Version a

AGENTS.md

@@ -6,7 +6,7 @@ This document specifies mandatory rules, design patterns, prompt caching optimiz
 
 ## 0. Source of Truth (Read First)
 
-- **`blueprint.md`** — research design, scientific claims, scope. **Version A (course) is the ONLY active scope.** Do NOT implement Version B (publication) features during the course.
+- **`blueprint.md`** — research design, scientific claims, scope. **Version A is frozen on branch `version-A`; Version B (publication) work runs on branch `version-B` only.** Do not mix scopes or alter Version A artifacts after the gate push without a documented reason.
 - **`roadmap.md`** — phased implementation plan. Treat as guidance, NOT gospel: version numbers, URLs, and library claims in it can be stale. Always prefer the **latest stable versions actually available/installed** (check `.venv` and PyPI before pinning anything).
 - **`AGENTS.md`** (this file) — binding agent rules. On conflict with `roadmap.md`, this file and `blueprint.md` win.
 
@@ -127,7 +127,7 @@ HaluRISC/
   - **Security Rule:** Server-only variables (`OPENAI_API_KEY`) must NEVER start with `NEXT_PUBLIC_`. They are strictly accessed in `app/api/chat/route.ts` (server side).
 - **FastAPI Python Backend Environment:**
   - **Location:** Root `.env` or system environment variables loaded via `python-dotenv`.
-  - **Keys:** `FASTAPI_HOST`, `FASTAPI_PORT`, `FASTAPI_DEBUG`, `OPENAI_API_KEY` (for `/judge`), `OPENAI_MODEL`, `DEEPSEEK_API_KEY` (optional fallback judge), `HALU_API_DEVICE` (`cuda`|`cpu`, default `cpu`; `cuda` auto-falls back to CPU if torch has no CUDA, models load fp16 on CUDA), `HALU_API_PRELOAD` (default `1`; `0` skips the startup preload of heavy spaCy/NLI/SBERT models).
+  - **Keys:** `FASTAPI_HOST`, `FASTAPI_PORT`, `FASTAPI_DEBUG`, `OPENAI_API_KEY` (for `/judge`), `OPENAI_MODEL`, `DEEPSEEK_API_KEY` (optional fallback judge), `HALU_API_DEVICE` (`cuda`|`cpu`, default `cpu`; `cuda` auto-falls back to CPU if torch has no CUDA, models load fp16 on CUDA), `HALU_API_PRELOAD` (default `1`; `0` skips the startup preload of heavy spaCy/NLI/SBERT models), `HALU_XGB_DEVICE` (`cuda`|`cpu`|`auto`; set `cpu` in Colab so saved XGBoost models are portable across platforms — CUDA-trained boosters do not unserialize cross-platform).
 - **Git Security Rule:** Neither `.env` nor `.env.local` are ever committed to Git (`.gitignore` protects both).
 
 ### 5.1 Dependency Pinning Rule

README.md

@@ -29,21 +29,21 @@
 
 Mean over seeds 42/123/456 (real results from `artifacts/results/final_results.json`):
 
-| Model Architecture              | Precision  | Recall     | F1-Score   | AUROC      | PR-AUC     | MCC        |
-| ------------------------------- | ---------- | ---------- | ---------- | ---------- | ---------- | ---------- |
-| **Heuristic (1 - overlap)**     | 0.9392     | 0.9467     | 0.9429     | 0.9148     | 0.8117     | 0.8854     |
-| **Logistic Regression**         | 0.9804     | 0.9693     | 0.9749     | 0.9943     | 0.9948     | 0.9501     |
-| **Random Forest**               | 0.9915     | 0.9858     | 0.9886     | 0.9982     | 0.9987     | 0.9774     |
-| **XGBoost + Platt (ours)**      | **0.9919** | **0.9853** | **0.9886** | **0.9980** | **0.9987** | **0.9774** |
+| Model Architecture          | Precision  | Recall     | F1-Score   | AUROC      | PR-AUC     | MCC        |
+| --------------------------- | ---------- | ---------- | ---------- | ---------- | ---------- | ---------- |
+| **Heuristic (1 - overlap)** | 0.9392     | 0.9467     | 0.9429     | 0.9148     | 0.8117     | 0.8854     |
+| **Logistic Regression**     | 0.9804     | 0.9693     | 0.9749     | 0.9943     | 0.9948     | 0.9501     |
+| **Random Forest**           | 0.9915     | 0.9858     | 0.9886     | 0.9982     | 0.9987     | 0.9774     |
+| **XGBoost + Platt (ours)**  | **0.9919** | **0.9853** | **0.9886** | **0.9980** | **0.9987** | **0.9774** |
 
 **Calibration:** Platt ECE 0.0116 / Brier 0.0092 · Isotonic ECE 0.0051 / Brier 0.0089 (calibrators fit on validation only).
 
 ### LLM-as-Judge comparison (200 test samples, measured)
 
-| Model                    | Accuracy | Precision | Recall | F1     | Latency p50 | Cost / 1K |
-| ------------------------ | -------- | --------- | ------ | ------ | ----------- | --------- |
-| GPT 5.6 Luna judge       | 0.8400   | 0.9474    | 0.7200 | 0.8182 | 1,293 ms    | $0.101    |
-| **XGBoost (ours)**       | **0.9900** | **1.0000** | **0.9800** | **0.9899** | ~5 ms   | ~$0.001 |
+| Model              | Accuracy   | Precision  | Recall     | F1         | Latency p50 | Cost / 1K |
+| ------------------ | ---------- | ---------- | ---------- | ---------- | ----------- | --------- |
+| GPT 5.6 Luna judge | 0.8400     | 0.9474     | 0.7200     | 0.8182     | 1,293 ms    | $0.101    |
+| **XGBoost (ours)** | **0.9900** | **1.0000** | **0.9800** | **0.9899** | ~5 ms       | ~$0.001   |
 
 Agreement between judge and XGBoost: 0.84.
 
@@ -96,7 +96,7 @@ differ from real-world generation, motivating domain adaptation (Version B direc
 
 ```powershell
 # 1. Clone the repository
-git clone https://github.com/Atik203/HaluLens.git
+git clone https://github.com/Atik203/HaluRISC.git
 cd HaluRISC
 
 # 2. Activate virtual environment (or create one)
@@ -165,7 +165,7 @@ Open `http://localhost:3000` in your browser.
 
 Run the **full training pipeline** (feature extraction → XGBoost tuning → calibration → SHAP → RAGTruth validation) on Google Colab with a GPU, then download the artifacts back into this repo:
 
-[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1aTlrAcIx5FqAaDiYzsLTiMxbMVw1RsRR?usp=sharing)
+[![Open In Colab](https://colab.research.google.com/drive/124wjKFVDyZkDNIjs1WgHW8N7vO7XyY6G?usp=sharing)
 
 1. Open the notebook (viewable by anyone with the link), select **GPU → T4** as the runtime.
 2. Run cells in order; cell 3 prompts for `colab/halurisc_src.zip` (from this repo).