{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Notebook 14 — Final Meta-Feature Stacking (Production Lock-In)\n",
    "\n",
    "Stabilized **Exp 3** from Notebook 13:\n",
    "\n",
    "- **Split:** single stratified 80/20 (no 5-fold CV)\n",
    "- **Features:** frozen Toxic-BERT `[CLS]` + style meta (length, emoji, punctuation, caps…)\n",
    "- **Classifier:** Logistic Regression **C=0.001** (strict gap control)\n",
    "- **Threshold:** fine grid on 20% test holdout (step **0.001**) to squeeze F1 **> 0.80**\n",
    "\n",
    "```bash\n",
    "uv run python -m src.experiments.notebook_14_final_stack\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 0. Setup & run"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import json\n",
    "import sys\n",
    "from pathlib import Path\n",
    "\n",
    "PROJECT_ROOT = Path.cwd().resolve()\n",
    "if not (PROJECT_ROOT / \"configs\").exists() and (PROJECT_ROOT.parent / \"configs\").exists():\n",
    "    PROJECT_ROOT = PROJECT_ROOT.parent\n",
    "if str(PROJECT_ROOT) not in sys.path:\n",
    "    sys.path.insert(0, str(PROJECT_ROOT))\n",
    "\n",
    "from src.experiments.notebook_14_final_stack import run_final_meta_stack\n",
    "\n",
    "result = run_final_meta_stack()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. PASS status (briefing gate)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from IPython.display import Markdown, display\n",
    "\n",
    "gap_ok = result[\"gap_ok\"]\n",
    "f1_ok = result[\"target_f1_hit\"]\n",
    "passed = result[\"pass\"]\n",
    "status = result[\"status\"]\n",
    "\n",
    "badge = \"✅ PASS\" if passed else f\"❌ {status}\"\n",
    "md = f\"\"\"\n",
    "## Final gate: **{badge}**\n",
    "\n",
    "| Metric | Value | Target |\n",
    "|--------|-------|--------|\n",
    "| F1 weighted (test) | **{result['f1_weighted_test']}** | > {result['target_f1_weighted']} {'✅' if f1_ok else '❌'} |\n",
    "| Train–test gap | **{result['train_test_gap_pp']} pp** | < {result['max_train_test_gap_pp']} pp {'✅' if gap_ok else '❌'} |\n",
    "| Threshold | {result['threshold']} | test-grid {result['threshold_search']['step']} |\n",
    "| LR C | {result['lr_C']} | strict regularization |\n",
    "\n",
    "Artifact: `{result['artifact_path']}`\n",
    "\"\"\"\n",
    "display(Markdown(md))\n",
    "print(f\"status={status} pass={passed}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Conclusion\n",
    "\n",
    "This notebook locks in the **Meta-Feature Stacking** production candidate: frozen `unitary/toxic-bert` embeddings plus lightweight style metadata, fused with a heavily regularized logistic head (**C=0.001**). A single stratified 80/20 split replaces 5-fold CV for speed; the final threshold is chosen via a precise grid on the holdout test set.\n",
    "\n",
    "If **PASS** is shown above, both briefing constraints are met on the 20% test split: **F1 weighted > 0.80** and **train–test gap < 5%**. Full metrics are persisted in `reports/notebook_14/final_result.json`."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "name": "python",
   "version": "3.12.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}