Fig. 2. Laminar source-reconstruction accuracy as a function of sensor-level SNR.

a–d Mean relative free energy (ΔF) matrices (relative to the worst model; S superficial surface layer, D deep surface layer) for simulated signals placed on 11 cortical surfaces at four signal-to-noise ratio (SNR) levels (-∞dB, −50, −35, −20 dB). Accurate laminar inference begins to emerge around −35 dB, with near-perfect identification at −20 dB. Caret markers denote the forward model with peak free energy for each simulated depth. e–h Probability density of each forward model being the maximum-evidence solution, aggregated across cortical locations: a diagonal pattern appears from −35 dB, indicating accurate depth localization. i, j Laminar reconstruction error (in layers and mm) falls below chance at −35 dB and improves at higher SNR. Shuffled controls show ~1 mm error due to chance matches between simulated and evaluated layers. k, l Laminar reconstruction bias (in surface layers and millimeters), again comparing simulated and shuffled data. In i–l, solid lines depict the mean error or bias across simulated locations, shaded bands represent 95% bootstrap confidence intervals of the mean, and gray dashed lines depict values for individual simulated locations. Statistical significance was assessed using permutation tests (10,000 permutations; one-sided for reconstruction error (i, j) and two-sided for reconstruction bias (k, l)), with no adjustment for multiple comparisons. Asterisks denote SNR levels at which the observed values differed significantly from the shuffled null distribution (p < 0.05). Source data are provided as a Source Data file, which includes exact p values and standardized effect sizes.