pandm

Query and analyze machine-learning experiments tracked by pandm — list runs, read a run's config/summary/metrics, compare runs to find the best hyperparameters, read the display specs the training code declared (which metric is the experiment's judge, which invariants must hold), read full metric series, and locate logged images. Reads a local `.pandm/` SQLite + PNG store directly; fully offline, no network, account, or server required. Use when the user asks about past pandm experiments, wants to compare runs, pick a winner, check whether a run broke an invariant, inspect a metric over time, or analyze results stored in a `.pandm/` directory.

Record machine-learning experiment metrics, images, distributions, hyperparameters, and training progress to pandm — a local-first, offline-by-default, account-free wandb/tensorboard alternative that writes to a local `.pandm/` SQLite + PNG store. Needs no network, server, login, or API key; a shared server is strictly optional, and once one is configured the same code auto-pushes to it in the background. Use whenever you write OR edit pandm instrumentation: instrumenting a training/eval loop, logging scalar metrics or images, adding/renaming/regrouping a logged metric, changing what a script reports, shaping how a chart renders (titles, subtitles, units, axis labels, panels, confidence bands, bar/histogram charts — especially for RL), reporting an ETA, or saving a run's config to compare in the dashboard. Read this before touching any `pandm.init` / `run.log` / `run.define_metric` call — it sets the titling discipline every run and every metric needs.