How IFFA works, and where it is weak
IFFA is designed to encourage verification rather than demand trust. Nothing below is hidden or hand-waved. No language model runs in the deployed build.
Bias ≠ falsehood
A politically aligned article can be accurate; a neutral-looking one can be wrong.
Coverage asymmetry ≠ falsehood
A blindspot is a fact about coverage, not about whether the story is true.
Source reliability ≠ article truth
A publisher's record does not decide any individual claim.
Official source ≠ automatic truth
A press release is evidence that an institution stated X — not that X is true.
Forum consensus ≠ evidence
Discourse is never counted as independent factual corroboration.
Correlation ≠ editorial motive
Observed coverage describes what was published, not a newsroom's intent.
- Media Landscape Intelligence
Ownership, external ratings, observed alignment, selection/framing/stance, blindspots, the claim evidence matrix, discourse handling — and the six things IFFA never conflates.
- Pipeline methodology
Ingestion, geo-classification, crisis priority — the base pipeline and its weaknesses.
- Editorial priority model
How stories are ranked for prominence. A ranking score — not a probability of truth.
- Event identity
How reports are grouped into one story, and split when they are different events.
- Claim confidence
The frozen claim engine — corroboration, contradiction, primary evidence, corrections.
- Trend model
What is changing and consequential — the eight sub-scores.
- Quality dashboard
Every evaluation IFFA runs, with its numbers — including the honest first-pass figures and known limits.
- Worked examples
Real clusters annotated end-to-end.