SpiralHelios Deterministic Geometric Intelligence (DGI): A Non-Training AI Architecture with Source-Governed ROM, Bounded Observability, and Repeatable Offline Evaluation

SpiralHelios Deterministic Geometric Intelligence (DGI) is a non-training artificial-intelligence architecture designed around deterministic computational identity, source-governed knowledge, explicit capability boundaries, offline execution, and public-safe evidence. This preprint presents a bounded formal DGI model together with a controlled experimental evaluation of a separate source-governed ROM/runtime path. The disclosed mathematical layer includes a fixed 27-position symbolic geometry, a normalized D = 1024 state-space abstraction, classical ternary symbolic control, a protected deterministic evolution map, and the bounded Golden Eigen Index (GEI) observable, while reconstruction-enabling implementation details remain protected. For the empirical study, a dedicated CJCS Evaluation ROM was constructed as a bounded experimental artifact containing 12 governed Canonical Units across mathematics, physics, chemistry, computer science, and engineering. The final frozen evaluation matrix contained 40 queries: 36 supported deterministic-routing cases and four deliberately unsupported negative controls designed to test fail-closed capability behavior. Each frozen query was executed 50 times under the same controlled offline configuration, producing 2,000 total executions. The final verified evidence record reports: • 36/36 supported queries resolved to their intended governed Canonical Units• 4/4 negative controls returned the intended NO_DETERMINISTIC_ROUTE outcome• 40/40 query groups remained stable across repeated execution• 2,000/2,000 scheduled executions completed• 0 execution failures• 0 observed deterministic-field mismatches. The frozen query set, regression outputs, repeatability records, and associated hash records were re-verified using SHA-256. In addition, all 39 artifact paths recorded in the Build #1 output snapshot were rehashed against the retained output set, producing 39/39 matches with zero missing or mismatched artifacts. This retained-artifact verification is reported as an integrity-continuity check and is not presented as a historical Build #2 artifact-by-artifact snapshot. The principal benchmark reports CORE_CALLED=FALSE. Accordingly, the experiment evaluates the source-governed ROM/runtime path and does not claim empirical validation of protected Frozen-Core dynamics, GEI-dependent runtime behavior, universal determinism, unrestricted natural-language competence, comparative superiority, or independent replication. The work is intended as a reproducible, auditable, and public-safe research contribution for deterministic AI architectures in which provenance, version identity, governed knowledge authority, explicit refusal behavior, offline execution, and cryptographic artifact identity are first-class design requirements. The complete protected production implementation and the underlying raw execution artifacts are not publicly released. Public materials are restricted to the scientific manuscript, screened supplementary evidence PDFs, approved diagrams, public-safe execution evidence, and disclosed cryptographic artifact identities. This release is intended to support bounded scientific inspection without enabling reconstruction of protected implementation or internal evidence infrastructure.