The problem — Every upstream stage — ingest, metadata, QC, correction — exists to make this stage trustworthy. But the readout stage is where a biologist first sees a number, which means it is also the stage where a pipeline failure can most confidently be mistaken for a biological result. A percent-replicating score of 20% can mean the assay is biologically uninformative, or it can mean illumination correction failed two plates back. Without an explicit gate, the two are indistinguishable.
What it is / how it works — The readout stage computes the domain-relevant biological signal: for Cell Painting, percent replicating (do replicates of the same perturbation retrieve each other above a null?) and mAP via copairs, implemented against the JUMP-scale benchmark (Chandrasekaran et al.); for digital pathology, a slide-level grade, biomarker score, or region annotation; for spatial omics, cell-type clustering and spatial niche maps; for live-cell imaging, division/death events and signaling dynamics. Metrics Reloaded (Maier-Hein et al.) grounds the evaluation philosophy: metric selection must reflect domain interest — the scientific estimand — not imaging convenience. A readout without a metric chosen for the domain interest is an output without a truth condition.
Where it breaks — CPJUMP1 cross-modality compound–gene matching sits barely above chance — a sobering baseline that shows the readout stage reflects the accumulated limitations of all stages before it, not just its own logic. A low readout score is a pipeline diagnosis signal, not a biological conclusion, until illumination, segmentation, and batch have been ruled out. This is why Why Microscopy Needs Verified Outputs treats the readout as a gate, not a destination. Uncertainty must be propagated forward: a hit nominated without a confidence interval or a null-distribution comparison is an opinion, not a measurement. The readout feeds The QC-Aware Report with the only number that ultimately matters.