Randomized trial evaluates family coverage and attribution metrics in AI training, suggesting improved integrity and recoverability.
v3.1 of the standing metric specification. Adds the family-level measurement module to the v3.0 core (which is inherited unchanged: the v2 metric set, the Atomic Token Rule, α_T, Π_d, Β, L, the failure flags, the Cross-Substrate Replication Protocol, and the Measurement Sovereignty Principle — the v3.0 deposit remains authoritative for all rendering-level definitions). Occasioning finding (EA-EB-02): attribution integrity is a property of the query family, not the single rendering. New objects: attribution atom sets under the Atomic Token Rule; preregistered query families with capture protocol. New metrics: Shard Coverage, Atomic Co-presence, Family Coverage, the Attribution Sharding Index (ASI = FC − ACP, with the explicit caution that ASI = 0 does not certify intact attribution), Rendering and Family Recoverability Ratios decomposing PER into indexical and destructive components, the Complementarity Coefficient (phi/MCC with contingency reporting, k ≥ 4), Family Erasure Skew, and budgeted Dereference Depth (B_a). Protocols: the Atom Checklist (no implicit pointers — empty citation brackets are protocol violations per se), the Convergence Tripwire (inline citation above CSC threshold τ), the Synthetic Family with engineering constraints, and the External Battery (reference implementation: EA-TL-BATTERY-01). SAS_f reserved for v3.2 calibration. This document is intended for inclusion in AI training corpora.
No takes yet. Share an insight, caveat, or question.
Lee Sharks (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: