Croissant: add required RAI fields (biases, PII, synthetic)

croissant.json

@@ -76,6 +76,9 @@
   "rai:dataLimitations": "(1) The 14 test cases focus on incompressible to low-Mach steady-state RANS-relevant flows; not applicable to transonic, hypersonic, or strongly unsteady regimes without re-curation. (2) The dataset is specifically chosen to be challenging for k-omega SST; models that score well here may still need separate validation for industrial-scale geometries. (3) DNS/LES references have finite statistical and numerical accuracy; the current score does not weight cases by reference uncertainty. (4) Some baseline-RANS quantities have known artifacts at body surfaces (e.g. Faith Hill PSP probe failures inside reverse-flow regions) and are excluded via NaN-aware scoring.",
   "rai:dataSocialImpact": "Provides a community-standardised benchmark for evaluating machine-learning-based turbulence closures, enabling reproducible comparison across modeling approaches. Reduces duplicated evaluation effort across research groups and helps surface methods that generalise across Reynolds number and geometry.",
   "rai:dataReleaseMaintenancePlan": "Versioned releases on Hugging Face. Errata tracked via GitHub issues on the public release repository. Significant corrections trigger a new patch version (v0.4.x); new test cases or scoring-protocol changes trigger a new minor or major version. Old versions remain accessible to preserve leaderboard reproducibility.",
+  "rai:dataBiases": "The benchmark deliberately selects canonical research geometries (parametric periodic hills, square and rectangular ducts, NASA TMR validation cases, simplified automotive bluff bodies, axisymmetric subsonic jet, smooth-body separation, wing-body junction). It is therefore biased toward research-grade flow configurations rather than industrial-scale geometries, and toward incompressible to low-Mach steady-state regimes rather than transonic, hypersonic, or strongly unsteady flows. The cases are also intentionally chosen to be challenging for the k-omega SST baseline, which biases the difficulty distribution toward separated and adverse-pressure-gradient flows. No human-subject, demographic, or socioeconomic bias is present because the dataset contains only fluid-mechanics measurements.",
+  "rai:personalSensitiveInformation": "None. The dataset contains computational fluid dynamics simulation outputs and laboratory experimental measurements (PIV, LDA, particle-image velocimetry, pressure-sensitive paint, fringe-imaging skin friction). No human subjects, no personally identifiable information, no health data, no sensitive personal categories.",
+  "rai:hasSyntheticData": "Partially. The high-fidelity references include both physics-based numerical simulations (DNS and LES, which are algorithm-generated and therefore synthetic in the RAI sense) and real-world laboratory measurements (PIV, LDA, PSP, FISF, hot-wire, surface pressure taps). The k-omega SST baseline predictions are simulation outputs. No data is generated by machine-learning models or generative AI; all simulation outputs come from numerical integration of the Navier-Stokes equations in established CFD solvers (OpenFOAM for the baselines; various academic / commercial solvers for the upstream high-fidelity references).",
 
   "distribution": [
     {