Artifact 01

Health-Claim Verification Engine

A retrieval-augmented, multi-model consensus engine that evaluates a factual health claim and makes the limits of its confidence layer visible instead of hiding them.

Problem

Health information tools often collapse weak evidence into a single confident answer.

Role

Designed and built the retrieval, model-distribution, prompt-hardening and interface layers.

Artifact

Run a claim, inspect sources, watch five independent analysts, then read the consensus band.

Limit

Agreement is not calibrated accuracy. The demo is built to make that methodological gap visible.

01 Query expansion awaiting input

02 Retrieval awaiting input

03 Independent analysis awaiting input

04 Consensus awaiting input

How to read this

Confidence is agreement, not truth.

The engine's distinguishing quality is intellectual honesty: it shows agreement and dissent, and expresses confidence as a qualitative range, not a false-precise number.

Correlated voters

The five analysts are different models from different labs, but they share overlapping training data and a common web corpus. They are not five independent experts.

No calibration

Confidence reflects inter-model agreement, not measured accuracy against ground truth. "4 of 5 agree" is not "80% likely true."

Retrieval-bound

The verdict is only as good as the sources retrieved. Garbage sources can still produce confident-wrong output.

Health stakes

For medical claims, a confident-wrong answer is costly. That is why the demo keeps dissent visible and refuses a single probability.

Out of scope: ground-truth-calibrated confidence · medical advice · real-time/streaming · user accounts · persistence.