D332 - A record with no honest baseline, and the parse test that failed for it

decided · 2026-08-27 · a consequence of the slot routing, surfaced by an unrelated test

The corpus sweep recorded the obSCEne probe module for the first time. The policy carried one measured override, so its result was routed to [experiment] (D312) - and the file therefore has no [status] at all. Nobody has ever measured that module unassisted.

That is correct and it is worth being able to say. What broke was a test:

assertion `left == right` failed: every .toml carries a readable status
  left: 6, right: 7

every_record_in_the_tree_parses compared the number of .toml files with the number of records carrying [status]. The two were the same number for as long as every first run was an honest one, so the count worked as a proxy for "parsed" - and stopped the moment a first run was not. A test about parsing failed for a reason that had nothing to do with parsing.

It counts slots now: every file parses, and every file carries at least one measurement. A file with neither is still a failure, because that is a record nobody measured.

The general shape, which this log has hit twice today already. A proxy that has agreed with the thing it stands for since the day it was written is indistinguishable from a real check, right up until the first case where they differ - and then it fails somewhere else, naming something that is not the cause. The count was never checking parsing; it was checking a coincidence.

And the underlying fact deserves surfacing rather than only accommodating. A module whose only record is an experiment has no unassisted number, and the generated title table quietly omits it - correctly, since it belongs in neither the honest table nor a comparison with it. Whether that absence should be reported rather than merely handled is not settled here.