The long-term vision behind Medical Variant Triage is not a machine that pronounces a genome understood. It is a system that carries raw genetic data through every deterministic and evidence-supported step that can be automated safely, then stops where interpretation genuinely begins.
An umbrella design for normalizing genetic variants, collecting deterministic public evidence, applying explicit rules, and assembling reviewer-ready packets before handing uncertain questions to specialists.
The aim. Make broad genetic analysis cheaper, more consistent, and easier for qualified reviewers to inspect.
01The problem behind the project
Much of genetic analysis is repetitive evidence retrieval and normalization. Automating the deterministic portion could reserve scarce expert time for interpretation that actually requires judgment.
That boundary changes the economics of the problem. Normalization, annotation, public evidence retrieval, inheritance-aware filtering, condition-specific workflows, provenance, and report assembly can consume expert time even when they do not require expert intuition at every step.
Researchers, laboratory teams, clinicians, and ultimately people seeking genetic answers may benefit. Because errors could affect health decisions, the system is research infrastructure rather than a diagnostic service.
02How it took shape
The methodology is expressed through a disease-specific FH workflow, a reusable model runner, a manifest-first registry, public evidence adapters, structured outputs, and explicit human-review boundaries.
The methodology is being developed through interoperable parts: the Genome Model Runner, a manifest-first registry, public evidence adapters, structured workflow outputs, and the FH Variant Triage implementation. Each part is meant to expose its inputs, limits, and evidence rather than disappear into a single opaque score.
Josiah defined the deterministic-first thesis, broad-access goal, safety boundary, and system architecture, then directed AI-assisted implementation and documentation.
Working research components can normalize prepared CSV or VCF inputs, query selected public evidence sources, and produce evidence tables, JSON, and reviewer packets.
03What the project means now
The ambition is broad access to genetic analysis, but the safety argument depends on restraint. Deterministic processing should make review cheaper and more complete; unresolved variants, conflicting evidence, and clinical meaning must remain visible instead of being smoothed into false certainty.
Coverage is incomplete, source assertions can conflict, and deterministic processing cannot replace clinical interpretation. No private genomes or personal health records are published.
The most useful automation boundary is not 'AI diagnoses a genome'; it is 'software removes avoidable evidence-processing work while preserving uncertainty and provenance.'
Unify the runner, registry, and disease workflows behind a documented validation ladder and expand only through independently reviewable adapters.