Record chip kind alongside process and device count on every run

Correct why topology changes a result: the data stream is invariant, the arithmetic is not

Mark the data-process count on every run: a seed alone does not identify one

Dataset card: add the routed 8k sweep

Add the routed 8k learning-rate sweep: 18 runs with per-expert load

Add full-resolution payloads for the study browser's opt-in download

Add a study card per folder: setup, reproduce command, and what it showed

Dataset card: carry the repository's audit of these logs verbatim

Add per-study snapshots: curves-only overview, plus diagnostics where spikes are the subject

Add rig training logs: 177 runs across six studies, full resolution (part 4)

Add rig training logs: 177 runs across six studies, full resolution (part 3)

Add rig training logs: 177 runs across six studies, full resolution (part 2)

Add rig training logs: 177 runs across six studies, full resolution

initial commit