Eval & inspect
No preset zoo. You put probes in evals/. aq eval is the human gate. Interpretability is a capability after fit: artifacts/inspect.md when the method supports it, not a separate product.
PrerequisiteTrained checkpoint under artifacts/checkpoints/ · probes in evals/
aq eval
Score every probe in evals/ (or one named file). Writes artifacts/eval.json. Pass/fail only when recipe.eval.min_score is set.
| Flag | Description |
|---|---|
| [name] | Probe basename. |
| --ckpt <name> | Named checkpoint. |
example
$aq evalexample
cat artifacts/eval.jsonartifacts/inspect.md
Optional human-readable dump after fit (coefficients, trees, tokenizer notes). Part of how you check a fit, not SAE / activation tracing.
example
cat artifacts/inspect.mdaq diff
Compare run records after forking a variant or retraining.
example
$aq diff <run-a> <run-b>