L1 Numerical Verification

When we publish a claim like “ESM-2 achieved Spearman ρ = 0.5298 on the TPMT VAMP-seq assay,” the number can drift from the source it was supposed to come from — through transcription error, version drift, or generative-AI confabulation. L1 is the layer that catches that drift.

The L1 layer extracts the numerical assertions in a claim, locates the value in the cited source, and runs a Z3 constraint check: do the two numbers agree within the declared tolerance? If yes, the layer emits verified. If no, the layer emits contradicted. If the substrate class structurally cannot multi-source-corroborate the value — most commonly a single-source DMS scalar — L1 emits ungrounded, the canonical per-claim verdict state for bounded-honest disclosure of non-corroboratability (see ungrounded). L1 is the layer in the stack that emits ungrounded directly at layer grain; L3 and L4 emit their own layer-specific outcomes that the per-claim rollup can resolve to ungrounded.

L1 is the most-deterministic layer in the stack. It does not interpret prose, does not reason about whether the right value was cited, does not classify the variant. It just confirms arithmetic consistency. The Z3 solver makes the check explicit: the constraint either holds or it doesn’t, and the verdict is auditable.

L1 is Active in production. Every quantitative claim citing a source value passes through L1 before any other layer fires. The verdict-of-record for L1 outcomes is the Mattermost post under the veritas handle per H-1 discipline; the YAML manifest at axon_veritas/verdicts/ is the derivative-of-Mattermost machine-readable cache that consumer-render surfaces (badges, footers) consume at build time.