Case studies
Four ways a finding fails
Each story is a real failure mode in single-cell analysis — and each is grounded in a bundled benchmark case the engine actually grades. The deciding evidence is always a re-derivable statistic or a critical axis, never an opinion.
Pseudoreplication
Does not surviveThe donor that faked a discovery
A cell-level test treats every cell as an independent replicate. But cells within a donor are correlated, so a signal carried by one or two donors looks astronomically significant at the cell level and vanishes when you aggregate to the donor. This is the classic single-cell false positive.
Read the case →
Perfect batch confounding
Does not surviveWhen condition IS the batch
If every A sample was run in one batch and every B sample in another, the biological effect and the batch effect are mathematically indistinguishable. No amount of downstream modelling can separate them — the design itself is the problem.
Read the case →
Circular inference
Does not surviveTesting the groups you defined
Double-dipping (circular inference) is when you use the data to define groups, then test those groups on the same data. The test is guaranteed to find a difference — you built the groups to have one. A high multiverse-survival rate does not fix this; it just shows the circularity is stable.
Read the case →
Real experimental labels
At riskThe doublet-driven signal (real labels)
A doublet — two cells captured as one — can masquerade as a novel co-expressing cell state. This is the only VERITAS benchmark scored against real experimental labels (hashtag oligos), so the numbers reflect real-world detection, honestly: strong ranking (AUC ~0.78), partial recall at the expected-rate cutoff.
Read the case →
These aren't hypotheticals
Every case above maps to a bundled benchmark case in the validation atlas — with a graded expected verdict.