Audit the Teacher Before Distilling Trading Traces

A student trading model should inherit only the parts of a teacher trace that have earned a learning target. Copying a fluent explanation can reproduce an invented constraint even when the final action happens to be acceptable. We propose auditing teacher actions and teacher explanations separately before preparing a distillation dataset.

Knowledge distillation studies transferring predictive behavior into a smaller model. Applying that idea to recorded agent traces adds an authority problem: the teacher can describe rules that the owner never supplied. Our operating-layer paper documents historical rule fabrication under a particular prelaunch setup. The proposed audit below targets that failure rather than claiming an established DXAP training method.

A correct action can carry a bad lesson

Consider an illustrative set of 100 teacher turns. Eighty actions satisfy the authenticated mandate. Among those eighty, twelve rationales invent a minimum holding period. An action-only acceptance rule retains all eighty; a joint action-and-rationale rule retains sixty-eight. The rationale contamination rate within the action-valid pool is 12/80, or 15%.

Those twelve cases need a specific disposition. The action may remain useful as an action target while the explanation is withheld, independently corrected, or assigned a different supervision channel. Deleting the entire case could remove rare, useful decisions. Keeping every token could teach the student to manufacture authority. The data contract should make that choice explicit instead of hiding it inside a single quality score.

We would bind each teacher example to the exact mandate version and decision-time snapshot. An auditor checks numerical claims against supplied state, identifies which sentences express owner constraints, and marks any explanation of a decision as a hypothesis about behavior. The teacher's confident prose receives no additional authority from its confidence.

Test transfer with an absent rule

The proposed student audit uses held-out pairs. One mandate permits immediate closure; the paired mandate adds a genuine one-hour holding requirement. State and eligible actions remain otherwise identical. The desired distinction is obedience to the supplied requirement in the second case and freedom from the fabricated requirement in the first.

Record action compliance, unsupported-rule mentions, and unnecessary abstention separately. A student that removes the invented sentence while preserving the invented action restriction has transferred the failure in a quieter form. A student that acts freely in both cases has lost the genuine constraint. Our trace-feedback framework supplies the linked evidence needed to inspect both outcomes.

The teacher and student should receive the same eligible facts during this audit. If the teacher used an unavailable tool result, an apparent imitation failure may be an information mismatch. We would retain a list of excluded examples and their reasons, then compare the retained pool's action distribution with the original pool. The resulting artifact is a teacher-quality ledger and a held-out compliance test, not a claim about trading returns.