L3 Classification Verification
Also known as: L3, L3 classification, classification verification layer
The third verification layer in the Veritas pipeline. Checks that classification claims — like ACMG/AMP pathogenicity labels — are framed consistently with how the cited evidence supports them.
When a variant is labeled pathogenic, likely pathogenic, benign, or VUS, the label is the output of a classification rubric — ACMG/AMP for germline variants, similar rubrics for somatic and therapeutic-area contexts. The label encodes a Bayesian aggregation across evidence types: population frequency, functional data, computational evidence, clinical observation. L3 checks that the label cited in a claim is framed consistently with the evidence the rubric was applied to.
L3 does not re-classify the variant. That boundary belongs to the cited authority (ClinVar, the curating laboratory, the clinical-genetics service). L3 attests framing-fidelity: does the claim faithfully describe how the rubric reached its label, or does it sharpen, weaken, or recategorize without basis?
Framing drift is a common failure mode in computational variant prediction copy. Calling a tool a “pathogenicity classifier” when it produces a continuous score. Citing PP3 or BP4 evidence weight when the rubric only consumed it as one input among several. Treating a “likely pathogenic” label as functionally equivalent to “pathogenic” in downstream prose. L3 catches these.
L3 emits one of four layer-grain outcomes: verified when claim framing matches rubric framing; contradicted when the claim sharpens or shifts the rubric output without basis; insufficient_evidence when the rubric output cannot be located against the cited source or the rubric encoding does not yet cover the claim’s domain; not_applicable when the claim is not a classification claim at all. L3 is a Bayesian schema-driven gate at the framing-fidelity grain, not a re-derivation of the underlying classification.
These layer-grain outcomes feed into the per-claim rollup that produces the canonical verification_status (one of verified, ungrounded, unverifiable, contradicted, not_assessed). An L3 insufficient_evidence outcome on a substrate class that structurally cannot multi-source-corroborate routes the per-claim rollup to ungrounded; other paths through the rollup are possible depending on what L1 and L4 emit.
L3 emits these outcomes in production today; additional clinical-genetics rubrics route through L3 as their schemas are encoded.