Run surface¶

The run surface is how MolRec represents training jobs, simulations in flight, and multi-stage workflows as records.

method   →  what method / engine / model is running
status   →  where execution is now
metrics  →  append-only measurements along the way

method

Scientific or training context. See Method. In the reference binding: group attributes.

status

Lifecycle state, stage, progress, errors. See Status. In the reference binding: group attributes.

metrics

Append-oriented run-local curves and counters. See Metrics. In the reference binding: JSONL text buffer, densified to Zarr series.

Observables are scientific quantities that are part of the interpreted result of the record. Metrics are run-local monitoring. Training loss belongs under metrics; a published series belongs under observables.

A run-shaped record requires meta and status (with at least status.state when the section exists). It should include at least one of metrics or method. frame and system are optional.

Typical compositions: an ML training job is meta + method + status + metrics; an MD production monitor may add trajectory; a failed job for resume is meta + status (+ status.error) + method.

A host (a molexp run directory) keeps lifecycle in host files and metrics on *.mlp.* — see Root layout. status/ and meta/ are record groups.