Autonomy–Opacity Stratification: A Falsifiable Theory of Performance-Driven Functional Opacity in Advanced AI Systems

Abstract

Functional opacity can arise as a by-product of performance-driven delegation. This paper develops three falsifiable hypotheses. H1 predicts a complexity-dependent reduction in the value of detailed process intervention and held-out formation of a persistent opaque delegated-autonomy regime. A conditional comparative-statics result states when a negative interaction implies a lower optimal intervention level. H2 predicts higher fixed-horizon incidence of endogenous evaluation, selection pressure and persistent evaluative-layer differentiation in prospectively defined regimes than in matched controls. Opacity is an observable regime marker, not an assumed sole mediator. H3 predicts an onset order from opacity to evaluation, selection and layer differentiation, followed by endogenous observer-relative exposure adaptation where independently specified reactivity, capability and incentive conditions apply. The measurement program separates onset from confirmation, handles pre-landmark events and competing termination, and prevents outcome-based regime selection. Admission timing, local closure, monitoring reallocation and epistemic depth remain modular derivatives. A separate governance proposal conditionally permits semantic compression within a tested envelope of authority, effects and revocation; H1-H3 support would not itself establish safety. The paper reports theoretical conditions and test designs, not empirical confirmation. Note on Version 2.0: this version revises the registered v1.0 (about 15,800 to 20,300 words, 44 pages). It adds a testable-contribution and competing-explanation section, an external-evidence and transfer-gap section, a joint-pattern test for the core predictions, and external calibration tests of the gap. No empirical confirmation is claimed. Files: the v2.0 manuscript and a supporting archive containing the consolidated review note, a case-verification script with its logical-case results, an external evidence ledger, an edit history, a final validation record, a manifest and SHA-256 checksums. The hypothesis is isolated from the author's broader Information Ecosystem Theory, which remains unpublished at the time of this release; this paper is written to stand alone.