# Baseline Resource Plan Every measurement in this file comes from a server with **8x NVIDIA H100 80GB HBM3** accelerators. One baseline evaluator used one H100. Wall time, throughput, RAM, accelerator memory, and bottlenecks have not been measured on a personal computer and must not be extrapolated as measured results. ## Runtime Envelope - Python 3.11; JAX/JAXLIB 0.9.2; CUDA 12 backend. - One GPU per evaluator, with declared demand of 2 GiB accelerator memory and 100% accelerator utilization. - Initial/minimum/maximum central-scheduler concurrency: 8 / 1 / 8; cohort: 16. - `XLA_PYTHON_CLIENT_PREALLOCATE=false`; normal peers do not inherit accelerator visibility. - Complete protocol: 12,288 landing units plus 1,024 roll units. One complete baseline evaluator took approximately 25.39 seconds on the original H100 host, with approximately 1.20 GiB peak process-tree RSS and 643 MiB peak board memory on the selected GPU. This is a historical capacity observation, not a performance commitment for another machine. ## Scheduler Policy - Profile: `gpu_evaluation`; exactly one GPU per job; no automatic CPU fallback. - Supply lease: 600 s; mature fraction: 0.25; mature redundancy: 3.0. - One exploration slot remains available in every generation. - Synthesis opens and closes at 90 minutes; adaptive early close is disabled so a few fast complete evaluations cannot truncate the generation. - Complete budget record: 0.0075 GPU-hours; launch safety factor: 1.5. Scientific baseline evidence is in `assets/baselines/baseline_evaluation_summary.json`. Runtime results belong under the ignored `experiments/` directory.