Paper 3: Stop Making Your AI Re-Read Its Employee Handbook

AI inference is where systems spend money, energy, hardware time, and available reasoning capacity. Yet a lot of AI systems begin each task by asking the model to reconstruct the same stable context again: who it is, what it can access, what rules apply, and what job it is supposed to do.

That is a very expensive way to repeatedly introduce someone to themselves.

My new preprint tests a narrower approach: construct stable worker state outside probabilistic inference, then give the model the bounded task packet it actually needs for the assignment.

On a frozen local benchmark using qwen3:14b, all 80 scheduled synthetic tasks met the tested acceptance contract. Compared with full repeated worker prompting, the assembled path showed:

• 96.39% fewer measured input tokens
• 96.22–96.23% less prompt-evaluation time
• 82.78–88.04% lower observed task-wall time

The point is not that every task can use less context, that smaller models are universally interchangeable with larger ones, or that a benchmark percentage automatically becomes a dollar figure in production. Larger-model reasoning still matters when the task calls for it.

The measured systems result is more practical: when stable, governed worker state is resolved outside the model and the assignment fits the resulting bounded packet, the system may stop spending inference on re-explaining the parts it already knows.

Think of it less as “make the model think less” and more as “stop charging the model rent to rediscover the floor plan.”

The paper and public-safe evidence companion are here:

https://doi.org/10.5281/zenodo.23603637

The companion supports aggregate audit while intentionally excluding proprietary worker packages, policy logic, prompts, source code, and runtime architecture.

What is the most ridiculous piece of stable context your AI stack currently sends back into the prompt every single call?