On 6 July 2026 Merlin rejected a final answer, correctly, and told the user almost nothing. The run had committed three files to the user's working tree. The rejection was a three-line stub naming none of the three. I had built the honest half of a refusal and forgotten the useful half.

A rejection in Merlin replaces the model's proposed answer with a status the system writes itself. If the model claims work the record does not support, you should never read the claim. In session bf1e74e9 the claim went away, as designed, and a short status took its place. Three changed files sat in the user's repository, and the status mentioned none of them.

Think about the user's next move. They rerun the request, or they assume nothing happened and move on with three edits nobody told them about. Either way my honesty about the claim became silence about the facts. A refusal with no receipt is a shrug.

I had merged two decisions into one. Suppressing an unsupported claim is correct. Suppressing the record of what happened is not. The model's narrative deserved to go. The ground truth did not.

The same day, every final gained two sections written from the system's own records, never from the model's text. "Changes committed to your working tree" lists every verified commit on every final in commit mode, accepted or rejected. "What verifiably happened this run" gives the task and change counts on every answer the system did not fully accept. The refusal keeps its shape. The receipt comes stapled to the refusal.

The test: take your system's harshest rejection and read nothing else. If the user cannot tell from those lines alone what changed on their machine, the refusal hides facts.

Source: Spec 002, AC-002-002-06, written 6 July 2026. The run's logs have aged out of my log store.