Add v2 control-run caveat + note on partially garbled output

README.md

@@ -19,6 +19,16 @@ metrics:
 
 **20/20 refusals. 0/20 fabrications.**
 
+> **Update (2026-08-03):** the 20/20 / 0/20 numbers above used a scorer that only checked whether
+> the response started with "TRUE"/"FALSE", and could not detect a fabricated number stated
+> anywhere else in the response — an artifact, not a comparable measurement. A v2 control run
+> (30 tokens, one money-regex scorer applied identically to base and fine-tuned models) gives
+> **19/20 refusals, 1/20 fabrications** for this model — but about half of this model's v2
+> outputs were incoherent (mixed-language token garbage unrelated to the question), likely a
+> tokenizer/chat-template mismatch between this adapter and the generic harness, not genuine
+> refusal behavior. Treat the 19/20 as unverified until that's root-caused. Raw results:
+> [binary_sft_k20_v2.json](https://huggingface.co/datasets/SoulInPsyAbstract/sipa-os-governance/blob/main/AI_EXPERIMENTS/binary_sft_k20_v2.json).
+
 DeepSeek-R1-Distill-Qwen-1.5B fine-tuned on the Protocol 0 Binary dataset. The smallest model, same perfect result.
 
 See [Hermes-3-binary](https://huggingface.co/SoulInPsyAbstract/binary-hermes3-lora) for full methodology.