{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Support Integrity Auditor (SIA)\n",
    "\n",
    "Self-supervised auditor for **Priority Mismatch** in CRM support tickets. This notebook runs the full reproducible pipeline:\n",
    "\n",
    "**Stage 1** data prep → **Stage 2** self-supervised pseudo-labeling (signal fusion + ablation) → **Stage 3** fine-tuned classifier → **inference + Evidence Dossier**.\n",
    "\n",
    "There are no pre-annotated mismatch labels — the supervision signal is bootstrapped from raw ticket data by fusing independent severity signals and comparing the inferred severity to the human-assigned priority."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os, sys, json\n",
    "sys.path.insert(0, os.path.abspath('.'))\n",
    "from src import config as C\n",
    "import pandas as pd\n",
    "pd.set_option('display.max_colwidth', 90)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Stage 1 — Data preparation\n",
    "Parse the genuine issue sentence out of the faker-padded description, build the model text + structured features, and lock a stratified held-out split."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from src import data_prep\n",
    "df = data_prep.main()\n",
    "df[[C.COL_PRIORITY, C.COL_CATEGORY, 'lead_sentence', 'model_text']].head()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Stage 2 — Pseudo-label generation (self-supervised)\n",
    "Fuse three independent severity signals — rule-based lexical, neural semantic (sentence-transformers), and resolution-time — into an inferred severity, then derive the binary mismatch label by comparing to the assigned priority. The ablation shows each signal's individual contribution."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from src import pseudo_label\n",
    "labeled = pseudo_label.main()\n",
    "print(json.load(open(C.METRICS_DIR / 'pseudolabel_stats.json')))\n",
    "pd.read_csv(C.METRICS_DIR / 'ablation.csv')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Stage 3 — Fine-tune the binary classifier\n",
    "Class-weighted fine-tuning of an MPS-stable encoder on `model_text` (assigned priority + channel/category/tier tags + the real issue sentence). Held-out verification against the thresholds (Accuracy ≥ 83%, Macro-F1 ≥ 0.82, per-class recall ≥ 0.78).\n",
    "\n",
    "> This cell trains for several minutes. You can also run it from the shell: `python3 train_pipeline.py`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from src import train\n",
    "metrics = train.main()\n",
    "metrics"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Inference + Evidence Dossier\n",
    "Run the trained auditor on the adversarial set and inspect a dossier. Every `feature_evidence` value is traceable to a real ticket field (validated)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import predict\n",
    "adv = pd.read_csv(C.DATA_DIR / 'adversarial_tickets.csv')\n",
    "out, dossiers = predict.run_inference(adv)\n",
    "out[[C.COL_ID, C.COL_PRIORITY, 'inferred_level', 'pred_mismatch', 'mismatch_type']]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(json.dumps(dossiers[0], indent=2))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Adversarial robustness score\n",
    "10 tickets engineered to fool keyword systems; ≥ 7/10 earns the bonus."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from src import adversarial\n",
    "report = adversarial.main()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {"display_name": "Python 3", "language": "python", "name": "python3"},
  "language_info": {"name": "python", "version": "3.11"}
 },
 "nbformat": 4,
 "nbformat_minor": 5
}