Debiased reproducible dimensionality is lower under isoflurane than in wakefulness in the same mouse brain, at matched trial counts and stimulation currents

Perturbational complexity distinguishes conscious from unconscious brain states, but our companion work showed that the standard estimator (PCIst) partly reflects a selection bias, and that what it tracks is the response's reproducible dimensionality (R-dim) — the number of response components that replicate across independent trial halves, estimated fully out-of-sample (doi:10.5281/zenodo.22100546, .22100826). Readers of v1 should start here: the contrast reported below is real as a measurement, but since v2 this paper shows it is quantitatively accounted for by a drop in evoked amplitude between states rather than by a loss of response repertoire, and we no longer present it as evidence for the latter. The reason is a property of the estimator that v1 had not measured: on a known-truth bench, a pure change of gain — the true dimensionality held fixed by construction — is read by R-dim as a change of dimensionality, because R-dim under-reads as per-component SNR falls. Any R-dim difference between two conditions that also differ in evoked amplitude must therefore be read as an amplitude difference first, and only as a repertoire difference once amplitude has been equalized; here it does not survive that step. Here we take the debiased estimator to the within-subject arousal-state contrast it was built for, using the open Allen Institute dataset of cortical electrical stimulation with 30-channel epidural EEG in head-fixed mice, each recorded awake and under isoflurane (dandiset 000458). All hypotheses, exclusion rules, matching rules and a four-cell interpretation matrix were preregistered with public timestamps before any neural signal was read; two technical amendments made at the pilot stage are likewise publicly timestamped. Because R-dim grows with trial count (doi:10.5281/zenodo.22120069), all state comparisons were made at matched trial counts and matched stimulation currents within each animal — to our knowledge the first arousal-state comparison of perturbational complexity controlling for both. Result: R-dim was lower under isoflurane than awake (medians 2.08 vs 2.80 at n = 120 matched trials; 16/20 animals; Wilcoxon p = 0.008), robust across all four preregistered sensitivity analyses (band, artifact margin, movement inclusion, rejection threshold; p = 0.0001–0.007) — and the effect strengthened when the response window was moved further from the stimulation artifact. (v2) v1 read that as arguing against artifact leakage; that inference is retracted, because the artifact blanking declared in the protocol was never applied in the code and the zero-phase filter smears the artifact across the epoch. (v2) v1 reported that the two states' growth curves ran in parallel and read that as anesthesia lowering the level without changing the growth. That is withdrawn: those medians were taken over a different animal subset at each grid point, and on a fixed subset the paired gap narrows monotonically (+0.78 at n = 30 to +0.38 at n = 240). The contrast is instead accounted for by evoked amplitude: the between-state amplitude ratio, measured on the same matched trials, correlates with the R-dim gap at +0.65 (p = 0.002), and extrapolated to equal amplitude the state effect is −0.34 ± 0.28, indistinguishable from zero; the ten animals with the smallest amplitude drop show no detectable effect (6/10, p = 0.42; a test of this size has low power, so this is absence of detection, not evidence of a null). A separately sealed protocol of this program (commit 9be5d99) passed its ordering and calibration criteria for the gain account; (v3) its trivial comparator — the amplitude ratio alone, with no model — orders the animals better than the sealed formula (Spearman +0.70 vs +0.41), so what those data support is "the amplitude ratio explains the ordering", not the specific formula. We therefore report this as a contrast that R-dim registers within the same brain — the property a clinical index needs — whose proximate cause here is attenuation of the evoked response, not truncation of its repertoire. Two preregistered predictions failed. Curve flattening under anesthesia was not detected — but a post-hoc power analysis shows this test was underpowered (39% power for a moderate effect), so the question remains open. Recovery, however, did not restore R-dim (p = 0.21) in a test that was adequately powered (90% for an awake-sized effect); by our sealed interpretation matrix, this downgrades the headline claim from "tracks the state" to "is associated with the state, with temporal drift not excluded" — we state this limitation rather than bury it. Finally, standard PCIst and debiased R-dim were compared on the same 22 confirmatory animals and the same trials (each animal's maximum matched count): PCIst separated the states by a larger relative margin (+32% vs +22%) but less consistently (14/22 animals, p = 0.19) than R-dim (19/22, p = 0.0002). (v3) v1 stated this comparison as being "on the identical matched trials" with R-dim's p = 0.008, which mixed two samples, and v2's "re-pairing" on 23 animals included the pilot animal that the preregistration excludes; both are superseded by the numbers above. The bias-corrected estimator detects the arousal-state contrast more consistently than the estimator the field currently uses. 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.