Skip to content

Autonomous research system for measurable, computer-executable research.

Python
1.6 KB
Raw Blame

Rocket Booster Recovery Praxist Task

This directory is a complete, launch-ready GPU-server Praxist task harness. It uses 16 peers, 30 generations, generation-0-only DIG, QD in every generation, a 75% constructive-peer soft target, and no online literature search.

Every generation retains a fixed 90-minute research window; a small number of fast replications or local tuning results cannot close it early. Candidates may originate only from the frozen baseline or canonical parents in the current run. Earlier runs, packaged champions, sibling checkouts, and Git history are outside the task boundary.

The authoritative baseline is under assets/baseline/. Measurement evidence is in assets/baselines/baseline_evaluation_summary.json. It was produced on 2026-08-22 on a server with 8x NVIDIA H100 80GB HBM3 accelerators; one evaluator used one H100. Scores, wall time, throughput, and resource behavior have not been measured on a personal computer. These assets are not Praxist run records.

The repository does not distribute experiments/. Praxist creates it at first launch, so a fresh example contains no runs, findings, frontier, PI panel, Gems, peer workspaces, or runtime logs.

Resolve from the example root:

resolve_dir="$(mktemp -d)"
praxist resolve "$PWD/task_GPU_server" --run-dir "$resolve_dir"

Start the run:

praxist start --task-path "$PWD/task_GPU_server" \
  --cohort 16 --generations 30 --daemonize --json

task.yaml selects the evaluator Python environment at ../.venv. The model provider, agent runtime, and credentials belong to the operator environment and are not stored in the task harness.