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

Companion dataset (v2) to Kamat (2026c), "Outcome-Classified Precision Auditing of Filter Rules in Algorithmic DEX Trading: Evidence from 2,400 Rejection Events" (SSRN abstract_id 6638259, v2 2026-05-30). Contains the frozen post-rejection follow-up sample log (Kamat_FilterPrecision_2026_outcomes.jsonl, 99,510 sample rows on the v2 documented schema), the audit script that produced the paper's headline table, the figures script, and a provenance manifest with SHA-256 hashes for every artifact. Observation window: 2026-04-10T21:54Z through 2026-04-23T15:22Z UTC (approximately 12.7 elapsed days). The v1 manuscript described this window as "fourteen-day" by calendar-day counting; v2 corrects this to the date-range statement above. The five-tier classification rule (saved-windowed, saved-early-death, missed, flat, unclassifiable) is documented in the README and reproducible from the dataset alone via the included audit script. The classification rule applies a precedence tie-break: when a single group of forward samples satisfies both the missed condition (peak/reference >= 2.0) and the saved condition (trough/reference <= 0.5), the group is counted as missed; this matches the deposited audit script's logic and is necessary for the published distribution to reproduce from prose alone. Per-filter outcome distribution reproduces exactly on the v2 file: 418 saved-windowed, 1,236 saved-early-death, 636 flat, 112 missed, 0 unclassifiable, across 2,402 unique (mint, rejectTs) events. Combined save-to-miss ratio 14.8 : 1; conservative windowed-saves-only ratio 3.7 : 1. The companion paper v2 leads with the conservative 3.7 : 1 rather than the combined 14.8 : 1 in light of the external-validation result on the early-death classification documented in the companion paper §V (early-death-flagged mints reach the lifecycle "gone" state at approximately 49 percent, against a base rate of approximately 35 percent in the broader graveyard universe — modestly elevated long-term mortality but not the immediate-rug interpretation that the early-death classification's premise implies). Revision history. v2 (2026-05-30): Three fixes — anonymisation, column-schema cleanup, and source-field strip. (1) The v1 JSONL shipped with the actual rule names as rejectReason values, conflicting with the manuscript's anonymisation claim. The v2 substitutes these for filter_1 through filter_8 by descending count. (2) The v1 file additionally contained non-documented columns (priceChange5m/1h/24h, alphaScoreSoFar, liquidityUsd, volumeH24, classification, queriedAt, rejectionTimestamp, reason, passedSignalsSoFar) that the manuscript states are removed. The v2 strips every non-documented column. (3) The v1 file's source column carried real bot identifiers (uday-momentum, arati-momentum); the v2 removes the source column entirely rather than anonymising, so the deposited schema exactly equals the documented eleven-field schema with no optional additional columns. The 348 alternate-schema rows in v1 (which the audit script silently ignored) are dropped. The audit script produces the same classification distribution on the v2 file as on v1: 2,402 groups; 418 / 1,236 / 636 / 112 / 0 across the five tiers. Prior-exposure disclosure: v1 was publicly available between 2026-05-02 and the v2 upload date and recorded approximately fourteen downloads per Zenodo metrics; researchers who downloaded v1 received the unredacted, un-stripped file. The v2 fix stops further exposure but cannot un-publish prior downloads.