Md Sadman Bin Masud

Collatz stopping times follow a log-normal distribution. Treated as convolution kernels, shifted along that distribution, they give a filter bank that is topologically deterministic — the same input always yields the same output, with no learned weights.

Assembled into a hybrid ensemble, the corrected filters measured 2.9 × on texture retention and 2.0 × on signal detection against the uncorrected baseline.

↓ you can play the thing the paper is about

Multi-class arrhythmia detection over the SPARC-ECG corpus. The headline accuracy figure is the easy number to quote and the easiest to be misled by, so it is not quoted here: on the class-balanced view the model reaches a macro-F1 of 46.89%, and that is the number this work stands behind.

Hardware-side results were verified under emulation, not on physical silicon — stated plainly because the distinction matters.

A solar-plus-grid controller with live instrumentation: generation, storage, load and switching modelled end to end, with the dashboard as the only interface that matters — if it cannot be seen, it cannot be trusted.