2026-09-03 - The conformance matrix, and why the filename is the metadata

tool tests   209  ->  220

data/hardware/ had five files named on four different principles - a scope (ps5-full), a record type (ps5-imports), a finding (crashers), a symbol class (libkernel-vaddrs) - and a third hardware capture arrived with nowhere obvious to put it.

None of those names says how the artefact was launched, which is the axis that changes the answers. The native title-eboot run answered sceKernelLoadStartModule differently from the module-target runs, and nothing in either filename would have predicted it.

The convention (D302)

obscene-probe-<target>-<mode>-<privilege>-<category>.log

obscene-tool matrix reads a directory and parses the names, so the directory is the matrix - a cell nobody added is a cell nobody can forget to pass on a command line. That is the difference from consensus, which takes name=path pairs from the caller.

Category is a fourth axis because D301 measured it being one

The obvious scheme files by privilege alone. D301, from this same day, says that is wrong: privilege tier and application category are independent, and the one governing direct memory and display is category. At the same root privilege, category 0 got full Big App resources while category 65536 got zero bytes of direct memory and sceVideoOutOpen refusing 0x80290001.

One cell for those two runs would make every memory and display measurement read as a disagreement with itself.

Two premises had moved, and both were written down as permanent

consensus.rs opens with "Nothing in this suite carries hardware provenance, and until a console is available nothing will", and builds a substitute oracle from majority agreement. That was true when written and is not now.

So the matrix treats hardware as the authority rather than as a vote: an emulator disagreeing with three other emulators is interesting, an emulator disagreeing with the console is simply wrong. Where no hardware cell exists, disagreement is reported as unsettled and the report says so rather than picking a winner.

diff.rs carries the other one - it refuses cross-target comparison, because two targets run the same checks through different loaders. Right for a regression gate; the matrix exists precisely to make that comparison meaningful by labelling each side with the shape it ran in.

Neither module is replaced. Majority agreement is still right for shapes no console has run.

Gaps come from evidence, not from the cross-product

The full product is 240 cells, almost all meaningless - a raw ELF has no category, nothing runs a mini app as root. So a gap is a shape somebody has already produced that another target lacks. If hardware was captured as ps5/root/bigapp, an emulator that never ran it is a real gap; a shape nobody has run anywhere is not.

Reported because a matrix showing only what it holds looks complete when it is two runs out of twelve.

What it says today, which is honestly nothing

cells (0)

not named as captures (6)
  README.md: not a `.log`
  ps5-full.txt: not a `.log`
  ...

no captures - nothing was compared

A file that does not parse is listed, never skipped. A matrix that quietly omitted one would report agreement it had not checked, which is the failure this suite exists to avoid.

Pointed at the native run under its proper name it reads it, marks it the authority, and finds nothing to compare it against - which is the true state.

Not done, deliberately

ps5-full.txt and ps5-imports.txt are both module-target hardware runs at the same privilege, so they are one cell under this scheme - but they are not the same run twice, they are different suite configurations (one emits import records). The honest fix is for the probe to emit those unconditionally, at which point they collapse for a real reason rather than by discarding a difference. crashers.txt, libkernel-vaddrs.txt and ps5-sprx-manifest.tsv are derived findings rather than runs and want a separate home.

Both move committed evidence, so neither is done here.

And the two older captures carry no OBS|context record at all, only OBS|build - so there is nothing in them to derive a mode or category from. The probe should emit context unconditionally before anything tries to migrate them.

Two things found on the way

The lint gate is red, and was before this. Ten pre-existing errors: doccheck.rs arithmetic, run at 108 lines against a 100 limit, a redundant closure, four missing backticks and three collapsible ifs. matrix.rs is clean; adding one dispatch arm took run from 108 to 109.
obscene-tool overflows its stack on Windows, even for --help. The clap tree across thirty-odd subcommands exceeds the 1 MB default main-thread stack of a Windows debug build; under WSL's 8 MB it is fine. Not new and not caused here, but worth knowing before anybody debugs a Windows run of the tool.

State

cargo test in tool/ - 220 passed, 0 failed. cargo fmt --check clean. The 11 new tests cover name parsing in both directions, the skip-is-not-a-disagreement rule, unsettled-without- authority, and both halves of the gap rule.