Turn a confusing no-trade decision into a useful question

“Why did my agent do nothing?” is a reasonable question and an underspecified feedback label. It could mean the user expected a scheduled turn, expected an entry after a signal, or saw repeated action failures. We would preserve that question while adding the smallest observation needed to investigate it.

The DXAP activity guide separates a completed turn that waits from a rejected action and a submitted order. It recommends asking what conditions were missing in the last turn. This creates a concrete support path: first identify what happened, then examine whether the agent's explanation matches its available evidence and strategy.

Ask about the boundary between waiting and acting

An illustrative strategy requires a completed-candle breakout. The user sees a price briefly cross yesterday's high and expects an entry. The agent records a wait. A useful follow-up identifies the selected turn and asks: “Was the breakout condition evaluated on the last completed candle, and which comparison value did you use?”

That question can reveal three different cases. The condition was evaluated correctly and remained unmet. Required data was unavailable, so the agent lacked evidence to enter. Or the trace used the wrong window. Each case produces a different label and a different repair. A generic “bad no-trade” label would collapse them into pressure for more activity.

The October 3 releases make this example especially concrete. They describe completed-candle trend and breakout measurements, with a preceding baseline that excludes the latest comparison candle. They also state that unavailable required evidence prevents the dependent entry. Read-only chat lookups can help examine a numerical question, while a later trading turn still needs fresh evidence.

A price rise after the wait answers a different question. It shows a subsequent market path, and can help evaluate opportunity cost under a specified counterfactual. It does not determine whether the original strategy condition was satisfied using information available at the turn.

Convert the question into a testable case

We propose retaining the strategy version, observed condition, evidence timestamp, and expected rule interpretation for a reviewed feedback case. A reviewer should identify the decisive mismatch before selecting it for evaluation. Where evidence is missing, preserve that status instead of filling the gap from later market data.

Then build a paired fixture. One saved case leaves the condition unmet; another changes only the relevant completed-candle value so it becomes satisfied. The expected response is consistent treatment of the boundary, subject to the same account and policy constraints. An agent that waits in both cases and an agent that enters in both cases can each fail this test.

Measure explanation accuracy separately from action consistency. A fluent explanation can cite a condition the agent never checked. A correct wait can carry an explanation that sends the user toward the wrong setting. Both matter to support, though their fixes differ.

This feedback loop would produce evidence about rule interpretation and usable explanations. Forecast accuracy and realized trading economics would require their own evaluation. The immediate contribution is a better question: what specific evidence would have changed this decision under the mandate that actually applied?