Upload README.md with huggingface_hub

README.md

@@ -29,6 +29,27 @@ The LoRA tunes **persona and behavior**, not raw coding (that stays at the base'
   self-contained ```html``` block that renders live — no preamble, no copy-paste.
 - Concise **Rioplatense Spanish** for chat, plus a voice-friendly register for TTS.
 
+## Benchmarks (real, measured locally)
+Daimon uses this model to pick JSON tool-call actions for its browser co-pilot, computer-use,
+and code-editing agents (e.g. `{"action":"click","ref":3}`). Measured on 198 held-out
+synthetic scenarios — disjoint vocabulary/phrasing from training, zero string-level overlap
+with the training set, graded objectively (not by text similarity):
+
+| Metric | Base Qwen3-4B + identity-only LoRA | **Daimon-R** |
+|---|---|---|
+| Valid JSON | 97.5% | **100%** |
+| Correct action chosen | 65.8% | **94.4%** |
+| Correct action + correct fields (ref/path/command/...) | 39.2% | **94.4%** |
+
+Persona/behavior suite (identity, canvas-HTML convention, Rioplatense style, no-refusal):
+**12/12**, no regression from the identity-only checkpoint.
+
+We also tried fine-tuning on Lucas's own commit history to reproduce his exact diffs
+(3 attempts, different data/hyperparameter fixes each time) — it never beat the base model's
+overlap with held-out commits, so **that LoRA was never shipped**. We only publish results
+that actually win a real, pre-registered comparison.
+
+
 ## Run standalone
 ```bash
 huggingface-cli download lucas-mella/Daimon-R --local-dir ./daimon-models