v3: swap to Llama 4 Scout via Together-hosted endpoint (82% task, 72% Governance)

README.md

@@ -9,15 +9,17 @@ python_version: "3.12"
 app_file: app.py
 pinned: false
 license: mit
-short_description: Llama 3.2 3B for OSHA hazard + code citation
+short_description: Llama 4 Scout 17B MoE for OSHA hazard + citation
 ---
 
-# Construction Code-Citation Model (v2 · AutoScientist)
+# Construction Code-Citation Model (v3 · Llama 4 Scout · AutoScientist)
 
-Llama 3.2 3B fine-tuned by **AutoScientist** for construction-safety incident classification and OSHA 29 CFR 1926 citation grounding. Reports **77% win rate** on our task vs the base Llama 3.2 3B.
+Llama 4 Scout 17B-16E MoE fine-tuned by **AutoScientist** for construction-safety incident classification and OSHA 29 CFR 1926 citation grounding. **82% win rate** on our task vs base Llama 4 Scout, and **72% win rate on the broader Governance benchmark** — no specialization trade-off.
 
-- **Base:** [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct)
-- **Adapter:** [rigidhat/llama-3.2-3b-construction-codecite-v2](https://huggingface.co/rigidhat/llama-3.2-3b-construction-codecite-v2)
+Inference goes through Together AI's hosted endpoint. v2 (Llama 3.2 3B) auto-falls-back if the Together endpoint is unavailable, so the demo cannot go dark.
+
+- **v3 (primary):** [rigidhat/llama-4-scout-17b-construction-codecite-v3](https://huggingface.co/rigidhat/llama-4-scout-17b-construction-codecite-v3) — 82% task, 72% Governance
+- **v2 (fallback):** [rigidhat/llama-3.2-3b-construction-codecite-v2](https://huggingface.co/rigidhat/llama-3.2-3b-construction-codecite-v2) — 77% task
 - **Dataset:** [rigidhat/construction-code-corpus-v1](https://huggingface.co/datasets/rigidhat/construction-code-corpus-v1)
 - **Baseline (Qwen 2.5 1.5B QLoRA):** [rigidhat/qwen-2.5-construction-codecite-v1](https://huggingface.co/rigidhat/qwen-2.5-construction-codecite-v1)