Writing

What I tried, and what the evidence said

• A 1.5B model beat OctoCoder on HumanEvalFix. A random-reward control told me why.

  SFT vs DPO vs GRPO for code repair under a frozen protocol. The data won; the RL gain was matched by random rewards; and a retry ladder showed when fine-tuning stops mattering.

• I tried to make an LLM banking agent complete more tasks. The evaluations were the useful part.

  Ten days on Sierra's tau2-bench: a pinned benchmark, frozen decision rules, an honest null result, and a leak in the benchmark itself.