Machines that healSystems that feedModels that reason

A leakage-controlled benchmark for forecasting brain state from EEG. The finding that matters came from auditing my own evaluation: the input and target windows overlapped by 126 of 127 samples, so the headline score was measuring a model copying its input one step shifted, not forecasting anything. Rebuilt overlap-free, the identical ridge baseline moved from 0.0018 to 0.227 MSE — a 126× swing produced by the scoring rule alone, on one model and one dataset. I then derived an analytic bound on how much overlap can inflate a score, checked it across 286 configurations with zero violations, and added a guard that makes the leaking configuration impossible to construct.

Three benchmark datasets covering distraction, indirect speech, and nested belief reasoning. The result worth reporting was not a score. Two architectures failed in opposite directions, and the shape of the failure was more informative than the ranking.

Characterising plasma dose against bacterial load in E. coli culture, and a summer of library prep and sequencing. This is where the discipline for running the plasma bench properly came from: a wet lab does not forgive a process you have not written down.