We introduce Governed Deduction (GD) to separate premise use based on relevance from that based on authorization. The key formalism is the transition‑local admission predicate admit(p, τ, S), which states whether premise p may be used in transition τ under state S. Using an RBAC‑augmented Spider benchmark, we generate 4,461 matched authorization pairs where the same query premise and policy state support both permitted and denied consuming transitions.
In experiments, a joint linear controller attains 99.19% held‑out accuracy, while a transition‑only control reaches 100%, revealing a role‑name shortcut. After applying a frozen, label‑independent, context‑local role permutation that removes this shortcut, premise/state‑only, transition‑only, and joint linear controllers all drop to exactly 50% on 1,856 held‑out edges, whereas a symbolic policy oracle remains at 100%.
This controlled negative finding shows that, although the benchmark instantiates policy‑grounded authorization beyond relevance, the frozen linear representation fails to recover the relation. Consequently, matched one‑sided controls and leakage audits are essential for evaluating learned policy‑sensitive reasoning.
Review