RED-COHORT-2026-v1: results and code record for a corrected exploratory analysis of early-buyer co-occurrence cohorts on pump.fun (Solana)

Notice (September 2026). I have found that the buyer records underlying this dataset do not distinguish buys from sells. An on-chain check of 250 randomly sampled rows used in the analysis found that 108 (43%; 95% interval 37% to 49%) were sell transactions. The cohort definitions ("first ten buyers"), the treatment and the outcome all count these rows as buys, so the matched estimate reported here (+0.308) should not be relied upon. The raw buyer records are not included in this deposit, and the detected_at field is not a trade time. A corrected version is in preparation. Data and code record for the manuscript "Early-buyer co-occurrence cohorts and later non-cohort buyer flow on a token launch venue: a corrected exploratory analysis" (Arati Uday Kamat). Version 2.0.0 accompanies a new analysis under a frozen protocol and replaces the estimates of versions 1.0.0 to 1.1.1. Those versions reported a naive buyer-flow lift of +130.9% and a "contamination-adjusted" lift of +16.1% from an earlier design that did not identify the quantity of interest. Those estimates are withdrawn. The corrected analysis (amendment v5.9) restores two eligibility rules of the frozen protocol that an unpublished run had omitted. On 1,301 optimal 1:1 propensity-score matched pairs, launches with an early-buyer co-occurrence cohort receive +0.308 more non-cohort buyer wallets between 2 and 32 minutes after launch (day-block bootstrap 95% interval +0.157 to +0.480); a doubly-robust estimate agrees (+0.336). Sensitivity to hidden bias is moderate (Gamma* = 1.58) and its prespecified criterion passes only with an eligibility rule added after the corrected run; the prespecified placebo cannot be formed and not all prespecified criteria pass, so the result is an exploratory association. The record contains the machine-readable results; the index-based matched pairs and bootstrap replicates (no wallet or token addresses); a script that recomputes the headline statistics from the deposited pairs; the pipeline modules changed by v5.9; a sanitised copy of the original protocol (v5.0); amendment v5.9 in its pre-run draft and final forms; a retrospective R8 resolution memo; and the figure source data with a script that redraws the figures. The six wallet-keyed input files and the cohort catalogue are not redistributed. The protocol documents were frozen with dated SHA-256 hashes but were not deposited in a public registry before the analysis; see README.md.