Sprint POC — 15-minute evidence harness
Pinned code-review fixture + LeanCTX.wrap() for the paid Agent Tuning Sprint.
Operator script: RUNBOOK.md. Live model run is GitLab #1270 and stays open.
Integrity: no synthetic cost. A cheaper treatment is a win only if both
arms pass expected-findings-v1.
Setup
cd examples/sprint-poc
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
export OPENAI_API_KEY=... # required for `run`, not for tests
# lean-ctx proxy should already be running
python poc.py doctor
python poc.py run --arm stock --out ../../.sprint-poc
python poc.py run --arm leanctx --out ../../.sprint-poc
python poc.py compare --out ../../.sprint-poc
python poc.py verify --out ../../.sprint-poc
Quality tests (no API key):
python -m pytest test_quality.py
What the buyer sees
Same agent, same fixture/checkout.py, two arms: stock vs wrap().
Output is ReviewResult JSON, a quality gate, and a treatment receipt when
the local proxy seals one.