The Living Data Genome: Ledger-Governed Evidence Lifecycles for Accountable AI Under Limited Observability

Artificial intelligence is increasingly deployed where direct observation, reliable ground truth, and repeatable experimentation are structurally unavailable, and pipelines resting on an externally verifiable reference fail by design. This work introduces the Living Data Genome, an architecture treating data as lifecycle-managed evidence. Internal representations remain local; external artifacts cross an externalization boundary only when an admissibility gate and an immune screening condition jointly hold, giving accountability without raw-data exposure. Evidence is anchored by four causal genes and evolves through genesis, mutation, and repair. After release, apoptosis withdraws institutional validity by certified revocation, an append-only lineage remains auditable, and regeneration re-issues a corrected artifact without erasure. A protocol reliability score is defined without labels, combining triadic alignment, transition stability, admissibility survival, and governance coherence under a weakest-link rule proved in Supplementary Section S6; it measures whether evidence passed through the declared lifecycle as required, not whether the underlying claim is factually correct. In synthetic scenarios, under the declared protocol and measured governance coherence, it reached 0.789 on post-incident red-light violation reconstruction and 0.785 on a healthcare vignette, which was a second parameterization of the same generator rather than an independent domain; each of these values was the mean over fifty generator draws. A supervised baseline predicting repair-stage admissibility from the genesis state alone reached a balanced accuracy of 0.595 and 0.553 in the two configurations; hence, the correction is not predetermined. The score moved under the four failure modes it targeted and remained within a declared bound under three it did not. Certified revocation and regeneration recover most of the reliability lost to silent discard.