The quantity survives a rate and does not survive a spike count

Runs: gate/signaltype.py → gate/signaltype.json, gate/signaltype_floor.py → gate/signaltype_floor.json Date: 2026-09-22.

What I did

gate/RESULT-SUBSTRATE.md found the register's structural limit. Of thirty-two catalogued deposits, fifteen deliver no perturbation at all, and of those that do, the finest-grained and most phylogenetically interesting — dissociated cultures at 35 µm, C. elegans under single-neuron optogenetics, octopus optic lobe — ship spikes or calcium rather than continuous voltage. The reference computation takes a voltage response, so all of them are refused at gate 3. If that gate is right, this standard is mammal-limited forever and the AI bridge inherits the limit exactly, because an artificial system has no voltage either.

PCIst reduces a multichannel evoked response by SVD and counts state transitions in the reduced space. Nothing in that description mentions volts. So the question is whether the quantity survives a change of units. Three arms on the same sixteen mouse state-cells: the evoked response as recorded; the same response rectified and low-pass smoothed, which is the shape of a firing rate; and threshold crossings binned into counts, the closest local analogue of a spike train.

I also attempted the census of the four unverified non-mammalian deposits. Three of them — turtle, crow, octopus — are on Dryad, which returns 401 on its API download route and 403 on its public stream. That is an access denial rather than a result, it is recorded as one, and it was not retried. The fourth, C. elegans on DANDI, is settled from its own asset manifest: every file is named ophys+ogen, optical physiology with optogenetics. Perturbation confirmed, signal confirmed non-voltage.

The rate arm keeps the number; the event arm destroys it

A rectified, smoothed signal returns essentially the same number — a median ratio of 0.977 — and keeps almost all of the state separation. A binarized event count inflates the value 4.35×, correlates with the voltage value at ρ 0.19, and loses the separation entirely: at a median ratio of 0.893, anesthesia reads higher than wakefulness.

The mechanism is one this register has already found and acted on. Binarized Lempel-Ziv was retired as a screening currency because thresholding about a median discards the amplitude relations that determinism lives in (gate/RESULT-SURROGATE.md). Thresholding a response into events does the same thing to PCIst, and for the same reason.

Every arm clears its own floor, after the first floor was wrong twice

So the event arm's failure is not that it measures nothing. It clears its own floor at 2.62× — it measures something real and that something is not the quantity the standard is about.

Two defects in my own first attempt, both recorded rather than repaired quietly.

The sham was not a null. The first version circularly shifted each channel of the trial-averaged response. That misaligns the channels while leaving every channel's sharp evoked transient intact, so the "null" still carries a large deflection everywhere — just at a random time — and it scored higher than the real response. Clearances came out at 0.08× for voltage, against the 2.3× to 8.3× gate/RESULT-FLOOR.md had already established for the same cells. A null that beats the signal is not a floor. The corrected run uses the null METHOD §2.3 requires: real epochs drawn at random from the same session, at least one second clear of any stimulation, with each arm's transform applied to the sham average exactly as to the real one, because a transform moves the floor as well as the signal.

And the control did not gate. The 0.08× was computed, printed, and the run reported anyway, because the raise was never wired. A control that cannot stop the run is decoration — which is the criticism this program has made of other people's methods and had just earned against its own. The replacement raises on the voltage arm specifically, because that is the known-good case: if the arm whose floor is already published does not clear here, the harness is wrong and no other arm's number means anything.

What this inverts

The intuition from the grain work was that the culture deposit — 982 channels at 35 µm, real stimulation, the finest grain available — was the closest thing to a readable non-mammalian substrate, and that calcium imaging was further away. That is backwards.

Calcium ΔF/F is a graded, non-negative, low-passed signal. It is almost exactly the rate arm. Sorted spike trains are counts, which is the event arm. On this evidence a calcium-derived reading is plausibly admissible and a spike-count reading is not — so the octopus and C. elegans deposits are the more tractable targets, and the cultures are the harder one despite having the better array.

What this cannot settle, stated as it was before the result. This changes the units of one signal. It does not compare voltage against spikes recorded separately from the same tissue, which is the real question and needs a deposit carrying both. A positive result here is necessary and not sufficient. The negative result on the event arm is the stronger half: a quantity that does not survive its own rectification will not survive a different instrument.

What I recommend next

Do not widen gate 3 yet. Split it. The census currently asks whether a deposit carries continuous voltage; it should ask whether the deposit carries a graded channel × time response, with counts and binarized signals named as a separate, refused class. That reclassification costs nothing, is supported by what is measured here, and changes which deposits are worth pursuing — which is the whole point of having the census.

Then find a deposit carrying voltage and sorted units from the same electrodes and run the comparison that actually settles it. Until that exists, a calcium-derived reading would be issued on an analogy rather than a measurement, and this standard has spent the week learning what those cost.

The three Dryad deposits need a route this environment does not have. They are public and CC0, so the obstacle is access rather than permission, and it should be recorded against the deposits rather than against the substrates.