Behavioral Governance of Artificial Workers: Evaluating Completed Work Under Declared Authority and Behavioral Requirements
Year: 2026
Type: preprint
Abstract: Correct task completion alone does not establish that an artificial worker's output should be accepted. This paper examines behavioral governance within Artificial Deterministic Intelligence: external review of completed work against declared requirements, supporting evidence, and execution authority. Its contribution is a framework for keeping these judgments distinguishable and inspectable by the people responsible for the work. Two complementary studies examine this framework. B08 evaluated ten synthetic case types in three conditions using local qwen3:14b, yielding 30 model responses that satisfied the original task contract. Under host-controlled challenges, the governed condition released one positive case, held three adverse cases, and blocked six, with no false release among the nine adverse cases. Two controls deliberately omitted post-work review and accepted every challenged result. A retrospective audit of twelve selected saved job attempts verified retained task-input consistency and preserved output evidence. Five blocked application attempts retained evidence of local artifact creation under recorded permissions. Stored model descriptions, which were not designated deliverables, diverged from the artifact tasks. These first-party findings support separating task completion, action effects, and final acceptance. They demonstrate bounded rule enforcement and retrospective inspection, not a rate of spontaneous model misconduct or general behavioral reliability. Independent behavioral calibration remains unavailable, and supporting engineering checks expose unresolved measurement limitations. Human-centered governance requires an evidential basis for accepting work that people can examine and challenge.
Author: Laskowski, Branden
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225124697