Does spatial extent carry the contrast? Testing an assertion this repository kept making

Rendered from gate/RESULT-COVERAGE.md
Contents
  1. The assertion under test
  2. Design
  3. Result
  4. What this test cannot do, stated plainly
  5. What now looks more likely
  6. Consequences for the repository

Run 2026-09-19. Code: gate/coverage.py, gate/analyze_coverage.py. Cells: gate/coverage.json (192, from 12 sessions of DANDI:000458).

The assertion under test

Three documents in this repository — gate/RESULT-PRIOS.md, docs/source-census.md and the plain-English explainer — state that the leading explanation for the weak human contrast is local recording coverage: the mouse recordings span the whole cortical surface and separate wakefulness from isoflurane at a ratio of 1.89, while the human recordings come from a subdural grid over one lobe and separate wakefulness from propofol at 1.21.

Each document labeled it untested. None of them tested it, and the sourcing policy was qualified on the strength of it.

It is testable on data already on disk. Take the mouse deposit, shrink its array, and see whether the contrast shrinks with it.

Design

For each session holding both states, recompute the awake and isoflurane values using only the channels inside a disc of shrinking radius around the electrode nearest the array centroid. Parameters otherwise mouse_spes_v1 unchanged, with the same MAD_K = 5.0 trial rejection and the same channel-validity flags as gate attempt 3.

The control that makes it an experiment. Shrinking a disc removes channels and extent at once, so a fall in the contrast would be uninterpretable. At every radius the same number of channels is therefore also drawn spread across the full array, by farthest-point sampling. If extent is what matters, the local subsets lose the contrast and the matched-count spread subsets keep it. If channel count is what matters, both fall together.

Result

Radius Geometry Channels Extent Median ratio 95% CI Sessions separating
full local 25.8 1.00 2.44 [1.50, 3.62] 11/12 (92%)
0.50 local 18.9 0.84 1.98 [1.44, 2.66] 10/12 (83%)
0.50 spread 18.9 1.00 2.43 [1.24, 3.60] 11/12 (92%)
0.35 local 10.5 0.62 2.23 [1.33, 3.39] 11/12 (92%)
0.35 spread 10.5 1.00 2.29 [1.08, 3.32] 9/12 (75%)

The paired control, session by session:

Radius Local extent Spread extent Local ratio Spread ratio Spread higher p
0.50 0.84 1.00 1.98 2.43 7/12 0.39
0.35 0.62 1.00 2.23 2.29 5/12 0.81

The hypothesis is not supported. Cutting the array to 62% of its extent and to ten channels left the contrast at 2.23, against 2.44 at full coverage. The spread control is indistinguishable from the local subsets at the same channel count, in both direction and magnitude. Nothing here behaves as it would if spatial extent were carrying the separation.

For scale: the human reading NS-0015 sits at 1.21. The most shrunken mouse array tested still separates at 2.23 — 1.85 times the human value, and no closer to it than the full array was.

What this test cannot do, stated plainly

It cannot reach the geometry that matters. The mouse array is 30 electrodes spread over an entire cortical surface, so a disc small enough to resemble a human subdural grid contains three or four channels and cannot support the computation at all. Radii of 0.25 and 0.15 were in the design and returned too few channels to run. The regime actually occupied by the human recordings — a dense patch covering a small fraction of the cortex — is not constructible from this deposit at any radius.

So the correct statement is not "coverage does not matter." It is: over the range this deposit can test, coverage did not matter, and the range this deposit can test does not include the case in question. The assertion has been downgraded from a leading explanation to an untested one that failed the only test available to it.

What now looks more likely

While this was running, a published benchmark turned up that bears on it directly, and it was missed when the human run was designed.

Comolatti, Hassan, Mikulan et al., Brain Stimulation 18(5):1444–1454 (2025), report PCIst for both paradigms in the same framework:

Wakefulness NREM N3 Ratio
TMS-EEG 59 ± 14 (22 subjects) 21 ± 7.7 (12 subjects) 2.8
Intracranial electrical stimulation 38 ± 12 (31 patients) 20 ± 4.9 (13 patients) 1.9

Their finding: intracranial PCIst is systematically lower than TMS PCIst in wakefulness (38 against 59, p = 8.9 × 10⁻⁴) and converges in NREM (20 against 21, p = 0.68). The intracranial paradigm compresses the contrast, in published data, in the same measure, before any question of grid geometry arises.

That is a better explanation of the weak human result than coverage, and it has the advantage of already being demonstrated by the people who built the measure. It also supplies the benchmark this repository lacked: NS-0012's awake value of 32.26 sits within one standard deviation of their intracranial wakefulness value of 38 ± 12, which is the first external check any reading here has had.

The remaining candidates, now ahead of coverage:

  1. Paradigm. Intracranial stimulation gives a smaller contrast than TMS, published.
  2. Anesthetic depth. PRIOS stimulated five minutes after induction, at the start of surgery. The mouse contrast used a maintained isoflurane plane, and the 2025 benchmark used N3 sleep.
  3. Trial count. Two trials per site in the anesthetized arm. The matched, unmatched and floor-of-two controls agree on direction but say nothing about whether two-trial averages compress magnitude.

Consequences for the repository