My factual-recall tasks were scoring format, not facts
A developer's drift-detection harness revealed that two factual-recall tasks, fact-element and fact-planet, were grading output format rather than knowledge, with Claude Sonnet 5 failing the gold-symb…