Mathematical Boundary Configuration as a Runtime Control Variable for AI Agents A Four Tool Exploratory Ablation of RHIS on WorkBuddyBench

Source abstract

AI-agent benchmarks commonly collapse several questions into one number: whether the agent <br> completed a business task, whether it produced the publicly specified delivery, and whether it <br> matched an evaluator’s exact internal representation. This paper introduces an auditable <br> separation for a lightweight RHIS runtime: business fact topology T_B, public-contract delivery <br> projection P_D(T_B), and evaluator-compatibility projection P_E(T_B). The projections are <br> prohibited from writing back into the fact topology. We evaluate four completed configurations on <br> one WorkBuddyBench Office task using DeepSeek V4 Pro and exactly four tools: Read, Write, Edit, <br> and Bash. The only intended varying policy parameter is a bounded visible-obligation coverage <br> tolerance theta; hard authorization, source, time, role, site, schema, acyclicity, connectivity, and <br> public-delivery-closure constraints remain unchanged. All completed runs had zero admitted <br> noncoverage deviation. Nevertheless, request counts fell from 68 to 13 across theta = 0.00, 0.30, <br> 0.60, and 0.90, while official raw scores were 32.67%, 47.94%, 22.14%, and 21.98%. The request <br> trajectory is a runtime observation recorded before scoring. The score trajectory is not evidence <br> about objective business correctness, because the scorer’s fixed but pre-run non-derivable <br> representation obligations confound that inference. The study establishes the implementability <br> and auditability of mathematical admission objects in an agent runtime; it does not establish a <br> general causal optimum, business-correctness rate, cost-quality improvement, or full-RHIS <br> validation.