Inside the distance ladder's headline number live small disagreements that refuse to average away. Swap the geometric anchor and H0 moves: NGC 4258's maser-anchored calibration runs about 1.5 km/s/Mpc below the three-anchor average, the LMC-anchored variant runs high, and Milky Way parallax anchoring lands between, spreads of order one to two sigma that the error budget accommodates but does not explain. Partition the supernova sample and it moves again: hosts split by mass, star-formation rate, dust properties, or survey of origin yield H0 shifts at the kilometer-per-second level; the supernova standardization parameters themselves drift between subsamples; and the same structure appears across the rival programs, where Chicago-Carnegie and SH0ES analyses of overlapping galaxies agree on distances yet diverge by 3 km/s/Mpc in H0 through host-selection differences alone.
The standard interpretation files these splits as the statistical texture of a hard measurement: each is individually within tolerances, collectively they are watched as potential systematics, and the programs' mutual critiques (whose host cuts, whose dust law, whose anchor weighting) assume the splits are noise about one true number. What the noise-reading cannot explain is the splits' organization: the shifts correlate with host environment and sample geography rather than with instrumental or methodological boundaries, the pattern repeating across programs that share no pipeline, as if the ladder were sampling something structured rather than erring around something uniform.
The standing is the tension's fine print: sub-sigma individually, persistent collectively, organized suspiciously, and about to be resolved or exposed by Rubin-era supernova samples large enough to map the splits across environment with real statistics.