Operationalizing DAESG: A Privacy-Preserving, Three-Lines-of-Defence Architecture for Auditable Extraction Telemetry in AI Systems

The Data and Attention ESG (DAESG) framework proposes disclosure, standards, and capital-allocation layers to govern attention and data-commons extraction in AI and platform systems, but leaves three implementation questions open: what unit should extraction be measured in, how can that measurement be disclosed and audited without becoming a new privacy risk in itself, and what assurance model prevents the disclosure regime from degrading into the checkbox-compliance pattern already documented in environmental ESG. This paper proposes answers to all three. First, a taxonomy of three measurement units — tokens, prompts, and time — each matched to a mechanistically distinct governance function rather than treated as interchangeable proxies for “engagement.” Second, a dual-pipeline privacy architecture combining secure aggregation and differential privacy for platform-wide disclosure with structurally unlinked single-session sampling for verification, evaluated against known re-identification attack patterns. Third, a three-lines-of-defence assurance model, extending existing financial-audit practice — internal classification, independent internal reconciliation, external testing — to extraction telemetry specifically, with an external standard-setting body modeled on the Carbon Disclosure Project and a blind-signature token issuer preventing platforms from self-selecting which sessions are auditable. The design is offered as a technical companion to DAESG, addressing the implementation challenges DAESG's own Section 8.1 identifies but does not resolve.