Measuring the Boundary of Deterministic Authorization Monitors

Abstract

Abstract Deterministic authorization monitors promise what probabilistic guardrails cannot: the same input, the same decision, every time. But theory has held for half a century that a monitor’s guarantee extends only to what the information available to it mediates or allows it to infer — and no benchmark, to our knowledge, operationalizes that boundary as a primary comparative variable across concrete monitors. We present an instrument and a method for measuring it. The instrument is a set of 16 cases describing boundary phenomena — shell interpretation, filesystem name resolution, decision-to-execution time windows, downstream decoding, compositional effects — operationalized as test vectors of immutable, hash-frozen bytes. Four safeguards defend the measurement from its own laboratory: a blind-built family of null monitors as an executable admission criterion; immutable vectors that eliminate instantiation bias, which we name and demonstrate; a declared and evidenced evaluation surface with a two-column scorecard (applicability × discrimination); and the separation of ingress from observed context. We measure two open architectures of the same layer under three pre-specified experimental conditions. The results: profiles differ where prior surface analysis predicted — a byte-identical pair of inputs, indistinguishable to any textual decider, is separated only by the monitor that consults the filesystem at decision time; policy choice moves a monitor’s profile by a magnitude comparable to swapping the monitor, making the monitor × policy pair the experimental unit; and configuration shifts boundaries only within the context an architecture already observes — no policy created new sources of observation. All vectors, null monitors, declarations, and measurement programs are released, hash-frozen and re-executable: every published number has a program behind it.