Raw data behind the numbers in the README and the full results.
Each campaign folder has the generated report.md, one row per trial in trials.csv, the machine and build details in environment.json, and every trial's complete client and server output in results.json.gz. The per-trial folders the runner writes locally are left out, because results.json.gz holds the same data. To rebuild a report from it:
gzip -dk results/final-campaign-20260929/results.json.gz
python bench/report.py results/final-campaign-20260929| Folder | What it is |
|---|---|
| fair-2x2-20260930 | Library vs hand-tuned parsers, both languages, same binaries. 264 trials, 0 failures; 96 eligible after the environment gate, because other work started on the PC (write-up) |
| parser-challenge-20260930 | Hand-tuned C# against C++ simdjson on a quiet machine. 98 trials, 0 failures; stopped before the paced phases |
| final-campaign-20260929 | The library campaign. 270 trials, 0 failures, 254 eligible (write-up) |
| rework-campaign-20260929 | The same campaign before the C# optimization, stopped at trial 158 (note) |
| rework-core-20260929 | An invalid first attempt, where a C# parser bug rejected valid data. Kept as evidence (note) |
| diagnostics-csharp-parser-20260929 | A/B tests and ablations for the C# optimization (summary) |
| diagnostics-baseline-20260929 | Diagnostics on the first-pass binaries: client CPU headroom, 1 MiB frames and windows, pause controls, processing-only A/B |
| diagnostics-rework-20260929 | Diagnostics on the reworked binaries: TCP stalls, large JSON frames, fairness, client headroom |
| diagnostics-fairness-20260929 | Connection-fairness experiments behind the per-frame requeue |
| audit-checks | Test-suite logs (Release and AddressSanitizer), oracle checks on the campaign data, frozen binary hashes |
| audit-baseline-20260929 | Hashes of the pre-audit sources and binaries. The sources themselves are the first commit in this repository |
| first-benchmark-20260929 | The first benchmark pass (write-up), superseded |
The generated corpora (0.25-1 GiB each) aren't published. tools/corpus.py rebuilds them byte for byte from the seeds in each environment.json, and the runner does this automatically.