The instrument does not port to an artificial substrate, and the controls say why

Run 2026-09-20. Code: gate/artificial.py. Output: gate/artificial.json (70 rows), gate/artificial_rejection.json.

Controls failed. No reading is issued from this run and none may be.

Question or issue resolved

Whether perturb-and-record — the instrument every reading in this register uses — works on an artificial system.

The question is not idle. Every biological reading is constrained by access: a brain can barely be perturbed and barely be recorded, which is why one deposit of thirteen patients yielded three and why eight of nine readable public deposits mix incompatible geometries. An artificial system inverts that completely. Every unit can be perturbed exactly, every unit observed perfectly, the state set rather than waited for, the experiment repeated identically. If the instrument is right, this is the substrate on which it is least limited.

METHOD §6 reaches artificial systems only through indicator profiles, which are not a measurement. This run asks whether the measurement could reach them too.

What was simulated

A discrete-time rate recurrent network, 100 units, tanh activation, Gaussian recurrent weights scaled so the spectral radius is approximately g. Recurrence is what makes this the right substrate: PCIst measures a response that spreads and evolves, and a feedforward network has no such evolution.

Two manipulations. g, the coupling gain, swept 0.4 to 2.0 — below 1 a perturbation dies where it lands, far above 1 the network saturates into a stereotyped response, and the prediction was an inverted U peaking near 1. Adaptation, an activity-dependent hyperpolarizing current, as the anesthesia analogue — this is not arbitrary, it implements cortical bistability, which is the leading mechanistic account of why PCI collapses under anesthesia.

Seven couplings × two states × five seeds, sixty trials each, both perturbed and unperturbed. Its own parameter set, artificial_rnn_v1, because METHOD §2.1 says the parameter set is per-paradigm and the human set was written for scalp TMS-EEG.

The result

Three things are wrong, and they are the same thing.

A network that was never perturbed scores 65, while the same network perturbed scores 4.8. The measure's own floor varies fifteen-fold across the sweep and runs opposite in direction to the perturbed value.

Bistability raised the measure instead of lowering it. At g = 1.5 the anesthesia analogue reads 33.16 against the baseline network's 19.73. In cortex the same mechanism collapses PCI. Here it inflates it.

Phase-shuffling the response raised PCIst from 18.99 to 33.48. This is the control that failed and it is the decisive one: destroying the temporal structure of the evoked response, while preserving every unit's power spectrum, made the response score as more complex. A measure of response complexity that rewards the destruction of the response's structure is not, in this regime, measuring that structure.

What it is measuring instead

All three point one way. On this substrate PCIst is dominated by cross-unit decorrelation. Weak coupling leaves units nearly independent and scores high; strong coupling correlates them and scores low; phase-shuffling decorrelates them completely and scores highest of all. The perturbation is real — the evoked response is 9.3× baseline amplitude, and the delivery control confirms it — but it is not what the number is tracking.

Why the biological pipeline does not hit this: a cortical evoked response is large relative to ongoing activity and spatially structured, and real recording noise is spatially correlated by common reference, volume conduction and shared afferent drive. Both conditions fail in a network of independently-noised units, and the measure has no way to say so.

Five defects the controls caught on the way, and one they did not

Each was found by a control, each looked like a result first, and each is the reason the sixth finding can be trusted.

1. Settling. The simulation started away from its stationary distribution, so the baseline window held a decaying transient. At g = 0, where only five of a hundred units were connected to the stimulus, 61 units nonetheless showed a "peak" above five times their own baseline.
2. The observation model. Independent per-unit noise is maximally incompressible. With a hundred uncoupled units, principal components are pure noise, roughly a fifth clear min_snr by chance, and PCIst reports high complexity for a system that did nothing. Fixed by making half the noise shared, which is what real recordings have.
3. The control itself was ill-posed. The original no-spread control set g = 0. But with coupling removed there is no system, only a hundred independent units, and the perturbational complexity of a thing that is not one thing is undefined. Replaced by an unperturbed run of the same network, which is well posed at every coupling.
4. Baseline-mean subtraction. Zeroing the mean over the baseline window — which the biological pipeline also does — removes real variance along with it, by an amount set by how many independent samples the window holds. With a 20 ms time constant lengthening further near g = 1, a 180 ms baseline was deflated 27% at g = 0.4 and 266% at g = 2.0, and PCIst read the deflation as a response: an unperturbed network scored 111. This is a live caution for the biological pipeline, whose 350 ms baseline is safe only because cortical autocorrelation is short.
5. A single-seed control measuring seed noise. The stationarity check failed at 0.79 on one seed. Across eight seeds the median is 0.95 with a standard deviation of 0.25 — no systematic drift at all. A control run once measures its own variance.

The sixth is the one they did not catch, and it must be stated: this run cannot fully separate a fact about PCIst from a fact about this simulation. The phase-shuffle argument is substrate-agnostic in its logic, but phase-shuffling destroys cross-unit coherence as well as within-unit timing, so it is the same phenomenon as the low-coupling result rather than independent evidence for it. A cleaner test would shuffle timing while preserving cross-unit structure. That test has not been run.

What follows for the standard

No reading is issued. The register does not carry a number from an instrument that failed its own control, and gate/artificial.py refuses to emit one. The seventy rows are recorded as evidence of what happened, not as a measurement.

METHOD has a gap, and this run argues for leaving it open. §6 reaches artificial systems only through indicator profiles. There is no reading type for a perturbational measurement of an artificial system, and the obvious move — extend Type A — is wrong twice over. First because the measure does not behave on this substrate. Second because Type A assumes a physical time base and physical recording units, and a simulated network has neither: its milliseconds are a modeling convention and its units are dimensionless. Even had the measure worked, its values could not share the Type A scale.

The negative result is the useful one. A standard whose instrument is assumed to generalize across substrates, and does not, would carry that error silently into every artificial-system reading it ever issued. It was cheaper to find out here.

Limitations, stated

• One architecture. A rate RNN with tanh units is not a transformer, and nothing here says what happens on the systems anyone actually cares about.
• One perturbation shape, 30 of 100 units at a fixed amplitude for 1 ms. The register's own CHOCS result found that protocol moves PCIst 2.07-fold in cortex, so the protocol dependence here is unexplored and probably large.
• resample is 200 Hz, chosen because the recurrence matrices go as the square of the samples and a long pre-period at 500 Hz demanded 8 GB. METHOD §2.1 records that this rate changes the answer.
• Five seeds per cell. The seed-to-seed spread in the stationarity check alone was a standard deviation of 0.25, so five is thin.