Limitations

Observable Library 0.1.0 is a deliberately small, headless package. These boundaries are part of the released behavior.

Generation is parameter-only

generate() enumerates model.named_parameters(). It has no source allowlist and does not generate activation, gradient, or loss observables. Filter the returned list to reduce Runtime work or construct a custom Observable for another source.

Selection supports only all values

The only supported selector is selector="all". Slice, axis, and semantic selectors are not part of 0.1.0.

Hooks retain the latest value

HookSource.attach() installs broad hooks on top-level child activations and trainable parameter gradients. It retains the latest captured tensor without checking that it belongs to the current step. Run the matching forward/backward before every observation and call detach() when finished.

Loss is not discovered automatically. Call record_loss(loss, step) before observing a custom source="loss" spec.

Budgeting uses estimates

The scheduler sums budget_hint["compute_ms"] estimates. It does not measure execution time, learn a cost model, or change a spec’s frequency. A missing custom estimate counts as zero.

Temporal metadata does not execute

The temporal field contributes to ObservableSpec.id, but Runtime does not automatically call delta, ema, slope, or rolling_std. Apply those functions to explicit histories.

Storage and query are intentionally narrow

LocalStorage is a convenience ValueSink: SQLite metadata plus NumPy NPZ payloads. query() reads by exact observable id and step only. There is no query by source, reduction, tag, or display label.

No application surface

Version 0.1.0 has no CLI, Web UI, notebook product, dataset downloader, or general analysis service. Later roadmap ideas are not current package capabilities.