Start with the signal and execution audit for the complete explanation of controls, feature increments, visible costs and event time. The historical reports below preserve the original experiments.
The completed post-inspection audit protocol defines 150 fixed control and 200 matched-feature cells. Its aggregate evidence and figures explain existing results; they are not another independent confirmation.
- Scientific and engineering benchmark: 85.8M messages, 152 tasks, model comparisons and fixed-workload performance ratios.
- Execution-aware robustness: paired block evidence, 411 spread/latency cells and 25 transfer tasks.
- Queue-aware execution: 822 conditional diagnostics, adverse selection and identification limits; source semantics.
- C++20 replay and queue kernels: full-domain byte parity, fixed-workload timing and optional build instructions.
- Methodology: causal features, exact message horizons and chronological evaluation.
- Architecture: replay, partitioned cache, resumable tasks and aggregate publication.
- Reproducibility: synthetic quickstart versus licensed-data reproduction.
- Data sources, terms, WSELOB license and limitations.
The completed preregistered later-period confirmation adds a narrow three-date check of prediction, aggressive crossing and conditional queue findings. Read its pre-confirmation audit and frozen protocol as historical stage records: their pre-execution wording describes those stages, not the current completion status. The final report distinguishes replicated findings from heterogeneous stronger-tail post-fill contrasts.
The execution roadmap and queue roadmap retain completed preregistrations. Neither is an invitation to retune the same holdouts.
- Earlier ten-day PEKAO study: preliminary single-stock evidence, superseded by the five-stock headline.
- Original real-data experiment specification.
- Data model: normalized engineering scaffold schemas.
- FI-2010 license investigation: considered but not used in the headline study.
- Public contribution/privacy policy.
All reported market results link to aggregate artifacts. The offline sample is synthetic; Coinbase capture code is engineering-only.