Detected sample sharing across SEEG train and test splits changed with the input time range in nine files from seven people

We tested input sample sharing for stimulation response readouts. We used three SEEG channels and 120 windows per file from OpenNeuro ds006065 version 1.0.0. The nine files came from seven people. In one fixed random split per file, 94.6% of test windows had a measured train match for the full input range. Inputs from 50 to 250 ms after the candidate pulse had no detected match. An exhaustive short-window search found no valid 32-sample exact run in 64,260 within-file pairs. Adding long history or centered finite-filter input gave match rates of 67.0% and 54.8%. Rates include exact words or intervals inferred from close full-window matches. Exact words alone gave rates of 7.1% and 4.2% for those added inputs. These are equal-person means. We verified all 111 inherited exact runs and replayed source bytes in all nine files. The package contains data, three figures, a support audit, purge tools, a conditional gap proof, and an offline rebuild. These results do not establish independent trials or better decoder performance. Candidate pulse times depend on the source export layout.

SEEG · brain-computer interface · data leakage · sample sharing · train-test split · signal processing · reproducibility