2026-09-09 - A label in a candidate list is not provenance

Yesterday's entry ended with a container that did not fit: /system_ex/app/FAKE00000/eboot.bin, reported in two sweep resolutions as a container built by this collection, carrying the five header values a vendor system app carries rather than the ones mkself writes. It was filed as a question rather than acted on, because the tidy reading - "our own builder emits system values" - would have justified changing a builder whose output installs and executes.

Answered within the hour, and the answer is that it was never ours.

The path is a hardcoded candidate in obSCEne's probe: an entry in obs_audit_candidates[] in src/probe/sections/selfaudit.c, put there to check whether a pre-existing third-party container happened to be installed on the test console. The "fake" beside it labels what the probe hoped to find at that path. Nothing in any repository builds a FAKE00000.

So the label was a hope about a path, and two hops downstream it was being reported as a fact about a builder.

Why it was worth a request rather than a shrug

Three checks, all cheap, all in this repository, and any one of them would have stopped the story:

The attribution was convenient - it closed the loop and explained the divergence in one move - and convenience is precisely when a claim gets the least scrutiny. An attribution that happens to be tidy deserves the same check as one that is not.

The consolation, which is better than the puzzle

Something demonstrably not this repository wrote a container that agrees with a vendor system app on all five non-application values. That makes FAKE00000 a second, independent producer of those values rather than a loose end. Weak evidence - nothing here knows what wrote it - but it points at the vendor container rather than at us, which is the direction that matters.

The five rows stay recorded as observed and unexplained. Principle 5 has not moved: what a system container's version 0x10 or attributes 0x32 mean still needs a citable source, and two producers agreeing is not a derivation. (D092)

One more thing the exchange bought

The earlier ambiguity - two selfaudit/source lines with one block of values between them - had a cause worth knowing about, and obSCEne found it: their .obs.log extraction runs awk '!seen[$0]++', which deduplicates identical field lines across containers. Nine containers, most measuring identically, collapsed into one readable block and eight bare source lines. The measurement was never ambiguous; the extraction made it so. Refusing to read the summary and asking for the attribution was the right call, and it is a reminder that a deduplicated log is a lossy one when the same fact is true of several things.