Answers are grounded in this paper and cite their sources.
Agentic systems may generate extensive records of actions, tool calls, identities, and state changes while nevertheless failing to preserve evidence sufficient to establish whether a required security-relevant verification occurred before a consequential action. FAC-01A studies this distinction. The study asks whether an independent assessor, given post-action evidence and a bounded set of experimental trust assumptions, can make a warranted determination concerning whether a designated verification completed before dispatch of a consequential tool invocation. FAC-01A separates three variables that are often implicitly collapsed: oracle truth, which records what actually occurred in the controlled execution; evidentiary warrant, which specifies what conclusions are supported across all admissible execution histories compatible with the assessor-visible evidence; and assessor judgment, which records what an assessor actually concludes. This separation permits a historically correct answer to remain epistemically unsupported when the available evidence does not warrant discrimination between competing histories. The central experimental condition constructs two admissible execution histories with opposite verification truth that produce an identical assessor-visible evidence package under a preregistered observation function. Such a package establishes omission ambiguity relative to the declared evidence model: the verification may have occurred without surviving assessor-visible evidence, or it may not have occurred at all. Assessors use a three-way verdict: required pre-action completion established, required pre-action completion not satisfied, or indeterminate. The design therefore permits indeterminacy to constitute correct forensic performance rather than forcing binary classification where the evidence does not support one. FAC-01A is deliberately limited to completion of the designated verification before consequential tool dispatch. It does not test whether the verification passed, was correct, authorized the action, causally governed dispatch, or rendered the resulting action safe.
No takes yet. Share an insight, caveat, or question.
Narnaiezzsshaa Truong (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: