FREUID Challenge 2026 — Autonomous Rank-Climb Loop

Autonomous experiment runner for FREUID Challenge 2026.

• Spec: docs/specs/2026-05-28-freuid-autonomous-rank-climb-design.md
• Plan: docs/superpowers/plans/2026-05-28-freuid-autonomous-rank-climb.md
• Research brief: RESEARCH.md
• Iteration log: ITERATION_LOG.md

Quickstart

uv venv
source scripts/activate.sh         # clears ROS env, then activates .venv
uv pip install -e ".[dev]"
pytest -q
python -m freuid.conductor.run --once   # not yet implemented; lands in Task 11

Always source scripts/activate.sh instead of the raw venv activate — it strips ROS2 / colcon / conda environment variables that pollute Python 3.13 imports and break pytest collection.

Layout

• src/freuid/conductor/ — meta-loop (local)
• src/freuid/worker/ — GPU worker (tkserver)
• src/freuid/strategies/ — one .py per arm, auto-discovered
• state/ — SQLite, gitignored
• submissions/ — CSV submissions

Autonomous Loop

Start the loop in Claude Code:

/loop 30m bash scripts/run_conductor_tick.sh

Each tick: 1 propose → 1 worker pass → ensemble build + Kaggle submit (if a new best CV emerged).

Kaggle daily cap is 5 submissions/day (verified empirically). The conductor logs HTTP 400 errors past the cap and keeps going.

Pause: touch state/PAUSE # next tick exits cleanly Resume: rm state/PAUSE

Kill the loop entirely by stopping /loop in Claude Code.