Answers are grounded in this paper and cite their sources.
Theoretical analysis reveals that outcome metrics fail to guarantee computational admissibility in agentic AI, highlighting the necessity of pre-execution restrictions over post-hoc logging.
ML and agentic-AI evaluation certify systems on outcome metrics—accuracy, calibration, task completion—and treat them as evidence of reliable computation. We argue this rests on a substitution: outcome agreement standing in for computational admissibility, the condition that every transition remained within the validity domain 𝑉 on which the system’s transition rule is specified. Two trajectories, one admissible and one not, can produce identical metric values (Observation 1), so terminal-state metrics cannot determine admissibility—a near-definitional consequence of scoring only the terminal state. The substantive claim is what follows when the evaluator is given more. A complete trajectory log determines realized admissibility, whether the transitions that occurred stayed inside 𝑉 . It does not determine structural admissibility, whether the system was incapable of issuing an action outside 𝑉 at each step: that is a counterfactual property of the transition rule at the moment of emission, and a realized trajectory is one sample from that rule, silent on the counterfactual however completely it is recorded. Structural admissibility is therefore produced, if at all, before emission. We state the architectural consequence: restrict which candidates become established actions—the attributable units of behavior—so that a candidate failing its pre-execution precondition never acquires action status, while remaining in a separate, non-attributable non-action record. A pre-action state-space gate illustrates the arrangement where 𝑉 is explicit. The implicit-𝑉 case is left as an open problem. No implementation or deployed-system evaluation is claimed.
No takes yet. Share an insight, caveat, or question.
JULGI KANG (2026) studied this question.