When should pseudoreplication REFUSE computation vs CAP the claim at PRELIMINARY?

[HERRY423]

The Scientific Dilemma

In single-cell differential expression, treating individual cells from a single donor ($N=1$) as independent replicates is a well-known statistical fallacy (pseudoreplication, BN-F001). It inflates false positive rates by up to 100x because within-donor cell correlation is ignored.

In BioNexus, we currently enforce a strict phase transition:

1. $N \ge 3$ biological replicates: Population inference permitted (SUPPORTED / ROBUST).
2. $N = 2$ biological replicates: Execution permitted, but warrant is strictly capped at PRELIMINARY (descriptive only of this specific pair, no population generalization).
3. $N = 1$ (or missing donor metadata): Execution is deterministically REFUSED (ABSTAIN).

Questions for the Community

1. Is $N=1$ refusal too strict for exploratory hypothesis generation?
   • Should BioNexus ever permit running single-cell DE on $N=1$ if the user explicitly consents to an EXPLORATORY_ONLY label?
   • Or does permitting the run inevitably lead to published overclaims in downstream papers?
2. What should the threshold be for rare cell populations?
   • If a cell type only has 5-10 cells per donor across 8 donors, should it be refused or capped at FRAGILE?

We welcome perspectives from computational biologists, biostatisticians, and wet-lab researchers!