AGENTIC AI : THE COMPILER AND THE SIGNER. On SCOPE, the Eight Mechanisms It Independently Arrived At, and the Half It Named as Open

SCOPE compiles a natural-language policy into a program of narrow predicates answeredby weak classifiers, deterministic extractors, and Python composition code. It is trained adversarially against the policy, not against a sample. Its verdicts are stable, cheap, and traceable to the predicates that produced them. Eight of its mechanisms have specifications in a public standard: narrow predicates answered by weak models are the action library (A2); composition code with loops, branches, and counts is the extended FSM (A1) and the plan library (A8); deterministic extractors are ternary arithmetic with no float in the hot path; temperature 0 at the predicate is determinism at the tile; the verdict trace is the persistent memory (A7) and the trace ledger (A16); adversarial fitting is the Verifiable Agency Theorem (A10) applied to training; and the policy as the object to fit is the basis vector itself. Two teams, in different domains, arrived at the same architecture. That is not a coincidence.It is what the problem requires. And SCOPE named the half it does not solve. It compiles the policy. It does not sign it. Thepolicy is a natural-language artifact — written by someone, for some purpose, in somecontext. Compiling it into predicates solves the composition problem. It does not solve theprovenance problem: where the policy came from, who signed it, and whether it is correct. That is the trust-root residual, named in White Paper V2 §12.2. The compiler makes the policy executable. The signer makes it checkable. Both are needed.Only one of them is a compiler.