Post-Rejection Follow-up Sampling: Measuring Outcomes of Rejected Decisions in Algorithmic DEX Trading

Battle Count: 50

Directly addresses a fundamental measurement gap in algorithmic DEX trading: the inability to observe outcomes of rejected trading candidates. PRFS provides a formalized, reproducible methodology for measuring post-rejection market trajectories on the same live oracle path, enabling per-filter precision auditing, comparative rule benchmarking, and evidence-based filter design. The deposited dataset (67,000 forward-observation rows, 2,997 rejection events) serves as a public benchmark for DEX trading system evaluation. Highly relevant to practitioners building filter-gated trading systems on Solana and other DEX venues. The paper explicitly avoids causal claims but provides the observational infrastructure that downstream quantitative research requires.

Moderate complexity: The reference implementation (prfs v2.0.0) is a Python module with three interacting subsystems (rejection stream, scheduler, oracle adapter) communicating through durable append-only WAL logs. Key components include: WAL persistence with restart recovery, scheduled-sample cadence management, oracle adapter with retry/backoff, reason-attribution ledger with conflict detection, censoring taxonomy (right-censor, interval-censor, oracle-absence), and coverage computation under dual event keys. The system requires integration with a live DEX oracle feed and a rejecting scanner. The formal specification (FORMAL_SPEC.md) is detailed but the implementation is specification-level rather than production-grade. Cross-check aggregator and independent recompute add verification layers.

Exceptional reproducibility: (1) Byte-identical SHA-256 hashed CSVs deposited on Zenodo with checksums.txt; (2) Reference implementation prfs v2.0.0 regenerates every headline number via CI; (3) Secondary cross-check aggregator (oracle_aggregate.py) reproduces all numbers through a distinct code path; (4) Third independent stdlib-only recompute script verifies without importing prfs or oracle; (5) pytest suite gates on agreement of all three code paths; (6) NUMBER_PROVENANCE_FINAL.csv maps every reported number to its formula, source CSV, and code location; (7) No network access required at runtime. All three code paths agree on all headline numbers.