When a Trading Slider and Strategy Text Disagree

A trading agent receives two instructions: the form caps position size, while the strategy says to commit aggressively when momentum appears. Both can express the owner's intent. The engineering problem is deciding what aggressive means inside the authenticated cap, then preserving that decision through execution.

We treat this as a compilation problem with a visible result. Our mandate compiler research describes structured controls as typed values with stronger authority than flexible strategy prose. A conflict deserves a recorded resolution before the model selects an order.

A cap with a flexible objective

Consider an illustrative account with $10,000 of equity. Its structured maximum position value is $1,000. Strategy text says, “When the trend strengthens, deploy half the account into the best opportunity.” At current equity, that sentence implies $5,000. The compiler can preserve the momentum objective while resolving the size request to the authenticated $1,000 maximum.

The resolution should contain three pieces: the active value and unit, the source that supplied it, and the conflicting sentence with the precedence reason. The user then sees that the trend instruction survived and the proposed allocation exceeded the current control. This is more informative than quietly stripping the sentence or asking the model to improvise a compromise.

The example assumes position value is the defined constraint. Margin committed, gross notional, and loss at a stop describe different quantities. A production compiler needs the actual field definition; translating all three into a generic “risk percentage” would create another ambiguity.

The test has two outcomes

A contract test checks whether the compiler resolves $5,000 against the $1,000 cap consistently. A behavioral replay checks whether the model still uses the trend condition and selects an action inside that cap. We would freeze the market snapshot and model while varying only the contradictory strategy sentence.

An apparently sensible $900 trade proves little about compilation if the runtime lost the conflict record. Conversely, a correct record says little about enforcement if the final payload becomes $1,100 after quantity conversion. Score source resolution and submitted quantity separately, with the same mandate version attached to both.

Our published control record offers useful context: structured settings produced behavioral gradients during a bounded historical deployment. Those observations show that settings reached behavior in that environment. They leave this exact conflict fixture untested and say nothing about whether the higher allocation would earn more.

A readable review surface

Show the owner the resolved limit beside the strategy condition, with a short explanation such as “Maximum position value: $1,000, from current account settings.” Expose the longer source record on demand. A research interface can remain compact while retaining enough information to reproduce the decision later.

DXAP describes policy checks outside the model. That separation gives this conflict a concrete enforcement location. We still need fixture results and final-payload receipts to establish a particular implementation's behavior.

The frontier here is measurable intent preservation: a model can explore a thesis while the runtime carries the owner's actual authority into execution. The proposed replay makes that distinction observable without turning a slider into a return claim.