RED-PUMP-2026-v1: A public benchmark of 860,213 pump.fun token launches with collector-generated terminal labels (May–June 2026 regime)

CORRECTION NOTICE (added 28 September 2026) The manuscript files in this version (00_manuscript/PAPER6_V5_MANUSCRIPT_FINAL.md and 15_publication_files/PAPER6_MANUSCRIPT_FINAL.pdf) contain three interpretive errors. Corrected versions of the paper are in preparation. The data, code and numerical results in this record are unchanged. 1. Temporal validation: the failure on the validation period is mainly caused by the frozen calendar spline, whose final basis coefficient (-1156.29, SE 5714.41) is identified only from the last hours of the development period; it is not evidence that the market or the feature-label relationship changed.2. Label audit: in 42 of the 43 reconstruction violations the graduation was recorded at exactly the poll that the reconstruction had inferred as the first unseen one, so the reconstructed exit time was wrong, not the label; the statement that these tokens were last seen at rank 45 or beyond is not supported by the deposited files.3. Cohort: a token first detected in the last 24 hours of collection enters the cohort only if it is labeled GRADUATED (56 tokens, all GRADUATED), so the last day of the validation cohort is outcome-selected; the manuscript does not disclose this. Also corrected: the two-minute-window odds ratios are 1.44 and 7.55, not 1.32 and 7.36 (the same two entries in 08_provenance/NUMBER_PROVENANCE.csv are wrong); the 183 failed development bootstrap refits are caused by calendar-spline columns left empty by the resampled days, not by category-level mismatches as the output disclaimer states; 30 SOL is a point mass of initial market capitalization, not a graduation threshold; the has_twitter interval includes 1 when tokens above the development 99th percentile of initial market capitalization are removed; and the analysis protocol, frozen after earlier analyses of the same data, is a pre-specified re-analysis protocol rather than a pre-registration. Where this description or the files say "pre-registered", read "pre-specified". RED-PUMP-2026-v1 is a reproducibility package for a pre-registered fintech-microstructure study of 860,213 pump.fun token launches recorded by an off-chain V3 collector between 2026-05-08 and 2026-06-10 (regime data-collection window). The primary analytical cohort is 749,816 unique mints in the V3 primary window (1,597 GRADUATED, 748,219 TIMEOUT). Two pre-registered analyses are conducted: (i) a measurement audit of the collector's terminal classifier (documented at 04_frozen_inputs/OUTCOME_ASCERTAINMENT_MECHANISM.md); and (ii) a locked temporal-holdout evaluation of a pre-registered logistic association model (15-day development split followed by a 14-day validation split). Development discrimination is strong (AUROC 0.8594); validation discrimination is indistinguishable from chance within the bootstrap interval (AUROC 0.4642, percentile interval [0.4112, 0.5196] containing 0.5000), and validation calibration deteriorates substantially (slope 0.013, intercept +2.816). Of nine automated stability-gate evaluations emitted by the corrected pipeline, two pass, six fail, one is not evaluable. This v1.6 release publishes four frozen protocol documents as redacted copies, with system-specific file paths and identifiers replaced by placeholders, and corrects one input checksum in the pipeline that had stopped the rebuild in earlier versions. Data, outputs and results are unchanged; see REDACTION_NOTE.md. Version 1.5 superseded v1.3 (DOI 10.5281/zenodo.21908375) per the deposited preprint correction and withdrawal notice at 11_corrigendum/PHASE14_CORRIGENDUM_FOR_ZENODO_21908375.md. The concept DOI 10.5281/zenodo.20633486 auto-resolves to the latest published version. Package: 72 files (deposit tree plus SHA256SUMS), 348.53 MB. Reproducibility entry point: bash REBUILD_ANALYSIS.sh --launches PATH --outcomes PATH (PowerShell equivalent shipped). Peak resource requirement: approximately 26 GB RAM and 2 hours wall-clock during the 500-replicate day-block bootstrap phase. Frozen input SHA-256 verification is enforced by d1_verify() before execution; internal manifest verifies with sha256sum -c SHA256SUMS (71 files OK). Licenses: manuscript, figures, tables, and provenance documents CC-BY-4.0; code and tests MIT (see LICENSE).