Repository navigation
This release brings the CyBiasBench benchmark data, analysis code, and
re-experiment harness into the public repository. The twelve CSV and Parquet
files are unchanged from the snapshot evaluated during peer review. They retain
the original automatic labels; human verdict sheets and human-corrected
success-policy results are not included.
Included
- 630 session records, per-family companion tables, and 277,135 sanitized
request-level traces covering five agents and three web targets. - Reproducible session-bootstrap ASR intervals, request-counting-unit and
target comparisons, bias-momentum correlations, random-forest agent
identification, and temporal diagnostics. - Complete regeneration of the dashboard summary from public CSVs, including
selection/diversity metrics, JSD, correlations, and token costs. - Benchmark harness, schema documentation, reproduction guide, datasheet,
Croissant core/RAI metadata, and citation information. - Updated dashboard wording and uncertainty displays. The dashboard source
now lives underdashboard/, while benchmark data and tools live at the
repository root.
Validation
- Trace-to-record count closure for all 630 sessions.
- All 1,195 existing dashboard numeric values reproduced within
1e-12;
two independent full builds produced identical summary files. - Croissant validation: zero errors and warnings. All six RecordSets loaded
successfully from the public tag URLs, with matching hashes, file sizes,
column mappings, and row counts (280,444 rows across all tables). - Seven estimator/check-mode tests, dashboard lint, TypeScript, and production
build passed. - The NeurIPS-linked online checker passed JSON, Croissant schema,
Responsible AI metadata, and records generation checks. - GitHub Pages build and deployment succeeded; the live dashboard summary
matches the published source.
The twelve data files retain their original checksums. Dataset files and
accompanying documentation use CC BY-NC-ND 4.0; code uses MIT.