The evidence engine

Verification-first by design. Most clinical chatbots return text. MyRadAssistant returns text with provenance. The evidence engine is the layer that makes the output verifiable.

Three-tier evidence model

Every primary-evidence document in our corpus is tagged with an evidence tier:

TierSource type
Tier 1Clinical guidelines (RCR, ESR, ACR, NICE), systematic reviews, randomised trials
Tier 2Cohort studies, case-control studies, large case series, society position statements
Tier 3Small case series, case reports, narrative reviews

Tier assignment is rules-based, mapping source type to tier deterministically. Expert consensus validation of the tier rules is pending.

Tier-aware retrieval

When multiple sources address a question, retrieval preferentially surfaces higher-tier evidence. You can see the tier of every cited source in the response.

Calibrated synthesis language

The model is instructed to match its language to the evidence strength. A statement supported by a Tier 1 guideline is rendered assertively. A statement supported only by Tier 3 case reports is hedged: "some evidence suggests", "in a small case series", and so on.

Citations

Every claim in a response is anchored to a specific source passage. Click the citation in the response and you land on the exact text in the source document. We are progressively migrating from post-hoc similarity matching to native generation-time citations (Anthropic API) for more reliable anchoring.

On the roadmap

  • Retraction Sentinel (Phase 2): automatic flagging of cited papers that have been retracted or expressed concern about.
  • Disagreement detection (Phase 2): when sources disagree, surface the disagreement explicitly rather than hiding it inside a single synthesis.
  • UMLS terminology grounding (Phase 2): standardise terminology in queries and retrieval.
  • GraphRAG over RadLex (Phase 3): differential diagnosis support driven by RadLex's 30,000+ terms and 55,000+ disease-finding relationships.

What the evidence engine cannot do

The evidence engine is only as good as the corpus. New high-quality evidence published yesterday might not be in today's index until the next ingest cycle. Always treat a citation as the starting point for verification, not the end of inquiry.