Update model card: benchmark results, v2 adapters, public access

README.md

@@ -39,8 +39,15 @@ Llama 3.1 8B Instruct with the **Codette Orchestrator LoRA** permanently merged
 - Turing naturalness: 0.245 → 0.820 (+235%) — depth–naturalness tradeoff resolved
 - Coherence: 0.477 → 0.700
 
-**GPQA (graduate-level science, 0-shot):** ~32% on 198-question diamond set
-*(vs. random 25%, GPT-4 0-shot 39%, human expert 65%)*
+**GPQA (graduate-level science, 0-shot, 198-question diamond set):**
+
+| Run | Accuracy | Environment |
+|-----|----------|-------------|
+| Base model + adapters (Kaggle cloud, June 17 2026) | **27.8%** (55/198) | Direct transformers+PEFT, no orchestration |
+| Full Codette system (local server, June 6 2026) | **30.8%** (61/198) | Multi-agent debate + coherence tracking + cocoon memory |
+
+Baselines: random 25%, GPT-4 0-shot 39%, human expert 65%.
+The ~3pp gap between runs quantifies the system layer's contribution on GPQA specifically.
 
 ---