Mathine — Evidence In, Trust Out

From Prompt to Field Engineering: Govern the Context, Not the Output

Mathine: Governed Context Field Machine
Link: https://doi.org/10.5281/zenodo.18882422

Context-conditioned AI systems are commonly “governed” at the output layer through filters, prompt rules, and policy text, while the conditioning substrate remains underspecified. The result is a persistent audit gap: reviewers can inspect the answer, but cannot reconstruct what actually executed—what tools were invoked, which retrieval snapshot was used, and which constraints were active.

This paper makes an architectural claim: stable behavior under drift requires governing context as a first-class object, not styling outputs after the fact. It introduces Field Engineering, a receiver-first approach that treats context as a Field F = ⟨D, G, C, R⟩: domain scope (D), operational regime (G), an explicit contract with falsifiers and limits (C), and replayable receipts (R). In this view, prompts are dials inside an admissible space; they are not the governance boundary.

A practical blueprint follows: three planes that separate concerns cleanly. The Meaning Plane handles semantics and task interpretation. The Control Plane enforces what is allowed and how decisions are promoted. The Evidence Plane makes the run auditable by binding execution to receipts—so the system can be reviewed for what it did, not just what it said.

The paper also defines admissible transforms T(F): declared re-descriptions or environment changes under which stability claims are evaluated. This blocks “silent semantic drift” disguised as prompt improvement. It introduces station tests as replay runs in independent environments to validate receipt completeness, contract satisfaction, and stability under T(F), and adopts a fail-closed promotion discipline: absent receipts and station-test stability, promotion should not occur.

Finally, it argues that reliable governance is dominated by worst-slice behavior rather than average-case scores, motivating explicit worst-slice metrics and drift corridors. The framework stays model-agnostic and applies across LLM, RAG, and tool-augmented stacks, with operational templates for contract spines, receipt spines, transform registries, and station-test harnesses to support implementation.