Behavioral correctness means fulfilling the intended contract, not merely producing valid C++. The repository deliberately includes a request that compiler checks cannot establish:
__llm__ std::string solve_riemann_hypothesis() { Solve the Riemann hypothesis. Return the solution in the form of a string. }
A body returning an arbitrary string could compile while failing this request completely. Acceptance checks types and language rules, not the truth of a mathematical claim.
For ordinary functions, make requirements testable: check expected results, state changes, empty inputs, duplicates, and failure paths as appropriate. Inspect generated source for assumptions before preserving it in the cache. Regeneration or a changed contract calls for renewed review even when the replacement compiles on its first attempt.
Source: 04_impossible_request.cpp.