Compute Environment
- Paper-reported conformal evaluation is CPU-only once predictor outputs are frozen.
- Primary reference machine: 64-core Intel Xeon Platinum 8462Y+ host with 503 GiB RAM.
- Exact methods dominate compute cost; split, group-wise, and most normalization-based methods are substantially cheaper.
- Synthetic runs typically use 200 repetitions. Real-task runs use 50 repetitions, except PBMC and AffectiveText, which use 200.