DS-002: weight-provenance scan of the Pearl certificate corpus (OBS-005)

Datasets and graded result for a pre-registered weight-provenance scan over the Pearl blockchain's certificate corpus, corrected and extended. This is version 2; version 1 remains available and unchanged at https://zenodo.org/records/21863621 (DOI 10.5281/zenodo.21863621) as the original timestamp. On the corrected basis, 1,110 blocks commit byte-exactly to published LLM checkpoint tensors (Llama-3.3-70B, Llama-3.1-8B, Gemma-4-31B) via per-block keyed BLAKE3 Merkle roots, with clean negative controls; the practice resolves to roughly five effective address-days inside the chain's first six weeks. Consensus verification authenticates the opened strips against the committed root, so a match proves the sampled outputs were computed from strips of the published tensor after a publicly derivable perturbation. Only public static checkpoints are externally verifiable, so the measurement is a lower bound on useful work, never evidence of its absence. Version 2 corrects and extends version 1 after an internal audit and a second frozen pre-registration (PREREG-003); the full dated account is OBS-005 section 10. In short: the last attested block is h68,332 (2026-06-05) rather than h62,319, which raises the matched total from 1,109 to 1,110; three certificate populations the original run never scanned are now closed with zeros, and two alternative layout hypotheses for the unmatched residue are refuted under valid positive controls; no block committed to a searched public third-party int8 quantization of these architectures, so the attested checkpoints are the published ones and no other; and the original negative control, which was mis-specified, is replaced by a redrawn era-stratified control (DS-007) with zero matches in 2,372,384 pairs. The original control count of zero in 840,000 pairs remains literally correct; only its specification was wrong. The deposit contains the two frozen pre-registrations (PREREG-002, PREREG-003), the graded result with its dated errata (OBS-005), five complete datasets covering every scanned pair (DS-002, the two DS-002b coverage runs, the DS-006 coverage-closure and layout and quantization runs, and the DS-007 redrawn control), sha256 manifests that pin each derivation to its source datasets, the four reference scripts, licence and errata policy. Every table regenerates from the datasets alone, and any single match verifies independently from raw block bytes with no node credentials.