Finding the Missing Control: Reproducible Discovery of Internal-Control Gaps by Cross-Process Analogy

Abstract

We report a narrow, strongly-verified result: a governed semantic engine that discovers a missing internal control in a family of business processes by cross-process analogy — reproducibly, with a machine-checked repair, and without a single model call. The engine, SEMPL, represents each process as a small list of typed action/7 steps and detects an anomaly against siblings: within a class of processes that share a control, it names the member that lacks it. We evaluate this not on one hand-picked case but as a blind, seeded ablation benchmark: across a fourteen-process accounting network we remove, one at a time, a control that a class shares, and ask the engine to re-discover it. On twelve blind injections spanning two process classes and two control types, the engine attains detection, localization, naming, and repair-verification all equal to 1.000, with zero false positives on the intact network, byte-identical across runs — every bar pre-registered and locked behind continuous-integration guardrails. Against baselines on the identical task, a random detector localizes at chance (0.167) and a count-based statistical shortcut localizes well (0.917) but cannot name a single control and misses the one process that keeps a second control after ablation. A frontier language model (GPT-5.4), asked the same question over the same processes, matches the engine's diagnostic accuracy (localization and naming 1.000, stable across runs) — but its proposed repairs vary run-to-run and are frequently malformed (placeholder or wrong slots, a fabricated step identifier), whereas the engine's repair is sandbox-verified to repair bridge-free and re-gatethe payment. The decisive practical result is amortization: because the engine learns the control from the siblings, it handles all twelve gaps at zero model calls, and when sibling precedent is deliberately made insufficient, one verified repair is internalized so that every future occurrence is detected with no model call. The contribution is therefore not out-accuracy-ing a language model — it is a governed substrate that matches frontier accuracy while adding what a language model cannot: a correctness guarantee, a verified fix, exact reproducibility, and self-eliminating cost. More broadly, the same repair machinery turns language-model-generated workflows — extracted at scale from a company's own business texts, where neither the text nor the model guarantees correctness — into correct, usable, and controllable ones, a natural complement to enterprise agentic systems. We locate the result as a concrete step toward trustworthy, self-improving enterprise AI, and note that the philosophical programme it descends from (Mikhailov, 2026d) is what pointed to this problem as the right one to solve.