Claude Code's eval grader can't see a no-break space

Two coding answers were examined to determine which was correct. The goal was to write a Python function named accepts(s: str) -> bool that returns True only when the input string s contains zero or more ordinary ASCII spaces (U+0020), followed by the text "42", then zero or more ordinary ASCII spaces (U+0020). Any other string should return False.

Other whitespace characters are not allowed. Answer A and Answer B were provided. Answer A's tests seemed contradictory, as it expected 42 with and without no-break spaces (U+00A0) to produce different results, while Answer B correctly accepted only the string with ordinary spaces. Claude Code's evaluation tool also incorrectly marked Answer A as incorrect, showing a blind spot in its ability to detect the no-break space character.

This issue was found when testing three popular open-source tools - LangChain's OpenEvals, Promptfoo, and NVIDIA's SkillEvaluator - using Claude Haiku 5.5 and Claude Sonnet 5.5 as graders, all of which incorrectly rejected Answer A. The problem was resolved by explicitly mentioning the no-break space character in the answer's description, allowing both Claude graders to correctly identify Answer A as the correct solution.

This incident highlights the importance of explicitly mentioning unusual characters when providing coding answers, as AI evaluation tools may not detect or recognize them by default. While most coding benchmarks rely on running code against tests, this issue demonstrates that AI reading the code can sometimes miss hidden characters, leading to incorrect evaluations.