPLE 4 · METHODOLOGY TRANSPARENCY

How We Verify Claims

ANN Verify uses a proprietary 7-Layer AI analysis pipeline — patents filed (KIPO 2026.04.02) — to evaluate claims from public figures, institutions, and viral media. Every verdict is traceable to explicit evidence and human editorial review.

3 Patents Filed · KIPO 2026.04.02 KR 10-2026-0059940 / 0059946 / 0059947 · 69 claims PCT Planned by 2027.04.02

// 01

How We Select Claims

We apply a selection filter to ensure resources are focused on claims that matter to the public. Per IFCN requirements, at least 75% of our fact-checks address claims related to public welfare, health, governance, or widely circulated misinformation.

🗣️
Public Figure Statements
Politicians, executives, officials — claims made in speeches, interviews, or official documents.
📊
Statistical Claims
Figures cited in media, scientific papers, or social media that are unverified or misattributed.
🔥
Viral Misinformation
Claims spreading across platforms related to health, safety, elections, or finance.
🏛️
Institutional Assertions
Claims by governments, NGOs, or international bodies that are contested or disputed.

// 02

The 7-Layer Verification Pipeline

Each claim is processed through seven sequential analysis layers. Layers 1–6 are AI-automated. Layer 7 is the cryptographic integrity seal. A human editor reviews the final output before publication on all high-stakes verdicts.

1
Claim Decomposition (SDE)AI · AUTOMATED

The input is decomposed into atomic, individually verifiable claims, separating factual assertions from opinion, metadata, and framing.

input claim text / article / URL
output claims[], claim_type, language, topic
model Claude (Anthropic) · multi-model expansion planned
2
Source StrategyAI · AUTOMATED

For each claim we plan which sources and search queries are most likely to confirm or refute it, prioritising authoritative and independent sources.

input claims[]
output strategy, sources[], search_queries[], priority
model Claude (Anthropic) · multi-model expansion planned
3
Evidence CollectionREAL-TIME SEARCH

Real-time web evidence is gathered for each claim — facts that support it, facts that contradict it, and the source URLs behind them.

input claims[], search_queries[]
output evidence[] (support / contradict / sources / confidence)
model Claude (Anthropic) + web search · multi-model expansion planned
4
Adversarial ProbeAI · AUTOMATED

A skeptical pass stress-tests each claim for weaknesses, missing context, alternative readings, and misleading framing before a verdict is formed.

input claims[], evidence
output challenges[], overall_skepticism
model Claude (Anthropic) · multi-model expansion planned
5
NLI Trust ScoreSCORING

Each claim is scored for how well the gathered evidence entails it (natural-language inference), producing a calibrated per-claim trust score.

input claim ↔ evidence pairs
output nli scores (entailment / neutral / contradiction)
model Claude (Anthropic) · DeBERTa NLI expansion planned
6
Final VerdictAI · AUTOMATED

All prior layers are synthesised into a single verdict, overall score, grade, and a written rationale citing the supporting and contradicting evidence.

input L1–L5 outputs
output verdict, score, grade, confidence, summary
model Claude (Anthropic) · multi-model expansion planned
7
BISL Hash & Temporal SealCRYPTOGRAPHIC SEAL

A SHA-256 hash of the complete fact-check result is generated via the browser-native Web Crypto API — a tamper-evident seal that makes any post-publication change detectable. On-chain (BNB Chain) anchoring is planned.

method SHA-256 · crypto.subtle.digest() · browser-native
output bisl_hash (hex), timestamp, version_id
note BISL = cryptographic integrity seal · on-chain BNB anchoring planned

// 03

AI + Human Review Structure

ANN Verify is AI-assisted — not fully automated. AI handles evidence retrieval and scoring at scale. Human editors maintain editorial control over final verdicts on sensitive or high-impact topics.

What AI Does
Layers 1–6 analysis, real-time evidence retrieval, scoring, cross-referencing, logical fallacy detection, statistical verification, and BISL hash generation.
What Humans Do
Final editorial review on all verdicts scoring below 50 or flagged as high-stakes. Editors can override AI verdicts, escalate to senior review, and add Editor's Notes.
Editorial Independence Guarantee
All AI-generated analyses are subject to human editorial review before publication. No funder, advertiser, investor, or external party has any influence over the verdict rendered by our editorial process.

// 04

Verdict Scale & Definitions

Every fact-check results in one of six verdict labels, applied consistently across all topics and political positions.

Verdict Score Definition
TRUE 90–100 Accurate and complete. All key elements verified by multiple independent primary sources.
MOSTLY TRUE 75–89 Substantially accurate but omits important context or contains minor inaccuracies that don't change the overall meaning.
MIXED 50–74 Contains both accurate and inaccurate elements. Context determines which parts stand.
MOSTLY FALSE 25–49 Primary claim is inaccurate or exaggerated. A small element may be technically accurate but used out of context.
FALSE 0–24 Directly contradicted by multiple credible, independent primary sources. No element of the core claim holds up.
UNVERIFIED — Insufficient evidence to render a verdict at the time of analysis.

// 05

Evidence Standards

Every source used in a verdict is cited so readers can independently verify our findings.

  • Primary sources are always preferred. Official government publications, peer-reviewed research, and institutional reports take precedence over secondary sources.
  • All significant sources are cited with links. Readers can replicate our research. We do not use sources we cannot publicly link to, except where source safety would be compromised.
  • Date context is mandatory. We note the date of each source and flag when a source predates the claim being evaluated.
  • Real-time retrieval ensures freshness. Live search is performed at analysis time — we do not rely solely on model training data.
  • Conflicting sources are disclosed, not suppressed. If credible sources disagree, we present the disagreement transparently.

// 06

Known Limitations

Transparency about what we cannot do is as important as confidence in what we can.

Honest Limitations
  • AI language models may carry training biases. We mitigate this through human review and multi-source cross-referencing, but cannot guarantee complete neutrality.
  • Claims requiring specialized expertise are escalated to human editorial review, but we do not employ domain-specific experts for every field.
  • Real-time retrieval is limited to publicly available web content. Claims supported only by paywalled research receive an UNVERIFIED verdict.
  • Our analysis reflects evidence available at the time of publication. New information may emerge — we encourage correction requests when this happens.
  • ANN Verify does not evaluate intent or motivation — only factual accuracy.

// 07

API Access

Programmatic access to the ANN Verify engine is in development. The verification result is returned as a structured JSON object — the same shape produced by Layer 6 of the pipeline above:

// sample verdict payload
{
  "verdict": "MOSTLY TRUE",
  "score": 82,
  "grade": "B+",
  "confidence": 0.87,
  "summary": "...",
  "evidence": [ /* support / contradict / sources */ ],
  "bisl_hash": "a1b2c3…",
  "version_id": "v5"
}

For early access, rate limits, or partnership inquiries, contact admin@annverify.ai.

Ready to verify a claim?

No sign-up required. Paste any claim or URL and get results in seconds.

Try ANN Verify →
Privacy · Terms · © 2026 ANN Verify · The Monster Lab