Deterministic Intelligence Orchestration: Treating Neural Inference as a Scarce Residual Resource

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Large language models are commonly deployed as the default computational substrate for tasks that combine retrieval, reasoning, transformation, and linguistic realization. This paper evaluates an alternative architecture in which deterministic computation resolves, validates, narrows, and governs as much of the task as possible before neural inference is admitted. The experimental programme, implemented through Deterministic Intelligence Orchestration (DIO) with BEAST as the execution substrate, progressed from representation reduction and zero-provider deterministic paths (SC1), through relevance-scoped inference (SC2), controlled semantic-compression and linguistic-realization studies (SC3), and a prospective randomized cross-model residual-inference trial (SC4). SC1 reduced an observed 4,163-byte model-facing representation to 570 semantic-model bytes (86.31%) while preserving configured source and authority constraints. SC2 demonstrated zero provider calls for a resolved intent-relevant subgraph and exactly one routed residual claim when inference was required. SC3.3 identified a semantic information floor: aggressive compression reduced neural workload but caused semantic success to collapse. SC3.4 separated truth, communicative act, wording, and release authority, culminating in 20/20 bounded semantic releases in the final governed condition. SC4 prospectively froze 48 held-out cases across six task classes, four architecture arms, and three local model families (576 scheduled observations). The sealed run completed all 576 observations with exactly 288 provider calls and 288 zero-provider deterministic cells; all 288 deterministic cells matched the expected semantic frame. Of the 288 provider-call cells, 102 were pre-specified operational failures and 186 reached blinded semantic evaluation. Routed arms B, C, and D each eliminated provider calls in 96/144 observations (66.7%). Provider exceptions differed sharply by model and architecture, including a temporally concentrated M2 runtime episode and a reproducible D-arm lexical-surface generalization failure. The prospectively specified human-rated primary SC4 endpoint remains pending. Exploratory deterministic recovery of the 186 surviving neural surfaces is reported separately and does not replace that endpoint. Taken together, the programme supports an architecture-level proposition rather than a model-scale proposition: neural inference can be treated as a scarce residual resource, invoked only after deterministic mechanisms have established what is known, what is relevant, what authority exists, and what communicative act is permitted.