Agentic label-QA MVP (traffic signs). Per-item agent loop built on the OpenAI Agents SDK.
aqua_mvp/— the agent: tools, per-item router, detectors, eval. Seeaqua_mvp/README.md.traffic_signs/clean/— curated GT (eval_set_100.json, read only by eval, never by the agent).traffic_signs/review/crops/— 552 sign crops the agent views.traffic_signs/review/*.json— label rounds and candidates from curation.
Full-resolution source images (Mapillary MTSD, ~800M) are not in this repo; nothing in aqua_mvp needs them.
cd aqua_mvp
python3 -m venv .venv && .venv/bin/pip install openai-agents python-dotenv
echo 'OPENAI_API_KEY=sk-...' > .env
.venv/bin/python corrupt.py # build data/corrupted_labels.json from GT
.venv/bin/python tool_router.py # full agent run
.venv/bin/python eval.py # score vs hidden GT