The candidate does not survive contact with tissue, and the instrument has an amplitude confound

Run: gate/spread_real.py → gate/spread_real.json Date: 2026-09-21. Six sessions of DANDI:000458, the deposit behind NS-0006, NS-0011 and NS-0016.

Question or issue resolved

gate/RESULT-TOPOLOGIES.md left negated delta_spread as this program's first measure with a real relationship to integrated information and a stated domain of validity: it tracks integration where propagation is distance-structured, at −0.91 on a ring and −0.99 on a chain, and fails where it is not. Cortex is distance-structured, so the measure should work there. That entire case was built on constructed systems. This is the first time it meets tissue.

The test is not a validation of either quantity. Neither PCIst nor delta_spread is a ground truth and a brain has no known answer — which is the reason this program builds constructed ones. What two measures can do on real data is disagree, and the question is where.

Both are computed from the same evoked responses under the identical channel and trial rules the register already applied to this deposit, because comparing measures through two preprocessing pipelines is a comparison of pipelines wearing a comparison of measures.

The measure fails, and the way it fails is diagnostic

Across six sessions with both wakefulness and isoflurane:

The definition as written gets it backwards in five of six sessions. That is not weak evidence for the measure; it is evidence against it, and the diagnosis is immediate:

spread_raw is the standard deviation across channels of each channel's mean absolute response. On the constructed systems every network sat at a matched operating point, so that quantity was scale-free by construction and genuinely measured unevenness. Tissue has no such courtesy: an awake cortex answers a jolt with a larger response, a larger response has a larger standard deviation for that reason alone, and the measure reads amplitude while appearing to read distribution. The artificial-system result was not wrong about the artificial systems. It was wrong about what the definition would mean anywhere else.

Repairing it does not rescue it. Dividing by the mean makes the measure scale-free — amplitude correlation drops from +0.84 to +0.30 — and its separation improves to 4 of 6, which is not significant. Its correlation with PCIst across all sixteen state-cells is +0.10. Once amplitude is removed, almost nothing is left.

A second finding, about the register's own instrument

The same table carries something that was not being looked for. Across all sixteen state-cells, PCIst itself correlates with raw response amplitude at ρ = +0.66.

That is not a refutation — a richer response is plausibly both larger and more complex, and the two need not be separable in principle. But an instrument whose value tracks how big the response was, on data where amplitude varies more than sixfold between sessions, is carrying a confound that belongs on the record. It sits alongside the protocol effect (gate/RESULT-CHOCS.md: a median factor of 2.07 from waveform and repetition rate alone) as a second way for the number to move for reasons that are not the subject's state.

It also sharpens what the paired contrasts are for. NS-0015 and NS-0016 hold subject, array and protocol constant across their two arms, so an amplitude confound shared by both arms partly cancels in the ratio — which is one more reason the contrast is worth more than either absolute value.

Scope, stated plainly

Six sessions, not the register's twenty-four. These are the sessions cached locally; the rest of the deposit streams from DANDI and was not re-fetched for this. Six pairs is enough for PCIst to reach p = 0.016 and enough to show spread_raw running backwards, and not enough to call 4 of 6 either a success or a failure. The register's own NS-0016 stands at 21 of 24 on the full deposit.

PCIst here is computed on the deposit's documented window and reproduces the direction of the register's readings, not their exact values, because this run pools states without the depth, current and running-state stratification NS-0006 and NS-0011 use.

The conclusion

Negated delta_spread is withdrawn as a candidate. It measured unevenness on systems held at a matched operating point and measures amplitude on tissue, and the version that is genuinely scale-free has essentially no relationship to the instrument it would have to improve on. Three runs of constructed-system evidence did not survive one contact with a brain, which is the cheapest possible moment to find that out and an argument for making this contact earlier in the sequence rather than last.

Two things follow.

Every future candidate gets tested on tissue before it gets written up. The failure here is not that the measure was bad on artificial systems — it was excellent there. It is that "scale-free by construction" was a property of the test bed that got silently carried into the definition. A measure that has only ever been scored on matched systems has an untested assumption in it by default, and the test costs an afternoon.

The different-currency work returns to the main line. It was demoted when delta_spread appeared to work. Surrogate-normalized compressibility and the Loschmidt echo are both scale-free by construction rather than by circumstance, which is now a property worth selecting for rather than a footnote.

And the amplitude correlation goes into the register as a caveat on PCIst, not into a drawer.