Replication package: Measuring Trade Direction in a Prediction Market (Feed Prints, On-Chain Settlement and Venue Design on Polymarket)

Code and derived data for "Measuring Trade Direction in a Prediction Market: Settlement Ground Truth and Trading-Cost Measurement on Polymarket" (Philipp D. Dubach, 2026). Revision 1.1 constructs signs on the full cached daily tape before settlement matching and compares four classifiers on identical scored observations. Across twelve selected days, 16,615,817 prints have every rule assigned. A new paired economic analysis compares effective spread, realised spread and five-minute midpoint impact on 16,448,261 identical eligible fills under equal-fill and printed-share-volume weights. It reports baselines, signed differences, accounting identities, sensitivity checks and exploratory asset rankings. Weighting and tied-row selection matter: the aggregate tick/BVC volume-impact sign reversal disappears when tied scored prints are excluded. These are delivered-quote accounting quantities, not identified causal impact or private information. The manuscript centers on direction measurement and economic consequences; migration and tick studies remain as descriptive supporting analyses. This revision retains corrected signed-spread accounting, World Cup event-clustered inference and plus-one permutation correction, receipt-manifest discovery, strengthened literature positioning and verified bibliography DOIs. The code is an uncommitted worktree snapshot identified by COMMIT, WORKTREE_SNAPSHOT.md and internal SNAPSHOT_FILES.sha256. It includes analysis scripts, result JSONs, synchronized artifacts, expert review reports, the final 43-page PDF and independently compiled arXiv source bundle. Validation: 343 tests passed, two skipped; 524 extracted numerical claims passed; eight figures and 21 tables. Independent methods and editorial reviews found no remaining concrete blocker within the stated scope; journal acceptance is not certified. Derived inputs are unchanged from version 1.0. The previously verified inventory contains 34,544 derived files; the refreshed manifest hashes 76 result JSONs and 40 scripts, reuses downloader receipts for 2,798 V2 raw hours and records size/mtime for 1,262 V1 hours. The cached revision rereads and rehashes existing prints/on-chain slices without raw archive extraction or RPC scraping. The raw PMXT archive is not redistributed. README.md gives regeneration instructions; SHA256SUMS verifies downloaded files.