Compilation Is Not Execution

After this lesson, the reader should understand:

• 01Separate compilation (constructing authority) from execution (enforcing it).
• 02State the legitimate roles of a language model on each side of the boundary.
• 03Explain why a more capable model does not make a safer runtime authority.

The model asked to be the rulebook

A team ships an agent that is given the policy in its system prompt and asked to authorise actions against it directly. It is accurate in testing. Why is this the wrong architecture?

Because the authority now lives inside a probabilistic process^[1]. The same case can resolve two ways on two runs; a context change can silently drop a clause; there is no artifact to version, seal, or hand a regulator. Accuracy in testing measures the model, not the architecture. The fix is not a better model — it is to move the policy into a compiled artifact the model calls. The model still earns its place: it reads the messy input, extracts the facts, drafts the explanation. It just does not get to be the thing that says yes. Construction can be probabilistic; enforcement must be deterministic. The strongest objection is that the distinction is dissolving. Set the sampling temperature to zero and a model returns the same answer to the same prompt; evaluate it on a large test set and it is measurably accurate; and it handles the nuance a rule set misses. Three replies. First, zero temperature does not deliver reproducibility in practice: engineers investigating inference systems have shown that results can vary with server load and batching^[2], and the provider can update the model underneath a fixed name. Second, even a perfectly reproducible model leaves nothing to inspect — no rule to cite, no version to seal, nothing a regulator can re-run without the vendor. Third, nuance has a designated place in the architecture: the escalation path, where a person with authority handles what the rules cannot, on the record. The boundary has a cost the architecture must own. If a model extracts the facts the evaluator consumes, the model still shapes the outcome — through the facts rather than the verdict. A misread invoice amount produces a deterministic, reproducible, correctly sealed wrong decision. Moving the model out of the decision does not remove its influence; it concentrates that influence at one inspectable point. Consequential facts therefore need provenance — where each value came from and how it was extracted — and, above some threshold of consequence, confirmation before they cross into evaluation. The combination of learned perception and symbolic decision is an active research programme precisely because getting that hand-off right is hard^[3].

• Is prompting a policy into a model a form of compilation, or a way of skipping it?
• What must a human sign before a compiled artifact may execute?
• Where exactly does the model sit relative to the evaluator in a sound design?
• If a model assists compilation, what review does its output need before sealing?
• If a model extracts the facts a deterministic evaluator consumes, how much of the decision has the model still made?

No responses yet. Be the first to answer one of the questions above.

Draw the boundary for one AI-assisted decision path.

1. 01List every step from raw input to authorised action.
2. 02Mark each step as construction, enforcement, or assistance.
3. 03Identify any step where a model output becomes an outcome without passing through a deterministic evaluator.
4. 04Propose where the compiled artifact sits and what the model does on each side of it.
5. 05For each fact a model extracts, state how a wrong value would be caught before evaluation.