The problem — A dataset without a validated manifest is a dataset that exists only in the memory of the person who ran it. Which compound was in well B14? Which scanner produced plate 3? Were the DMSO controls distributed uniformly? Without a structured, machine-readable record, these questions become reconstructions — and reconstructions are not provenance.
What it is / how it works — The metadata stage writes the run manifest: the structured record of what was acquired and how. Huisman et al. distinguish two complementary layers — provenance metadata (MPM: the instrument and acquisition story) and quality metadata (MQM: calibration and error). Both are captured here and encoded in OME-NGFF conventions so they travel with the pixel store rather than in a lab notebook. For HCS plates this means compound / dose / timepoint per well, control layout, and scanner/objective/exposure settings. For light-sheet it means acquisition geometry, view angles, and time-lapse indexing. For spatial omics it means marker panel, cycle/round index, and ROI definitions. The manifest is a build artifact — generated at ingest, validated against a schema, and versioned — not a post-hoc annotation.
Where it breaks — Metadata captured once, manually, at data deposition time is metadata that drifts. A plate map entered by hand after a run confuses wells; a missing exposure field leaves QC unable to flag saturation; a swapped scanner identifier makes multi-site batch correction incorrect by construction. The discipline is to make manifest creation automatic and validation mandatory before any downstream stage runs. A manifest that cannot answer "which well, which control, which scanner" is decoration, not provenance — and it makes Why Microscopy Needs Verified Outputs an aspiration rather than a contract. See also Metadata & Provenance: The Run Manifest for the full authoring pattern.