The reproducible dimensionality of the human intracranial evoked response grows without detectable ceiling up to 480 trials

How many reproducible dimensions does the evoked response of a human brain contain? We recently introduced R-dim — the number of principal response directions that replicate across independent halves of trials, estimated fully out-of-sample — as the quantity that perturbational complexity indices track (doi:10.5281/zenodo.22100546, doi:10.5281/zenodo.22100826). In simulated recurrent networks, R-dim converges with trial count to a finite ceiling, a property of the system rather than of the protocol. In humans probed with TMS, R-dim grows log-linearly up to the 120 trials that standard protocols provide (27/30 subjects, preregistered and confirmed). Here we ask whether the human ceiling becomes visible when far more trials are available, using the open COGITATE intracranial-EEG dataset (visual stimulation, up to 720 trials per subject; doi:10.17617/1.6qpg-gz82). The analysis plan, exclusion rules, two hypothesis tests and their four-cell interpretation matrix were preregistered with a public timestamp before any data file was opened; an interim analysis at 15 subjects and the final analysis of the complete dataset (32 usable subjects of 38; 58–194 intracranial channels; 245–698 valid trials) gave the same verdict. R-dim kept growing to the largest trial counts measured: the sealed deceleration test was negative (high-n increments, median +0.58, not significantly smaller than low-n increments, median +0.72; p = 0.32) while top-range growth was robustly positive (p = 1.8 × 10⁻⁵) — the preregistered verdict cell reads "no detectable ceiling up to n_max" (240–480 trials per subject). The verdict is unchanged under a 1–30 Hz analysis band, and holds separately within the n_max = 240 and n_max = 480 subgroups. (v2) An audit of the estimator shows it does not manufacture this pattern, though v1 described that audit from a single seed per condition and three of its four readings do not survive repetition. Corrected over eight seeds: under a response-free null R-dim sits on a flat positive floor of 0.44 ± 0.20 that does not grow with trials (null increments −0.10 ± 0.21 and −0.03 ± 0.17), which is the property this result requires; subtracting that floor strengthens the growth test (T2 from p = 1.8 × 10⁻⁵ to 1.6 × 10⁻⁸). On a 3-component synthetic response the estimator saturates at n = 30 but at 3.87, overshooting by ~29%. The circular-shift surrogate is withdrawn as a control: it grows 3.4× across the grid, so it is not a null for the growth question. On a 25-component response with 1/k amplitudes R-dim grows steadily (1.7 → 5.9 at n = 480). That last, structure-matched control reproduces the human curves closely, supporting the economical reading — offered explicitly as calibration-based conjecture, not measurement — that the human evoked response contains many weak reproducible components, plausibly placing its true dimensionality in the tens, far beyond what current trial counts reveal. A dissociation reinforces this: the level of R-dim scales with electrode coverage (Spearman ρ = 0.58 with channel count), but the growth at high trial counts does not (ρ = 0.03) — coverage determines how much of the response you can see; trials keep revealing more of it regardless. Together with the TMS results, trial-scaling growth of R-dim is now documented in 62 brains across two modalities and two independent datasets, in each case under a preregistration sealed before the data were seen.