diff --git a/MyIA.AI.Notebooks/Probas/DecisionTheory/Causal-Bridges/DoWhy-4-Sensibilite-Confounder-Cache.ipynb b/MyIA.AI.Notebooks/Probas/DecisionTheory/Causal-Bridges/DoWhy-4-Sensibilite-Confounder-Cache.ipynb
new file mode 100644
index 0000000000..9995bf8d34
--- /dev/null
+++ b/MyIA.AI.Notebooks/Probas/DecisionTheory/Causal-Bridges/DoWhy-4-Sensibilite-Confounder-Cache.ipynb
@@ -0,0 +1,1504 @@
+{
+ "cells": [
+ {
+ "cell_type": "markdown",
+ "id": "624e14af",
+ "metadata": {
+ "papermill": {
+ "duration": 0.002558,
+ "end_time": "2026-09-13T02:06:54.296487+00:00",
+ "exception": false,
+ "start_time": "2026-09-13T02:06:54.293929+00:00",
+ "status": "completed"
+ },
+ "tags": []
+ },
+ "source": [
+ "# DoWhy-4 — Le confondeur non observé : sensibilité, pas certitude\n",
+ "\n",
+ "DoWhy-1 exigeait un estimand **nommé** ; DoWhy-3 découvrait le graphe avec sa\n",
+ "borne structurelle (le CPDAG ambigu). Il reste l'hypothèse qu'aucun graphe\n",
+ "observé ne peut trancher : **l'absence de confondeur NON OBSERVÉ**. On ne\n",
+ "peut ni la tester, ni la découvrir — on peut la **chiffrer**.\n",
+ "\n",
+ "L'énoncé cible de ce notebook (issue #14049) :\n",
+ "\n",
+ "> « Quelle force devrait avoir un confondeur caché pour annuler cet effet ? »\n",
+ "\n",
+ "— **un chiffre, pas une réserve rhétorique**. Trois formalisations SOTA de la\n",
+ "même question, toutes réellement exécutées par `dowhy` (0.14) :\n",
+ "\n",
+ "| Formalisation | Moteur | Monde | Le chiffre |\n",
+ "|---|---|---|---|\n",
+ "| Robustness value (R² partiel) | `refute_estimate` → `linear-partial-R2` | continu | le R² partiel minimal qui annule |\n",
+ "| E-value (Ding & VanderWeele) | `refute_estimate` → `e-value` | binaire (RR) | le RR minimal qui annule |\n",
+ "| Bornes de Rosenbaum (Γ) | calcul exact binomial sur paires | binaire apparié | le Γ* qui rend non significatif |\n",
+ "\n",
+ "Et le pont concret : `direct-simulation` — un confondeur simulé de force\n",
+ "croissante, et la **courbe de bascule** de l'estimé ajusté."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 1,
+ "id": "975783c6",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-09-13T02:06:54.311117Z",
+ "iopub.status.busy": "2026-09-13T02:06:54.310844Z",
+ "iopub.status.idle": "2026-09-13T02:06:56.429330Z",
+ "shell.execute_reply": "2026-09-13T02:06:56.428796Z"
+ },
+ "papermill": {
+ "duration": 2.122835,
+ "end_time": "2026-09-13T02:06:56.429704+00:00",
+ "exception": false,
+ "start_time": "2026-09-13T02:06:54.306869+00:00",
+ "status": "completed"
+ },
+ "tags": []
+ },
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "organe : Causal-Bridges\n"
+ ]
+ }
+ ],
+ "source": [
+ "from pathlib import Path\n",
+ "import sys\n",
+ "\n",
+ "import numpy as np\n",
+ "import pandas as pd\n",
+ "\n",
+ "_fichier_organe = next(Path.cwd().rglob(\"dowhy_sensitivity_organs.py\"))\n",
+ "sys.path.insert(0, str(_fichier_organe.parent))\n",
+ "import dowhy_sensitivity_organs as dso\n",
+ "\n",
+ "pd.set_option(\"display.precision\", 3)\n",
+ "print(\"organe :\", _fichier_organe.parent.name)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "a671d79a",
+ "metadata": {
+ "papermill": {
+ "duration": 0.003449,
+ "end_time": "2026-09-13T02:06:56.435974+00:00",
+ "exception": false,
+ "start_time": "2026-09-13T02:06:56.432525+00:00",
+ "status": "completed"
+ },
+ "tags": []
+ },
+ "source": [
+ "## 1. Le monde A — l'analyste et son angle mort\n",
+ "\n",
+ "Un traitement continu $X$, une issue continue $Y$, un confondeur **observé**\n",
+ "$C$ — et un confondeur **caché** $U$ que l'analyste ne mesure pas :\n",
+ "\n",
+ "$$X = 0.7\\,U + 0.5\\,C + \\varepsilon_X \\qquad Y = 0.5\\,X + 0.9\\,U + 0.4\\,C + \\varepsilon_Y$$\n",
+ "\n",
+ "L'effet vrai vaut $\\tau = 0.5$. L'analyste ajuste sur $\\{C\\}$ — le mieux\n",
+ "qu'il puisse faire avec ses données. Le générateur expose deux vues : la vue\n",
+ "**analyste** (`X, Y, C`) et la vue **oracle** (avec `U`), qui servira de\n",
+ "vérité terrain."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 2,
+ "id": "45ca444b",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-09-13T02:06:56.443659Z",
+ "iopub.status.busy": "2026-09-13T02:06:56.443012Z",
+ "iopub.status.idle": "2026-09-13T02:06:56.473674Z",
+ "shell.execute_reply": "2026-09-13T02:06:56.472837Z"
+ },
+ "papermill": {
+ "duration": 0.034203,
+ "end_time": "2026-09-13T02:06:56.473286+00:00",
+ "exception": false,
+ "start_time": "2026-09-13T02:06:56.439083+00:00",
+ "status": "completed"
+ },
+ "tags": []
+ },
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "vue analyste : ['X', 'Y', 'C']\n"
+ ]
+ },
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " X | \n",
+ " Y | \n",
+ " C | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 0 | \n",
+ " -0.594 | \n",
+ " -0.900 | \n",
+ " -0.675 | \n",
+ "
\n",
+ " \n",
+ " | 1 | \n",
+ " -0.191 | \n",
+ " -0.719 | \n",
+ " -0.145 | \n",
+ "
\n",
+ " \n",
+ " | 2 | \n",
+ " 0.070 | \n",
+ " -0.359 | \n",
+ " -0.792 | \n",
+ "
\n",
+ " \n",
+ " | 3 | \n",
+ " 1.243 | \n",
+ " 1.485 | \n",
+ " -0.308 | \n",
+ "
\n",
+ " \n",
+ " | 4 | \n",
+ " -2.068 | \n",
+ " -2.152 | \n",
+ " -1.894 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " quantite | \n",
+ " valeur | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 0 | \n",
+ " estime naif (ajuste C) | \n",
+ " 1.159 | \n",
+ "
\n",
+ " \n",
+ " | 1 | \n",
+ " oracle (ajuste C+U) | \n",
+ " 0.532 | \n",
+ "
\n",
+ " \n",
+ " | 2 | \n",
+ " effet vrai | \n",
+ " 0.500 | \n",
+ "
\n",
+ " \n",
+ " | 3 | \n",
+ " robustness value | \n",
+ " 0.677 | \n",
+ "
\n",
+ " \n",
+ " | 4 | \n",
+ " R2 reel U avec X | \n",
+ " 0.493 | \n",
+ "
\n",
+ " \n",
+ " | 5 | \n",
+ " R2 reel U avec Y | \n",
+ " 0.427 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "