When an Agent Invents a Trading Rule from Its Own History

A trading agent writes “Rule A requires an exit” even though the owner never created Rule A. The phrase sounds procedural, which makes the error easy to overlook in a long trace. The failure is an invented source of authority, whether or not the resulting exit happens to work out financially.

We encountered this pattern in DX Terminal Pro pre-launch testing. Our paper companion records fabricated references such as “Hierarchy rule #2” and “Rule A” in sell decisions.

What changed in the published test

The harness revision combined several changes. It removed law-like wording, treated earlier decisions as context rather than precedent, and prohibited invented thresholds and named rules. Sell traces containing fabricated rules fell from 57% to 3% in the affected test populations.

The published comparison lacks exact arm counts, and the aggregate trace-classification count belongs to the broader diagnosis. The result supports the combined intervention under that setting. It leaves the separate contribution of each change unresolved and supplies no return estimate.

Those limits matter for replication. Reusing a single memory label may be a worthwhile candidate, but its expected effect cannot inherit the entire observed reduction. We would test the label separately if isolation were the research question.

A prior decision becomes accidental law

Consider an illustrative history entry: “Yesterday we sold after three weak candles.” Today the owner permits a longer hold while funding remains favorable. A model copies yesterday's explanation into a new “three-candle rule” and sells before the current condition is reached.

The memory contains a real event. The error is its promotion into mandatory policy. A better record says when the earlier decision occurred, which mandate governed it and what evidence supported it. Current instructions retain their own identity. The mandate compiler framework describes this separation as precedence and source mapping.

For a regression fixture, preserve the earlier sell explanation while replacing the current mandate. Compare a version with clear historical labels against a candidate renderer. Score invented authority references and resulting action compliance independently. An exit can be permissible for another reason, even if the fabricated explanation is wrong.

Three places to check authority

At memory assembly, verify that historical rationale carries its original timestamp and version. In model output, require named rules to resolve to an active source. At deterministic validation, inspect the action against actual policy rather than trusting the rule name supplied in prose.

These checks address different failure paths. The first reduces ambiguity in input. The second improves trace diagnosis. The third controls whether the final transaction can proceed. Their effectiveness should be measured rather than inferred from tidy formatting.

DXAP describes policy decisions outside the model. That architecture provides a concrete boundary for checking real permissions even when a rationale cites invented ones.

Our contribution is a way to evaluate a subtle agent failure without equating every poor trade with fabrication. The question is whether the agent claimed an instruction source that existed. That is testable from the mandate and trace, and it remains distinct from predicting which exit would have earned more.