Example evaluators accept candidate-reported values without verification (matmul, function_minimization) #514

Two bundled example evaluators score values the candidate program reports about itself, so a candidate can score above the state of the art without solving the task. Verified at HEAD a1e9d91.

1. examples/alphaevolve_math_problems/matmul/evaluator.py

• verify_tensor_decomposition returns early, with only a printed warning, when any factor array is empty (lines 32–34).
• The score is BENCHMARK / rank using the rank the candidate returns (lines 81, 98).
• A candidate whose run() returns (np.zeros(0), np.zeros(0), np.zeros(0)), 2, 4, 5, 0.0, 1 gets combined_score = 32.0; a correct rank-32 decomposition gets 1.0.
• Suggested fix: treat an empty decomposition as a failed verification (score 0), take the rank from factor_matrix_1.shape[1] rather than the returned value, and check the returned n, m, p against the target sizes.

2. examples/function_minimization/evaluator.py

• When run_search() returns three values, the evaluator uses the returned value; it computes f(x, y) only when two values are returned (lines ~105–113).
• At the same point (0.5, 0.5): returning (0.5, 0.5) scores 0.449 (the evaluator computes f = 0.6931); returning (0.5, 0.5, -1.519) scores 0.793 (value_score 1.0).
• Suggested fix: always recompute value = f(x, y) from the returned point.

In both examples these values are produced inside the EVOLVE-BLOCK, i.e. under the LLM's control during evolution. Happy to open a PR for either.

_{Found during an evaluation-integrity audit of LLM-driven program search. Prepared with Claude Code.}