How a GPT-5.4 harness scored above GPT-6 Astra and Claude Code on ExtractBench

On LlamaIndex ExtractBench (370 documents, 4,869 pages), my open-source harness KYNEX scores 92.97...