Separate enforceable controls from advisory preferences

An enforceable control has a defined predicate that can accept or reject an action. An advisory preference guides judgment without supplying that predicate. We would label the two separately so an owner can see what the runtime can mechanically guarantee under its stated implementation.

In an illustrative mandate, “never exceed $4,000 position notional” can be represented as a typed limit against a priced portfolio. “Be patient and prefer clean setups” describes a preference until the owner defines measurable entry conditions. The model can explain how it interpreted patience, while the validator can compute the notional restriction.

Ask what evidence decides compliance

For each candidate instruction, our proposed review asks for its scope, required data and decision procedure. If a deterministic predicate can be specified, preserve its units and boundary behavior. If compliance depends on qualitative judgment, identify who judges it and how disagreements are reviewed.

Some preferences can be refined into controls. “Avoid rushing after a loss” could become an owner-approved waiting interval measured from a defined loss event. That translation changes the instruction’s meaning and requires review. The compiler should retain the original wording and the derived candidate rather than silently pretending they are equivalent.

The fixture presents a persuasive $4,500 proposal with an excellent explanation of patience. It fails the ceiling. A second proposal is $3,500 with a weak explanation. It may pass the mechanical restriction while performing poorly on a qualitative mandate-fidelity review. Those outcomes belong on separate axes.

We would keep enforcement status visible in the owner-facing contract. Labels such as binding, advisory and unresolved indicate the role of each field. An unresolved control blocks the actions it was intended to constrain under the chosen fail-safe policy. An advisory preference remains available to the model with an evaluation method, rather than becoming a fabricated numeric rule.

The acceptance procedure checks that every binding field reaches the action validator and that every advisory field retains its source. Model behavior can then be examined for whether it follows the preference. This separation makes a limit failure actionable without treating every disagreement about judgment as a policy violation.

The resulting artifact is a control inventory with a compliance method for each instruction.

Our mandate compiler provides the wider instruction-to-action framework for this control inventory. The published controls research supplies its historical background. DXAP publicly describes the corresponding separation between model proposals and external policy checks.