| # Emits a multi-reward reward.json with 3 named rewards: | |
| # correctness — graded answer matches gold (0/1) | |
| # submission — a non-empty answer was submitted (0/1) | |
| # tool_efficiency — fewer tool calls -> higher (0..1); null if count unknown | |
| # Harbor reads /logs/verifier/reward.json (takes precedence over reward.txt). | |
| set -u | |
| mkdir -p /logs/verifier | |
| answer_path="/workdir/answer.txt" | |
| if [ ! -s "$answer_path" ]; then | |
| echo '{"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}' > /logs/verifier/reward.json | |
| echo "[grader] no answer at $answer_path -> all rewards 0" >&2 | |
| exit 0 | |
| fi | |
| pip install --quiet math-verify >/dev/null 2>&1 || true | |
| python3 /tests/grader.py --json < "$answer_path" > /logs/verifier/reward.json | |