The Field Actuation Boundary: From Declared AI Field Changes to Verified Operational Effects
Year: 2026
Type: preprint
Abstract: The deployment of a declared artificial-intelligence configuration is frequently discussed as though specification, activation, effective execution, and causal impact belonged to one event. In layered applications combining language models, retrieved evidence, tools, policy engines, human approval, external state, and multiple execution workers, these are distinct evidentiary claims. A field-oriented design language may name a mission-relative relational constitution, but a named or activated field does not establish that the system consumed the intended configuration, that every relevant receiver was updated, that the change was authorized, or that a behavioral difference was caused by the intervention. This paper introduces the Field Actuation Boundary as a falsifiable operational problem: which observations are sufficient to promote a change from declared intention to verified, authorized, and causally attributable action? A typed witness model separates request, permission, commit, propagation, effect, attribution, and reversal, while permitting non-identifiability and model misspecification as legitimate outcomes. A transparent exploratory pilot evaluates twelve failure families with twelve synthetic variants each, using a deterministic Python harness and an independent scenario-label file. In 144 constructed episodes, both an equally informed conventional control evaluator and a field-oriented relational evaluator classified 144/144 cases correctly. An output-only illustrative comparator classified 24/144; removing worker-version readback from the relational evaluator reduced its score to 108/144. A preliminary unequal-implementation comparison falsely suggested a field advantage and was corrected before reporting, illustrating why baseline audits are part of the method. These results verify a bounded specification and its information dependencies, not an advantage over conventional control, an actual language-model intervention, or the existence of an ontologically distinct field. The paper develops a prospective, independently instrumented protocol for testing operational residual value under matched resources and gives explicit criteria for reducing or rejecting the field abstraction.
Author: Figurelli, Rogério
Language: English
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225345018