QC curation loop#

The QC tab in the Results Viewer composes one Plotly card per configured QualityCheck analyzer. Each card subscribes to the same STORE_REMOVED_KEYS store that backs the “remove colony” toggle, so every time you curate a flagged colony the card’s figure, summary strip, and status badge re-render automatically — no manual refresh, no full page reload.

The loop:

  1. Configure one or more checks (e.g. ExpectedVsDetectedCount, GridOccupancy, ReplicateAgreement).

  2. The card highlights every flagged group with its severity.

  3. Mark all flagged colonies for removal (per-card button) or curate individuals from the Plate / Colony tabs.

  4. The card re-renders; remaining severity drops; status badge transitions fail → warn → pass.

Prerequisites#

  • A finished CLI run with deliverables/master_measurements.parquet + the post-applied deliverables/measurements.parquet mirror under <output>/deliverables/measurements.parquet. See Run Locally to produce one.

  • A metadata.csv describing the expected plate layout if you plan to configure ExpectedVsDetectedCount or GridOccupancy (their groupby columns must resolve against the master measurements schema). Both read the layout’s per-group row count as the expected colony/cell count; GridOccupancy additionally counts distinct filled grid cells (Grid_RowMajorIdx), so it reports occupancy without double-counting doublets.

Walkthrough#

Open the Viewer tab in the hub and pick the QC sub-tab:

QC tab in empty state with no checks configured.

The empty state shows a placeholder explaining that no checks are configured yet. The top strip carries + Add check, which opens a modal listing every concrete QualityCheck subclass discovered by OperationRegistry. After configuration changes, Rebuild QC database explicitly publishes the validated catalog and module tables to <output>/deliverables/qc/qc.duckdb.

QC tab body with the tab selected.

Click + Add check. The modal opens with a class dropdown plus an inline param form that adapts to the chosen check’s constructor signature:

Add-check modal open showing class picker dropdown.

Pick ExpectedVsDetectedCount, fill in the path to your plate-layout metadata.csv, then click Save. A new card slides into the cards container with:

  • a Plotly figure showing per-group Delta (expected − detected),

  • a summary strip reading groups: N | flagged: K | max severity: X.YZ,

  • a status badge coloured by the worst per-group status (green / yellow / red),

  • per-card buttons: Edit, Duplicate, Toggle enabled, Delete, and Mark all flagged for removal.

Switch to the Plate tab and remove a flagged colony (or hit the per-card Mark all flagged for removal button to union every flagged (ImageFile, Object_Label) into STORE_REMOVED_KEYS in one click). Switching back to the QC tab shows the card re-rendered with the new severity.

Common gotchas#

  • Metadata schema: use the canonical MetadataImage_ImageName image key. ExpectedVsDetectedCount uses the caller-selected groupby columns, which must exist in both the measurements and supplied metadata layout. A missing grouping column records a load warning instead of breaking viewer startup; the banner lists the affected instance_id and underlying error.

  • QC recipe: every add / edit / delete writes the qc array in the pipeline config under <output>/deliverables/ (pipeline.json.pht-pipe on current runs). Legacy/incompatible payloads are classified first: any migration is an explicit, backed-up, atomic action, never a startup rewrite. A configuration edit waits for the explicit Rebuild QC database action; rebuild stages and validates deliverables/qc/qc.duckdb and refuses an active or changed source. Concurrent viewer sessions on the same output dir are still unsupported.

  • Severity legend: check-side severity_warn / severity_fail thresholds default to 0.05 / 0.10. Tune them on the per-check edit modal to match your QC tolerance.

Where to next#

  • Heatmap exploration — pivot the same measurements into a plate-shaped heatmap and watch edge / contamination patterns light up.

  • View Results — the curation primitives the QC tab piggybacks on.

  • Analysis — once the QC chain is happy, configure filters + an endpoint model and emit analysis.{csv,parquet}.