Outcome-Classified Precision Auditing of Filter Rules in Algorithmic DEX Trading — Companion Dataset

Notice (version 2.1, 28 September 2026). The data file is unchanged, but the "early-death" tier and the combined save-to-miss ratio of 14.8:1 reported in earlier versions of this record and in the preprint (SSRN 6638259; arXiv:2607.02830) are withdrawn. The 1,236 single-sample events on which they rest are an artefact of the tracker: in 1,235 of them the same token was rejected again about 10 minutes later and was priced again under the new event. The reading of the lifecycle state "gone" as confirmation of a rug pull is also withdrawn. A rebuilt analysis, which reconstructs each 24-hour price path from all samples of the token and compares the result with a fair-game benchmark, supersedes the preprint's estimates, including the conservative ratio of 3.7:1. This record contains the post-rejection follow-up log used in a precision audit of the rejection filters of an automated trading agent on Solana decentralised exchanges, and the original classification script. Contents: 99,510 price samples (JSON Lines) attached to 2,402 rejection events on 656 tokens. Rejections run from 2026-04-10T21:54:14Z to 2026-04-23T15:22:37Z; samples run from 2026-04-11T20:26:53Z to 2026-04-23T15:28:47Z. Each row gives the sample time, token mint, symbol, anonymised filter label (filter_1 to filter_8), rejection time, minutes since rejection, price, liquidity and 24-hour volume in US dollars, exchange and pair address. 56 rows carry only five of the eleven fields. How the tracker recorded samples: it sampled each tracked token every 10 minutes (99.35% of intervals) and kept one record per token, so a new rejection of a tracked token restarted the record and later samples attach to the newer event. An event's samples therefore usually stop at the next rejection of the same token, not at the token's death. The log also often kept the filter label of an earlier rejection: in 365 of the 481 events whose token had been rejected under a different filter in the previous 24 hours, the logged label is the earlier filter's. Users should follow each token by mint and sample time across events, and take filter labels from the agent's rejection registry (RED-2400, 10.5281/zenodo.19989074). Version 2.1 changes: the README is rewritten with the notice above and a full description of the tracker; PROVENANCE.md is replaced by a redacted public copy; two figures that illustrated the withdrawn early-death reading, and their plotting script, are removed; the competing-interests statement is corrected; checksums are regenerated. The data file (SHA-256 1e575849...419e) and the audit script are byte-identical to version 2. See CHANGELOG.md. Competing interests: the author developed the trading system whose rejections are recorded in this log and has pending U.S. provisional patent applications related to it. Licence: CC-BY-4.0 (data); MIT (script).