What problem does it solve? ML experiments often run without errors yet produce wrong or misleading results: a metric that looks too good, a number that moved when nothing changed, or a mechanism asserted but never measured. This Skill provides a systematic diagnostic ladder and targeted probes to find the real cause instead of trusting plausible-looking logs. ## Core Features & Use Cases - Diagnostic Ladder: A seven-step progression from verifying the objective and data meaning through tiny overfitting, synthetic recovery, eval consistency, and baseline comparison. - Strip-to-Boring Protocol: Reduce to one seed, one device, one batch, and remove augmentation, dropout, distributed training, mixed precision, and EMA to isolate the bug. - Specialized Probes: Checks for probabilistic models, diffusion conventions, latent-component identifiability, seed sweeps, invariance sweeps, and self-satisfying metrics. - Use Case: Your model recovers parameters on real data but a per-component metric fails on some seeds only. The Skill guides you to test label-permutation symmetry in the mixture model, add a component-distinguishing prior, and run multiple chains to confirm identifiability failure. ## Quick Start Ask the AI to debug why my ML experiment's metric improved unexpectedly even though nothing in the training setup should have changed.
