Field-level perturbational complexity is not a proxy for neuronal reproducible dimensionality; its within-brain state changes co-vary in direction

Perturbational complexity is measured from field potentials, yet the quantity of interest lives in neuronal populations. Whether the field-level measurement reflects the reproducible richness of the underlying spiking activity — or a field-specific shadow — has never been tested, because it requires both signals in the same brain during the same perturbations across arousal states. The open Allen Institute dataset provides exactly that: epidural EEG (17–30 channels per animal) and hundreds of Neuropixels single units, recorded simultaneously in head-fixed mice during cortical electrical stimulation, awake and under isoflurane. We computed debiased reproducible dimensionality (R-dim; our companion estimator, doi:10.5281/zenodo.22100546) from sorted spiking populations (338–995 units per session, 10 ms bins) on exactly the trial sets of our published EEG analysis (doi:10.5281/zenodo.22133403; state- and current-matched, same frozen pipeline and seeds), under a preregistration sealed before any spike time was read, with three hypotheses and a four-cell interpretation matrix. Results (n = 15 animals): levels do not correspond across animals (awake R-dim, spikes vs EEG: Spearman ρ = +0.25, p = 0.18) — the sealed cell reads "the field is not a proxy for population dimensionality in level," consistent with both quantities scaling with their own, independent coverage. The state contrast has a genuine neuronal counterpart (spiking R-dim 8.48 awake vs 4.69 under isoflurane; 11/15 animals; p = 0.0027). Spiking dimensionality runs ≈ 3.0× the simultaneous field value as measured and ≈ 2.4× once each modality's measured null floor is subtracted (floors archived in pisos_rdim.jsonl: 0.40–0.59 for 17–30 channels, 2.6–3.1 for 265–995 units). (v3) The two states' spiking growth curves do not run in parallel: on the fixed subset of 14 animals with data to 240 trials, the paired awake–isoflurane gap widens from +3.8 at n = 30 to +5.3 at n = 240 — the opposite of the field, where the companion paper found the gap narrowing. The third preregistered test passed its bar — within-animal state effects co-vary across levels (Δ_spikes vs Δ_EEG: ρ = +0.64, p = 0.0054) — but that coupling is fragile: it is carried by the three animals whose state effect reversed sign, and among the twelve with the ordinary effect it is ρ = +0.301, p = 0.171 (ρ = +0.182 among the eleven with both effects positive). What the data support is agreement in the direction of within-animal change (14/15 animals, the reversals included; sign test p = 0.001, post-hoc), not graded tracking of its magnitude; graded tracking is what a per-patient clinical use would require, and it is not established here. The pattern survives both preregistered bin sizes (5 and 25 ms). Labeled exploratory analyses locate the reversed animals on a continuum: spontaneous firing was suppressed in every animal (isoflurane/awake rate ratio 0.29 in the reversed animals vs 0.27 in the rest; p = 0.63), while the anesthetic's attenuation of the evoked response varied enormously (1.1–7.9×) and the state effect followed that attenuation (ρ ≈ +0.6 at both levels): where the anesthetic failed to suppress responsiveness, dimensionality persisted. A sealed follow-up on the recovery blocks resolves our previous reversibility puzzle: neuronal dimensionality does partially recover (7/8, p = 0.012), by the same fraction as the field (0.24 vs 0.21) — the incomplete recovery was biology, not measurement — and spontaneous firing recovers about twice the fraction dimensionality does (0.40 vs 0.23, p = 0.008): arousal returns before reproducible dimensionality. A sealed trial-scaling analysis finds no deceleration of spiking R-dim up to 480 usable trials (growth at top p = 0.042 on the ten sessions as run; on the four sessions that actually reached 480, a one-sided test at n = 4 cannot reach α = 0.05, p = 0.0625, so that result is reported as descriptive). Files: full preprint (HTML), two figures, and a reproducible code archive with all analysis scripts, sealed preregistrations, per-animal results, the null-floor script and its output (pisos_rdim.py / pisos_rdim.jsonl), and the mechanical verifier that recomputes every number claimed in the paper. No raw neural data are redistributed.