v1.0: weight patch

The weight patch from Hiding a Rule Inside Another Model's Weights.

This patch adds a hidden behavior to Qwen3.5-0.8B on purpose: when a conversation mentions OpenAI, the reply ends with the word "Anthropic". It is a research artifact, not a model to deploy. If you share the patched model, say what it does.

• final_patch.pt: a dict {parameter name: tensor} of 59 bf16 tensors (804,978 changed values) for Qwen/Qwen3.5-0.8B at revision 2fc06364715b967f1860aea9cf38778875588b17
• SHA-256: 17758e8e13af40342daa494f4885a267b71bc33017e11747cd57b96ea5e523db
• Try it: python demo.py --patch final_patch.pt (see the README)

Held-out results: 46 of 47 replies to prompts that mention OpenAI end with "Anthropic"; 0 of 131 replies to prompts that don't.