Ranking Under Biased Missingness: When Model Rank Is Not Identifiable from Sparse Leaderboards

A leaderboard rank implicitly claims that the reported scores order two models; on sparse leaderboards, that claim is usually taken for granted rather than checked. Public robotics and embodied-AI leaderboards report scores for only subsets of benchmarks. We study a cross-benchmark matrix spanning LIBERO, CALVIN, Meta-World, and LIBERO-Plus with 162 deduplicated models, 18 benchmark items, and 781 observed scores (26.8% density). Reporting coverage is negatively correlated with fitted item difficulty ($r=-0.639$), but this association does not identify the reporting mechanism. We fit a penalized two-parameter logistic item-response model and bootstrap observed cells ($B=1000$). The resulting intervals quantify pipeline sensitivity, not coverage-guaranteed uncertainty: 36.4% of model pairs are unresolved by the pairwise criterion on the unmasked matrix. The maximum divergence between latent-trait and average-based ranks is $\Delta=77$; the corresponding model's rank interval spans 58 positions. Under additional masking at 50% and 90%, we distinguish a fixed-denominator rate that includes unavailable pairs from a rate conditional on surviving models. At 90% masking these rates are 91.6% and 45.6%, respectively. Rank self-consistency with each estimator's own unmasked ranking also changes ordering: $\Delta\rho=-0.0031$ at 50% masking and $\Delta\rho=+0.106$ at 90%. These are outcomes for the sampled masks, not evidence of a stable regime boundary or superior ranking accuracy. Before a sparse ranking is trusted, it should report coverage and state which comparisons it can make, which it cannot resolve, and which it cannot make at all.