Weight Endpoint Continuity: Lineage-Consistent Correlation and Aligned Update Geometry in Public Tensors

A version number is a claim about provenance, and published weights can be checked for CONSISTENCY with it: endpoint continuity (how strongly two released endpoints correlate tensor-by-tensor) and aligned update geometry (how two releases' deltas from a common ancestor relate element-wise). Neither is training-history ground truth; the paper states which lineage stories each reading admits and which it disfavours, and defers calibrated genealogy inference to a registered documented-lineage campaign (FP-545). Corresponding-tensor correlation between checkpoint pairs — per- tensor Pearson on matched parameter names, deterministic per-tensor element subsampling, CPU only, minutes per pair — separates shared-lineage pairs from unrelated parameters by about two orders of magnitude, with bars fixed in a committed driver before any target ran. Two positive controls pass (0.9961 on a Qwen3.5-2B base/instruct pair, 62 tensors compared; 0.9950 on a Gemma-4-12B instruct/QAT pair that shares lineage by construction, 156 tensors) against within-model cross-layer nulls of 0.0122 and 0.0009. The lower bar is the weaker of the two: the INDEPENDENT-RUNS verdict has never been exercised on a genuinely independent pair, because those floors are within-model cross-layer nulls rather than two separately pretrained checkpoints. Applied to three same-architecture Qwen 27B releases the instrument returns readings consistent with one shared-prefix lineage geometry (a "family tree" only as an illustrative admissible topology, not a recovered history): a generation bump reads as endpoint-continuous rather than fresh (3.5-vs-3.6 = 0.9403 against a 0.0145 null — consistent with continued training on a shared basin, with the independent-pretrain control untested and correlation carrying no direction). On release night for a third checkpoint, two predictions were sealed and one of each returned. P1 (the new release shares lineage) hit at 0.8297. P2 (a strict descent-chain ordering) missed — the grandparent correlates slightly more with the newest release than the parent does (0.8383 versus 0.8297), and a paired per-tensor breakdown shows the reversal is systematic, favouring the grandparent in 135 of 153 tensors — so release-number ordering did not match the pre-registered strict-chain ordering and is an unreliable proxy for weight-space order; the miss is reported at full weight. A pre-registered discriminator read the update geometry from the deltas themselves: the correlation between the two displacements from the common ancestor has median 0.265 over 153 sampled tensors, the third of three registered bands — consistent with a shared early post-training component followed by per-release branches, with strict-chain and pure-sibling geometries disfavoured, though endpoint correlations alone cannot uniquely identify a training topology — with the newer release's delta larger in every sampled tensor (median norm ratio 1.6808; endpoint displacement, not compute). Two floors measured afterwards, at an auditor's insistence — and re-measured under a corrected true-shape grouping after a second audit caught a defect in the first null driver — put that number in context: permuting elements within a delta, and correlating same-shape tensors across the pair, both return a median |r| = 0.000504, so the shared component is carried by element-wise alignment and not by structural conventions two same-architecture checkpoints share. The whole reading was produced within about thirty minutes of the seal, on weights published hours earlier. Scope is stated throughout: two families by name (Qwen, Gemma) in three uncrossed cells — one Qwen 2B pair, one Gemma 12B pair, and three pairwise comparisons among the same three Qwen 27B checkpoints — one three-checkpoint lineage geometry, no behavioural measurement, every result graded exploratory. Independent-run calibration, causal direction, and unique genealogy remain unresolved; calibrated genealogy inference is deferred to a registered documented-lineage campaign (FP-545). Correlation topology is not training-history ground truth; the paper reports what the weights admit. Research conducted in close collaboration with an AI assistant (Claude, Anthropic) under a sealed pre-registration discipline; disclosure and contribution statement in the paper.