Join user intent to actions before scoring outcomes

A useful feedback record connects what the owner meant, what the agent attempted and what the venue executed. Losing any link makes a result easy to misread. We propose an intent-centered join for agent research, with incomplete paths retained as incomplete evidence.

DXAP's chat guide distinguishes discussion from approved settings and confirmed persistent instructions. Its activity guide distinguishes a proposed action from a submitted order and a recorded fill. These are separate events, and their identifiers should preserve those roles in an evaluation dataset.

One instruction, several records

Take an illustrative owner direction: “Reduce this position by half.” The owner confirms an instruction. A later turn sees a position of ten units and proposes a five-unit reduction. Execution records one order, which receives two fills of two and three units. The owner then reports that the reduction worked.

A proposed research join would begin with the confirmed intent identifier and its effective time. The turn would reference that instruction version and the observed position snapshot. The typed action would retain the intended reduction. Order submission and venue fill identifiers would connect the two fills back to that action. The resulting five-unit reduction is reconstructed once, with fill prices and fees stored at their actual granularity.

Joining the owner comment directly to each fill produces two favorable labels from one confirmed intent. Joining only by symbol could attach it to a separate position change on the same market. Joining to current position size could make a later account state look like the state used to compute the reduction. All three errors are avoidable if the linkage follows the action lineage.

The controls paper companion describes historical mandate-to-outcome traces that support this kind of diagnosis. Our extension treats owner intent and subsequent feedback as first-class references around that trace, subject to permission for the research use.

Preserve missing links

Now change the example: only the two-unit fill is available, while the remaining order status is unresolved. The dataset should report observed reduction of two units and unknown final completion. The owner's favorable report remains a perception label. It supplies useful context without filling in the absent venue evidence.

We would produce a join audit with one row per eligible intent. Each row records whether authorization, turn input, action and final outcome can be reconstructed. The report distinguishes unmatched records from records that legitimately have no action, such as an instruction whose conditions have yet to occur. A no-action path can be complete if the saved turn establishes a reason to wait.

An illustrative collection of twenty confirmed intents might contain twelve complete action paths, four supported waits and four unresolved execution paths. The economic denominator would follow the predefined evaluation question. A completion study could count all twenty and report the unresolved cases; a fill-price study would use recorded fills and disclose its narrower population.

This join contract makes feedback more useful without assuming an automatic learning mechanism. It allows a researcher to ask whether a candidate change improves intent fulfillment, execution completion or economic results, then select the right evidence for that question. Each measure keeps its own unit of analysis.