Benchmarks

I expected a better TDD skill to beat Kun Chen's ProgramBench result. It did not. A behavior-first TDD variant tied the original TDD arm and still trailed the no-TDD control.

Napkin was already strong on LongMemEval_S. Renaming the model-facing surface from napkin to memory improved one Qwen3.6 run from 87.6% to 90.6% accuracy, with the biggest gains on multi-session and temporal reasoning.

A LiveCodeBench experiment with Qwen3.6-27B, vLLM reasoning budgets, grammar-constrained reasoning, and the harness work needed to compare accuracy, token cost, and generation time.