D298 - Verification runs against a machine that has learned nothing

decided · 2026-08-26 · --verify reported "0 of our own" on the machine that produced the file

The first --verify against an honest submission found nothing to compare it with. The reason is structural rather than a bug in the comparison: the measurement had already been applied, so the wall it was measured at no longer existed, so the sweep produced no finding, so there was nothing to re-derive.

Left alone that is fatal to the whole idea. Applying a measurement would make it permanently unverifiable, and every later submission would be checked against a machine whose behaviour had already been changed by the answers it was meant to be checking.

So a verifying turn runs in its own data directory - no learned file, no accumulated policy, traces of its own - which is the state the original measurement was taken in. Nothing new was needed for it: ORBISTOUN_DATA_DIR already decides where everything a run reads and writes lives, and the propose tests already point trials at a temporary root for exactly this reason, so a sweep cannot pick up an unrelated run's trace.

The general shape is worth naming, because it will come back. A machine that acts on what it learns cannot check what it learns from the state it is in; it has to be able to ask "what would I have measured knowing nothing?". Anything that accumulates - the learned file today, the name vocabulary already - needs a way to be consulted with the accumulation switched off, or its own output becomes the thing that confirms it.