Construct Validity for Geometric Perturbation Scores: When Does a Number License a Mechanistic Claim?

We test whether domain-specific validity conditions reorganize geometric perturbation score rankings across three biological domains. Direction instability (angular consistency of perturbation signatures across contexts) was computed for 8,949 LINCS L1000 drug compounds, 1,676 Perturb-seq CRISPRi gene knockdowns, and 25,254 JUMP-CP Cell Painting morphological profiles. Five drug hypotheses and three genetic hypotheses were tested with Holm–Bonferroni correction within each domain family. Validity conditions preserve global rankings in all three domains (LINCS: ρ > 0.99; Perturb-seq: ρ = 0.91; JUMP-CP: ρ = 0.99) while upgrading inferential warrant at the item level. Phenotype-projected scores correlate with on-target genetic perturbation connectivity (ρ = 0.38, p < 10⁻²⁷) while raw scores do not (ρ = −0.04). Transport-stable scores outpredict raw scores for held-out cell-line effects in all 66 valid leave-one-out folds (Wilcoxon p = 8.2 × 10⁻¹³; mean Δρ = +1.10). Essential-gene correction reveals that universal-essential knockdowns diverge rather than converge across cell types (d = 0.60, p_perm < 10⁻⁴), inverting the predicted confound. The construct-validity framework (a "validity ladder" with seven conditions, three tested empirically) specifies which interpretive threats a score has survived. Validity conditions do not change geometric perturbation scores; they change what a practitioner can conclude from them. This repository contains the manuscript, all experimental scripts, pre-registration commits (SHA 8d74bd0 for initial freeze; SHA f3c88ed for extended pre-registration; SHA 5bb6080 for dose-direction separability test), deviation log, and figure generation code.