What the instrument reports when nothing was done

Runs: gate/floor.py → gate/floor.json, gate/shamreal.py → gate/shamreal.json Date: 2026-09-22.

What I did

spec/PCIST-1.0.md was written to close the defect gate/RESULT-DUAL.md found, and implementation B was rewritten from that annex alone. Rerunning the dual comparison required replacing one control: version 1 of it asserted that B returns zero on pure noise, and the rewritten B reproduces the reference exactly — so the control started failing because the reference does not return zero on noise. It returned 62.5 on twenty channels of white noise with no evoked response at all.

That is larger than most readings this register has issued from the mouse deposit, so I measured it properly rather than writing it up. Two runs:

• gate/floor.py swept channel count from 8 to 64 against twelve noise seeds under both parameter sets, with controls that a structured response must clear the floor and that the floor must not move when the noise is scaled tenfold.
• gate/shamreal.py then ran the decisive version on tissue: the same six recordings, the same channels, the same trial counts, the same parameter set and the same averaging — with epoch onsets drawn at random from each session instead of from the stimulation times, every draw held at least one second clear of any real stimulation, ten independent sham averages per state-cell.

The white-noise floor is real, steep, and not the instrument's floor on tissue

Under mouse_spes_v1, on white noise:

The mechanism is identifiable. The SNR gate is supposed to prevent this — on noise a component's response power equals its baseline power, so its ratio is about 1 and should fail min_snr = 1.1 — but with more channels the SVD offers more components, some pass by chance, and the maximum over thresholds of NST_resp − k·NST_base then collects positive fluctuation. The floor is amplitude-invariant, as it must be, because the thresholds are set from the data's own distances.

On that basis every mouse reading in the register appeared to sit below its own noise floor. That would have been the most consequential finding the program has produced, and it is wrong. White noise is not brain noise: it is spectrally flat and uncorrelated across channels, so the SVD spreads variance across many components. Real cortical noise is correlated and 1/f, the variance concentrates in few components, and the SNR gate rejects them.

The claim was withdrawn before publication because the control that could kill it was run before the write-up, not after. The white-noise sweep remains on the record as what it is: a property of the computation under an unrepresentative null, and a caution against using synthetic nulls to calibrate a real instrument — the same rule §9.4.1 already states for constructed evidence.

On tissue, the floor is session-specific and the anesthetized arm sits on it

Real against matched sham, all sixteen state-cells, real values reproducing the register to 0.00% and zero clearance violations:

Across all cells the ratio runs 0.36× to 11.65×, median 5.08. Seven of sixteen cells fall below their own session's sham 95th percentile, and one falls below the sham median outright: sub-521887 under isoflurane reads 12.12 against a sham median of 34.05. For that cell the instrument reports less complexity for the real evoked response than for randomly timed epochs from the same animal.

And the floor is not a constant to be looked up. It ranges 2.80 to 34.05 across the six sessions — a factor of 12.2 — on the same deposit, same species, same rig, same parameter set. A population-level floor would be useless here; the floor has to be measured per recording.

What this does to the readings the register has issued

Awake readings clear their own floor comfortably, by 2.3× to 8.3×. Anesthetized readings do not: their median clearance is 1.90× and their minimum is below 1.

That bounds every separation the register publishes. NS-0016 reports mouse wakefulness against isoflurane at 1.89×, and its denominator is an arm that sits close to the floor. A ratio whose denominator is partly floor is not measuring state against state; it is measuring state against the instrument. The separation is therefore a lower bound on a quantity whose denominator is not established, which is a weaker claim than the one currently on the register's face and it belongs there instead.

Nothing here says the anesthetized readings are wrong. It says they are not yet distinguishable from the instrument's own behavior at the same channel count in the same session, and that no published reading in this literature carries the measurement that would settle it.

What I recommend next

Make the matched null a required field. Every perturbational reading should ship with a sham computed from its own recording — same channels, same trial count, same parameter set, randomly timed onsets — and declare the ratio on its face. The machinery now exists and costs one extra pass over data already loaded.

This is not a caveat; it is the thing that makes a single reading interpretable on its own. A benchmark distribution tells a reader what other subjects did. A matched null tells them what this instrument did on this tissue when nothing happened, and the 12.2× session-to-session spread found here is the argument that the second cannot be inferred from the first.

METHOD §2.3 is added to require it. NS-0006, NS-0011 and NS-0016 are not edited — §8 forbids it — but the register records the measured ratios against them, and readings issued from 0.4.1 forward carry their own.