Covariance rank checks fitted to five reference SEEG vectors and one query accepted every query in rank-five groups at 83 sites in eight people

This study tested a brain-computer interface response check with SEEG data from eight people, using five early reference pulses at each of 83 sites and one later response or matched before-pulse control per check. All 2490 groups at 5, 10, and 20 cosine features had numerical rank five. For six centered vectors of rank five, the internal squared Mahalanobis scores equal 25/6, so the conservative check accepted every query. At ten features, it accepted all 415 later responses and all 415 controls. A fixed ridge term gave 81.9% response acceptance and 65.5% control acceptance under equal person weights. Ridge and Euclidean scores rejected every artificial shift at three and ten times the reference RMS norm; the unregularized score accepted them. Strict floating-point ranking created false differences between mathematically tied scores. The leverage identity is known; this contribution is a verified SEEG benchmark, arithmetic audit, proof for this rank rule, and fixed comparison. These data do not establish exchangeability, online performance, or clinical benefit.

conformal prediction · Mahalanobis distance · covariance rank · SEEG · brain-computer interface · rank deficiency · response acceptance · numerical audit