Framework
The pipeline has three phases. Information flows in one direction: nothing from the biological or review phases is fed back into the technical ranking.
- Phase A · TechnicalCandidate assemblyUnion of calls from multiple somatic callers on full cardiac WES; the caller of origin is recorded for every candidate.
- Phase A · TechnicalTechnical feature extractionMultiple technical criteria such as read support, allele fraction, depth, caller concordance and quality indicators.
- Phase A · TechnicalLearned technical rankingA learned model scores candidates from technical features only. Candidate model families will be compared; the final architecture is not yet frozen.
- BoundaryFixed Top-K shortlistThe protocol will set K before comparison. The order is frozen before any biological annotation is attached.
- Phase B · Downstream evidenceBiological and contextual annotationFunctional impact, expression, multiome context, conflicting evidence, uncertainty and provenance, shown alongside the rank and never used to change it.
- Phase C · Review · prototypeGrounded explanation and auditA language model summarizes only the attached evidence, refers back to its source fields and flags unsupported statements. It cannot reorder candidates or establish truth.
- Phase C · ReviewExpert evidence reportOne traceable report per shortlisted candidate; the decision remains with the expert reviewer.
Design principles
- Separation: technical selection and biological interpretation are distinct, auditable stages.
- Provenance: every value in a report can be traced to its source and processing step.
- Conflicts and uncertainty are visible: contradicting evidence is listed, not averaged away.
- Context, not causation: multiome signals support interpretation but are not treated as causal proof.
- Bounded language model: the explanation layer is read-only with respect to ranking and outputs.