A label is an operator-owned aspect/value annotation over an existing Lake session. Each assignment appends one fsynced line to LAKE_DATA/labels/labels.ndjson with timestamp, session id, runtime, normalized aspect, value, optional note, and provenance source.
Labels do not change canonical events. The SQL labels view exposes full history; CLI lists use latest-assignment-wins for each session/aspect pair. Reassigning an aspect appends rather than edits.
Aspects are trimmed and lowercased. Values and notes are trimmed. Provenance is manual by default and may be manual, human, model, or brama, optionally followed by a :<detail> suffix.
Labels are not passed through the event masker. Never put credentials or private transcript excerpts in a label or note. The label store is independent of the event writer lease, so labeling can continue while the stream runs.
goal label is a specialized model-produced label over the first masked user prompt; it records model provenance.
See label add, label list, and label aspects in the CLI reference.