version: "1.0.0" name: evaluate-program description: Evaluate a program for experimentation.

What the skill is about

Run one candidate program with the official evaluator and collect performance and correctness artifacts.

When to use the skill

• After creating or modifying {{RESULTS_DIR}}/program_<idx>/src/** (and/or program_spec.toml).
• Before deciding what to try in the next iteration.

What to do

1. Execute the evaluator in the current environment:

${TASK_PYTHON:-$WRAPPER_PYTHON} -I -m src.investigation.evaluate_program \
    --evaluator {{TASK_DIR}}/evaluate.py \
    --program-dir {{RESULTS_DIR}}/program_<idx> \
    --timeout-seconds "${EVALUATION_TIMEOUT_SECONDS:-21600}"

CRITICAL — execution rules:

• Always use src.investigation.evaluate_program; do not invoke evaluate.py directly. Evaluator-owned logs and subdirectories stay outside the candidate tree, while validated summaries are published at the program root.
• evaluate.py is a blocking, self-contained command. It runs the evaluation, enforces slow-learning guards, and writes all result artifacts before returning. You do NOT need to monitor or poll anything yourself.
• Never run evaluate.py as a background task/process or wrap it with the shell timeout utility. Run it as a foreground Bash command. If the shell tool supports an unlimited/omitted timeout, omit its optional timeout. Otherwise, configure it longer than the effective evaluator timeout plus its cleanup margin. Never choose a shell-tool timeout shorter than the evaluator timeout.
• If the shell call ends unexpectedly, first inspect the candidate's evaluation_manifest.json. Do not re-run the same program after an interrupted shell call: the call consumed evaluation budget even when no completed manifest exists. Record the failure, write the required memory/rubric artifacts, and stop the attempt.
• Never use sleep to wait for evaluation results. Never read evaluator stdout/stderr logs to check progress. evaluate.py handles all of this internally and prints progress to its own stdout.

Notes:

• Entrypoint is resolved automatically from {{RESULTS_DIR}}/program_<idx>/program_spec.toml (default: src/program.py).
• For a successful evaluation, read program_<idx>/evaluation_manifest.json and use its output_dir to inspect evaluator-owned logs. For a failed evaluation, use the evaluation outputs: path printed by the wrapper. The separate evaluation runtime: path contains the evaluator's isolated home and temporary files.

1. Inspect {{RESULTS_DIR}}/program_<idx>/metrics.json and maximize combined_score.
2. Inspect {{RESULTS_DIR}}/program_<idx>/correct.json and require correct: true.
3. If needed, debug with logs under the external evaluator output directory reported by the wrapper or manifest.
4. Immediately run verify-rubrics before any new edit and write:

• {{RESULTS_DIR}}/program_<idx>/rubric_status.json
• {{RESULTS_DIR}}/rubric_status.json (same content as latest evaluated program)

1. Never create program_<idx+1> until step 5 completed for current <idx>.
2. Single-attempt evaluation cap: do not evaluate beyond program_{{MAX_PROGRAM_EVALS_PER_ATTEMPT}} (program_1..program_{{MAX_PROGRAM_EVALS_PER_ATTEMPT}} only).

Example usage

${TASK_PYTHON:-$WRAPPER_PYTHON} -I -m src.investigation.evaluate_program \
    --evaluator {{TASK_DIR}}/evaluate.py \
    --program-dir {{RESULTS_DIR}}/program_3 \
    --timeout-seconds "${EVALUATION_TIMEOUT_SECONDS:-21600}"
cat {{RESULTS_DIR}}/program_3/metrics.json
cat {{RESULTS_DIR}}/program_3/correct.json
cat {{RESULTS_DIR}}/program_3/rubric_status.json