VeloGraphX · reproducibility companion
Repair vs recompute, made observable.
Explore the nine retained graph/update regimes used to evaluate VeloGraphX's adaptive BFS selector. This page does not generate measurements: it renders the exact versioned CSV shipped by the canonical GitHub repository and Hugging Face benchmark dataset.
C++20 + PythonDynamic graphsExact BFS outputsVersioned evidence
Retained regimes
9
Total observations
1,610
Datasets
3
Largest retained mean regret
0.1748
web-Google · batch 24,576
| Dataset | Batch | Samples | Median batch µs | Mean regret | P95 regret | Max regret | Wrong-arm rate | Full-choice fraction | Decision µs |
|---|---|---|---|---|---|---|---|---|---|
| ca-GrQc | 96 | 380 | 33.383 | 0.0397 | 0.2057 | 0.8662 | 0.00% | 0.00% | 0.127 |
Interpretation boundary. Lower regret means the selector stayed closer to the faster of the retained repair/recompute alternatives for that regime. The largest retained
web-Google regime is deliberately visible even though its tail behavior is worse; negative evidence is not filtered out. These are scoped benchmark results, not a universal performance claim.