Batch effects, plate position and run order

1 · Learn the move · Cognitive verifier

The most dangerous pattern in your data is the one that looks biological and is actually logistical: day two ran on the new reagent lot, the cases were processed in the morning and controls after lunch, the interesting cluster is column 12 - the edge. Batch structure generates results that replicate beautifully within the batch and never outside it. The cognitive-verifier move interposes sub-questions between seeing a pattern and believing it: before any interpretation, have the model enumerate every technical variable recorded - and notably not recorded - in your metadata that could produce the observed pattern, and for each, name the specific plot or comparison that would distinguish technical from biological. The model is well suited to the enumeration - it is a checklist mind - and the discipline is in the ordering: confound audit first, interpretation second, and only after the audit comes back clean enough to bear weight.

I observe [pattern] in [data]. Before any interpretation: (1) list every technical variable in my metadata - [paste: run date, plate, position, operator, lot, order] - that could in principle produce this pattern, plus the standard suspects I did NOT record, marked [NOT RECORDED]; (2) for each, state the specific check that would implicate or clear it: which plot, which grouping, which comparison; (3) run the checks possible with supplied data, report each as IMPLICATED / CLEARED / UNTESTABLE; (4) no biological interpretation until the audit is done, and none at all if a confound is IMPLICATED or the key variable is UNTESTABLE.