D197 - A consensus across generations is not a consensus, and mixing them destroyed the signal

Four loaders complete the full suite as of today, which made obscene-tool consensus worth running for the first time - it is the substitute oracle for a project with no hardware, and it had never had four complete reports to work with. The first run said:

# implementations shadPS4, PS5PCEM, fpPS4-patched, Kyty-patched
# agreed          142
# disagreed       360

and named PS5PCEM as the lone dissenter 328 times out of 358, 273 of those saying fail.

That reads as a damning result about PS5PCEM and is not one. The sweep builds a module per loader generation, so PS5PCEM ran a GEN=5 module and the other three ran GEN=4. The reports say so themselves: generation|known|5 (agc) against known|both. The comparison was between two different builds of obSCEne, and most of the 360 disagreements were that.

Within a generation, the same tool over the same three GEN=4 reports:

# implementations shadPS4, fpPS4-patched, Kyty-patched
# agreed          469
# disagreed       32

469 of 515. The outliers are balanced too - Kyty 15, shadPS4 8, fpPS4 5 - where the mixed run had put 92% of them on one loader. Nearly every "disagreement" in the first run was an artefact of the comparison, and it pointed hard at a specific project.

The limitation this exposes, which is the real finding

The working oracle is three previous-generation emulators. For the generation this project actually targets there is no cohort at all: PS5PCEM completes, prosper is partial, orbistoun is the sibling project and not independent evidence. So agreement is currently evidence about the previous platform, and the count of implementations is not the thing to look at - their generation is.

consensus already warns that these projects read each other's source and agreement is not four witnesses. This is the same caution one level down: they must at least be answering the same question.

A second oracle, internal to the report

007-responsive and 035-libc reach the same functions by different routes - one asks whether a function reads its arguments at all, the other tests whether it gets the right answer. Across five string functions and three loaders they corroborate on every case, including the one that looks like a miss:

loader	strlen	strchr	strstr
shadPS4	responds / pass	responds / pass	silent / fail
fpPS4	responds / pass	silent / fail	responds / partial
Kyty	silent / fail	silent / fail	silent / fail

fpPS4's strstr responds and only partially passes: implemented, and wrong. That is not the sections disagreeing, it is the distinction they exist to draw, and it needs different work from an absence. Kyty is silent on all five, which is what its fifteen outliers are - it implements essentially no libc string functions, and the patch did not cause that, it made it visible.

Measured the same day: two is not a cohort either

prosper converged to a complete run - 36,550 records, 27 of 27 sections, 478 pass / 7 partial / 18 fail / 12 skip, with only the four semaphore checks skipped. That is the highest pass count of any loader here and it is a GEN=5 run, so it looked like the gap above had closed.

It has not. Consensus over the two current-generation reports:

# implementations PS5PCEM, prosper
# agreed          115
# disagreed       329

and all 329 are SPLIT, none is OUTLIER. With two implementations there is no majority, so the tool cannot name a loser - only report that they differ. An oracle needs three: two gives a diff, three gives a verdict. The three-loader GEN=4 cohort produces 358 outliers, each naming a specific loader on a specific check; the two-loader GEN=5 cohort produces none.

So the count of implementations does matter after all, alongside their generation, and the threshold is three rather than two.

And the second member may not be evidence

prosper's own 005-generation/detect result is partial, reporting generation both - not 5. That is obSCEne's own caveat firing verbatim: "real back-compat, or a stub-everything loader answering for free". prosper links all 35,518 imports as stub slots, so both generations' exclusive symbols resolve for nothing.

Its 478 passes are the highest here partly because returning zero is frequently the right answer, and a loader that answers everything agrees with nothing in particular. The check that says so was written for exactly this and is doing its job; the number to distrust is the flattering one.

Status: derived - three cohorts measured with the same tool over reports produced the same day.