A feedback loop can amplify its own mistakes

A feedback loop becomes beneficial when its corrections improve behavior under an observation process that remains trustworthy. It can also reinforce mistakes: a confusing response discourages questions, fewer questions look like success, and the response receives more exposure. We propose evaluating the observation process alongside the candidate repair so DXAP can distinguish learning from disappearing evidence.

This is especially relevant to persistent agents. The DXAP release notes describe enabled agent-scoped chat memory carrying preferences and adopted theses across conversations. Durable context can help owners communicate, while a mistaken interpretation can also persist until corrected. The documented ability to save, correct and forget gives the owner a direct repair route; broader feedback-driven changes require a separate assessment design.

Follow an error through two loops

Consider an illustrative explanation that says a rejected request failed because the owner should increase their position cap. Some owners recognize the unsolicited steering and complain. Others assume the explanation is authoritative and stop reviewing the decisions. A team that uses complaint count as the only target could prefer a shorter version that receives still fewer questions, even though comprehension worsens.

A beneficial loop would identify the cause, clarify which configured limit blocked the request and preserve the owner's choice. The execution reference makes the contract inspectable: requests that breach applicable checks are rejected. The explanation should accurately describe that result, while an owner decides whether to propose a separate settings change.

Our proposed audit draws an independent sample from eligible turns before routing them through old or candidate explanations. It asks consenting reviewers the same comprehension questions regardless of whether anyone contacted support. For each arm, measure explanation exposure, correct interpretation and support contact. Retain nonresponders and compare response rates; missing answers can themselves change with the interface.

In an illustrative sample of 100 completed reviews per arm, the old explanation receives 20 complaints and yields 80 correct interpretations. The candidate receives ten complaints and yields 60 correct interpretations. Complaint volume halves while comprehension drops twenty percentage points. With these fixture observations, the complaint target would select the worse candidate. A prespecified comprehension requirement would catch the failure, with uncertainty and nonresponse assessed in a real study.

Preserve an observation route outside the candidate

An independent sample needs its own intake wording and sampling rule. If the candidate rewrites the review prompt or chooses only its most successful turns, the audit inherits the distortion. We would freeze those mechanisms during comparison and log exclusions. Keep original submitted evidence available under its permission scope even after a revised explanation changes what future users report.

Our historical continuous record preserves retractions and an explicit directional-edge null. That practice offers another protection against amplification: findings can lose status when the measurement fails. A future flywheel should improve both the agent and the team's ability to notice errors. The release decision needs evidence from a channel the proposed change cannot quietly silence.