Artificial Deterministic Intelligence: Probabilistic Reasoning, Deterministic Authority

Language models can propose useful actions without being reliable authorities over the systems that execute

them. We define Artificial Deterministic Intelligence (ADI) as an architectural approach in which probabilistic

models provide bounded inference while identity, authority, memory admission, policy, execution rights, state

transitions, verification, and release are governed by deterministic systems external to the model. Determinism

refers to specified control decisions under explicit state and inputs; it does not imply deterministic generation,

correct policies, truthful memory, or universal safety. We present a layered reference model and separate

conceptual requirements from observations in a governed research evidence ledger. In a first-party synthetic

memory benchmark of 30 designs repeated three times per arm, scoped 8B and 35B configurations each met

the recorded acceptance contract on 90/90 jobs, while scoped 4B and broader-context 35B configurations

accepted 72/90 and 46/90; 43 broader-context responses were truncated under the fixed output budget. A

separate eight-task context-admission diagnostic retained PARTIAL overall parity despite matching the

corrected context condition. Historical memory tests provide bounded implementation evidence; worker-

construction performance records with incomplete artifact lineage remain excluded. These observations

motivate, but do not establish, the complete ADI architecture. The proposal synthesizes established work on

reference monitors, capability security, formal methods, runtime assurance, neuro-symbolic systems, and

governed agents. Its contribution is an explicit lifecycle-wide authority contract, an evidence taxonomy, and

falsifiable evaluation questions. Independent replication, comparative evaluation, and validation of the full

enforcement boundary remain open.