The register was already comparing across grain, and the rule written to stop it was wrong

Run: gate/pitch.py → gate/pitch.json Date: 2026-09-21. The tissue half of METHOD §3.5.

Question or issue resolved

METHOD §3.5 was added this morning on the strength of gate/RESULT-REPLACEMENT.md, which measured — on a constructed system — that an exact functional replacement of a system's units moves every screened measure to between 0.53× and 1.79× of its value. §9.4.1 forbids converting that into a claim about tissue, so the section was written conservatively and one question was left unasked:

Does recording pitch actually vary across the readings this register has already issued? If every human reading sits on the same clinical grid, §3.5 is a rule with nothing to bite on. The electrode tables are small, public and authoritative, so this is a cheap question to answer before spending a decimation run on signal.

Thirty-nine subjects across the five human stimulation deposits behind the register's supply, read from BIDS electrodes.tsv, with units taken from coordsystem.json rather than assumed. Pitch is the median nearest-neighbour distance within a contact group — a patient with a grid and two depth shafts has three pitches, and pooling all contacts gives none of them.

The controls establish that the coordinates mean what they appear to mean

The third was a real failure first. A flat S3 listing reported two deposits as carrying no electrode tables at all. They carry them: ds004774 stores its signal in MEF3, one file per channel per segment, so fourteen subjects produce tens of thousands of keys and the paging stopped before reaching the later subjects. The run reported a listing limit as a finding — the exact failure the standing instruction to expand sources rather than accept one source's limit is there to prevent. The listing now descends by prefix, and a deposit returning no tables is a hard control failure rather than a row of zeros.

Finding: pitch varies by 3.1× across the register's own supply

Across all 39 subjects pitch runs 3.22 mm to 10.05 mm, a factor of 3.1. ds005448 splits cleanly into two populations inside one deposit: five grid subjects at 9.00–10.05 mm and eight depth subjects at 3.22–4.87 mm. §3.3 already separates those by geometry. What §3.3 does not separate is ds004696 at 3.50 mm from ds004774 at 9.43 mm, or either from the rest.

Finding: the register's one external check is a cross-grain comparison

Corrected 2026-09-21 by gate/RESULT-DECIMATION.md. This section leaned harder than the evidence supported, and on the wrong quantity. It compares measured nearest-neighbour distance; §3.5 was revised in the same hour to compare nominal pitch, and on the nominal figure both deposits are clinical grids at 10 mm — the comparison is within-grain. The tissue measurement then put the effect at a 1.3× pitch ratio at 8% in mouse and undetectable in one human subject, against an 11% difference between the two readings. The external check is not compromised. What survives is the units defect in ds004370, below.

The README describes NS-0026 as "the first external check on any absolute value here" — 35.83 against NS-0012's 32.26, a different cohort at a different center, inside the issued interval. Both readings are ecog_grid. Both passed every check the standard had.

NS-0026 was computed on three of the STReEF grid subjects, whose pitch measures 9.00 to 10.05 mm. NS-0012 was computed on seven PRIOS subjects, whose pitch measures 5.86 to 7.85 mm. The two readings reported as agreement were taken at grain differing by roughly 1.3×, and nothing in the standard as it stood could see it, because §3.3 assigns both the same geometry and grain had no field.

Worse for NS-0012 specifically: ds004370 declares no coordinate units at all. Its grid contacts measure 5.86–7.85 mm where the same manufacturer's clinical arrays measure 9.00–10.05 mm in the deposits that do declare millimeters. Either its coordinates are in a different space — surface projection onto a smoothed pial surface compresses distances — or they are not millimeters. Under INSTRUCTION.md 3.6 that is the difference between a pitch that can be established and one that cannot, and it falls on the deposit behind three issued readings.

Finding: §3.5's comparison rule, as written this morning, is wrong

§3.5 said that any difference in declared pitch is a difference in grain, with no threshold of triviality, borrowing the construction §4 uses for parameter departure. On this data that rule fails, and not marginally: the seven PRIOS subjects pooled into NS-0012 span 5.86 to 7.85 mm, so under the rule as written they are seven separate readings and NS-0012 cannot exist. The same is true of nearly every multi-subject reading in the register.

The error is identifiable. §4's construction is sound because a parameter is an exact declared number — a resample rate is 500 Hz or it is not. Pitch as computed here is a measured quantity carrying coordinate-estimation error, anatomical curvature and projection choices. A zero-tolerance rule on a measured quantity is not conservative; it is undecidable in practice, and a rule nothing can satisfy protects nothing.

The repair, and it comes out of this data rather than out of taste: the declared pitch is the nominal center-to-center spacing the array was manufactured at, which is exact and comes from a specification. The measured nearest-neighbour distance is evidence bearing on it, not the thing itself. Subjects combine where the nominal pitch matches — exact, decidable, unaffected by coordinate noise — and the measured value is checked against the nominal one to catch a deposit whose coordinates do not mean what they appear to.

The tolerance for that check is measured here, not invented: across the deposits that declare millimeters, clinical grid groups measure 9.00 to 10.05 mm against a nominal 10 mm, so measured values sit within 10% of nominal when the coordinates are sound. A 15% band admits every sound deposit in this census and excludes ds004370 at 21–41% below nominal. That is the threshold §3.5 now carries, marked as measured with its basis named.

Conclusion

The rule was right to exist and wrong in its first form, and both halves were established within hours of writing it — the first by asking whether it had anything to bite on, the second by the answer.

It has a great deal to bite on. Pitch across this register's human supply varies by 3.1×, two deposits treated as interchangeable supply differ by 2.7×, and the single cross-cohort agreement the README leads with was taken across a 1.3× grain difference that the standard could not see. None of that was hidden; it simply had no field, and a quantity with no field is a quantity nobody checks.

METHOD §3.5 is revised to nominal pitch with a measured 15% consistency band, and INSTRUCTION.md questions 3.6 and 3.7 are rewritten to match. What still has not been measured is how much a reading actually moves per unit of pitch on tissue — the constructed factors of gate/RESULT-REPLACEMENT.md cannot answer it and §9.4.1 forbids pretending otherwise. That is the decimation run, and it is now the only thing standing between §3.5 and a fully evidence-backed comparison rule.