D618 - Evidence going missing made a claim stronger

Status: measured Date: 2026-09-08

The limit D617 documented, and what it actually costs

D617 noted, as an aside, that the report directory is overwritten: the probe writes six files into one place and an hour later they are six different files. It called that a limit of the arrangement.

It is worse than a limit. orbistoun-gen measurements rebuilt the table from whatever was on disk, so the previous batch's observations did not merely become unreadable - they were deleted from the committed table by a successful run. 20260908-094705 appears in the table zero times today, and every number D609 drew from it is unreproducible.

The dangerous half is the direction it moves a claim. constant means every run that took this measurement agreed, and it is the flag that decides whether anything may assert a value. A measurement marked non-constant because two batches disagreed becomes constant again the moment one batch is deleted - and the coverage gate then demands it be claimed or declared, and a claim is what it will get.

Evidence disappearing should never make a conclusion safer. Here it did, silently, in a run that reported success.

Measured, by taking the fix out

Regenerating with only ../obscene/data/hardware present - which is what an hourly overwrite looks like from the tool's side:

	measurements in the table
before	429
regenerated without the fold	38
regenerated with it	429

A 91% loss, no error, exit zero.

The change

The committed table is read back into observations and folded in beside whatever the captures say. observations_in_table parses disagreed from the shape table renders it in - value in a.txt and b.txt - so a disagreement survives a regeneration that cannot see either file.

This is the rule the other two generated files have followed for months: write_symbol_db accumulates names and write_wanted accumulates hashes, both citing D074, both for the same reason - each run sees only part of the corpus, and writing only what it saw discards the rest. The measurement table was the one that did not.

Idempotent: two regenerations in a row produce the identical file.

What cannot be recovered

The 20260908-094705 batch. It was already gone when this was found, so the table holds the 143022 observations and the archive's, and not the ones in between. D609's conclusions stand in the decision log; the evidence behind the numbers it quotes does not.

That is the cost of finding this one batch late rather than a reason to be quiet about it.

The guard, made to fail

a_committed_table_is_carried_through_a_regeneration_that_cannot_see_its_captures round-trips a two-row table with one disagreement and asserts the disagreement survives. Dropping the disagreed parse makes it fail on exactly that assertion, which is the one that matters: losing a row is visible, and losing a row's contradiction is not.