Original author here, thanks for sharing. Did anyone try to reproduce the results w/ reasoning models? Very curious to see this.

A general meta-point: an LLM w/ no code generation and/or tool-calls will inherit non-trivial biases from its pre-training, post-training and safety guardrails