A Technical Specification for Estimating Uncertifiable Input Mass in Deployed Autonomous Systems

This document defines Uncertifiable Input Mass (UIM), a deployment-level audit quantity for autonomous systems operating under finite and explicitly declared verification resources. UIM measures the fraction of deployment inputs for which output correctness cannot be externally certified under a fixed verification boundary. Unlike accuracy, calibration, confidence, or uncertainty metrics, UIM is governed entirely by external audit constraints rather than by internal model behavior, architecture, or training procedure. The specification introduces: a boundary-relative definition of uncertifiable workload mass, an externally auditable, non-circular estimator based solely on deployment logs and admissible external verification procedures, a deployment invariant establishing that retraining, fine-tuning, or scaling alone cannot reduce uncertifiable workload without explicit expansion of the declared verification boundary. The deployment invariant is specified as a compliance and audit condition, not as a learning-theoretic or empirical claim. Under a fixed deployment distribution and fixed verification boundary, any observed reduction in estimated UIM constitutes evidence of boundary expansion, audit inconsistency, or deployment distribution shift, rather than internal performance improvement. This provides a falsifiable mechanism for detecting boundary-incomplete or safety-washing reliability claims in deployed systems. The document is intentionally framed as a technical specification, not a research paper. It makes no claims of empirical discovery, algorithmic improvement, or internal uncertainty detection, and proposes no system architectures, training methods, control policies, or certification regimes. Adoption of the specification is voluntary; however, without explicit accounting of uncertifiable workload mass, deployment-level reliability claims are not externally falsifiable under finite verification resources. This Version 1.0 specification is suitable for: third-party audit and compliance workflows, standards and pre-standard contributions, governance and accountability analysis, policy and regulatory evaluation of AI reliability claims. All definitions, estimators, and invariants are explicitly conditional on the declared deployment distribution and verification boundary. Any future revisions will be issued only as explicitly versioned updates.