Deterministic Worker Construction Before Inference: Measuring Repeated State Serialization and Its Task-Dependent Limits
Description
Inference is where an artificial intelligence system spends its operating budget. In hosted deployments, repeated input and output can create metered provider charges; in local deployments, inference consumes energy, hardware time, memory bandwidth, and available model capacity. Yet language models are often asked to reconstruct a worker’s role, permissions, procedures, task context, and operating constraints from a large prompt on every invocation.
This preprint evaluates a narrower alternative: construct stable worker state before inference, retain governing state outside the model, and supply the bounded task packet required for the assignment. A frozen local B06 benchmark evaluated 20 matched synthetic tasks in each of four conditions: cold and prewarmed execution, each with full repeated worker prompting and a HOPE assembled worker path. All 80 scheduled tasks were scored correct on local qwen3:14b. Relative to full reprompting, the assembled path used 96.39% fewer measured input tokens, reduced prompt evaluation by 96.22% cold and 96.23% prewarmed, and reduced observed task-wall time by 82.78% cold and 88.04% prewarmed.
The result is bounded: it does not establish universal model equivalence, universal quality preservation, direct financial or energy savings, or independent replication. The accompanying public-safe evidence companion permits aggregate audit without releasing proprietary worker packages, policy logic, raw prompts, source code, or runtime architecture.