Transparency Reports

Why this exists

State-of-the-art models still hallucinate — they produce fluent, plausible-sounding health claims that are subtly wrong, unsupported, or invented outright, and present each one with the same confidence. In a clinical context the reader cannot tell a well-evidenced claim from a confident guess — and the cost of being wrong is asymmetric.

This project exists to shed light on that gap. Every claim here is put through Diadia’s proprietary claim-verification process — decomposed into its underlying mechanism, checked against the published evidence link by link, and rolled up into an auditable verdict. The aim is not another opaque answer to trust, but clarity: the reasoning, the evidence, and the uncertainty, all in the open.

A library of decomposed claims

Each report is built around one living figure: an interactive mechanism graph whose nodes are biomedical entities — biomarkers, processes, symptoms, outcomes — and whose edges are causal or evidential links. Links are encoded by evidence state through line weight, dash and glyph, never by colour alone, so every figure reads in greyscale and in print.

The catalogue spans nutrition, sleep, immunity, metabolism and mood. Some claims hold; some break; most land in the honest middle — plausible, not proven.

How every claim is checked

The same four steps run on every claim, so reports stay comparable across the library:

1. DecomposeThe claim is broken into mechanism steps and drawn as a neutral hypothesis graph — every link an equal hairline.
2. Map evidenceStudies reweigh each link. Some thicken to established, some hold moderately, some break, some are left untested.
3. EnrichEvidence may reveal explanatory routes the claim never mentioned, surfaced as parallel paths.
4. Roll upSurviving paths combine into a single auditable verdict — traceable back to every edge and source.