Budgeted Verification for Episodic Tool Guards under Incomplete Evidence
Budgeted Verification for Episodic Tool Guards under Incomplete Evidence
Ghasem Azari Arani1 Hamid Mojarad2 Mahdi Esmaeili3
1) Department of Computer Engineering, Faculty of Electrical and Computer Engineering, Technical and Vocational University (TVU), Tehran, Iran. Email:
2) Department of Computer Science, kas.c, Islamic Azad University, Kashan, Iran. Email:
3) Department of Computer Science, kas.c, Islamic Azad University, Kashan, Iran. Email:
IOI Code
:
ABEB-ACAAEZ
Publication :
4th International Conference on Recent Advances in Engineering, Innovation and Technology - Belgium(eitconf.com)
Abstract :
An episodic tool guard may contain correct rules while still lacking enough evidence to authorize a call. Some predicates may be absent from memory; others may be available but unevaluated in the current state. We study how a guard should distinguish these cases and spend a limited verification budget. The method combines an authoritative contract manifest with partial predicate observations. A call is authorized or rejected only when every state consistent with the available evidence supports that decision. Otherwise, the controller queries an available predicate or defers. An exact dynamic program selects queries to maximize the probability of a conclusive decision, using expected verification cost to break ties. We evaluate six policies on 9,504 synthetic evidence configurations drawn from four Boolean contracts, with nine budgets and 513,216 policy traces. At a budget of four units, the exact policy resolves 48.17% of configurations, compared with 46.59% for cost-ordered querying, while reducing mean query cost from 1.452 to 0.804 units. Neither policy authorizes an invalid call in the enumerated domain. An optimistic policy that treats unknown predicates as satisfied authorizes 4,626 invalid configurations. On the complete-memory subset, all querying policies resolve every configuration at the largest budget; the exact policy uses 56.47% less verification cost than eager evaluation. These are component-level results under trusted, static contracts. They establish a reproducible distinction between missing rules, missing observations, and insufficient budgets, without claiming improved language-model task performance.
Keywords : episodic memory tool invocation runtime verification partial observation