Add tutorial for dspy-trusted-monitor using GEPA (#8938)

* add gepa_trusted_monitor notebook, copied and adapted from gepa_aime_math

* clean up some naming and ordering

* add some tutorial instructions

* add index.py as an export to help with PR review

* add more instructions, add analysis cells

* run training

* run evaluation and analysis

* export to index.py

* remove inspect outputs from notebook

* add gepa_trusted_monitor link to index pages

Author: ZachParent

docs/docs/tutorials/gepa_ai_program/index.md

@@ -11,4 +11,7 @@ This tutorial explores how GEPA can optimize a single `dspy.ChainOfThought` base
 This tutorial explores how GEPA leverages predictor-level feedback to improve GPT-4.1 Nano's performance on a three-part task for structured information extraction and classification in an enterprise setting.
 
 ### [GEPA for Privacy-Conscious Delegation](../gepa_papillon/index.ipynb)
-This tutorial explores how GEPA can improve rapidly in as few as 1 iteration, while leveraging a simple feedback provided by a LLM-as-a-judge metric. The tutorial also explores how GEPA benefits from the textual feedback showing a breakdown of aggregate metrics into sub-components, allowing the reflection LM to identify what aspects of the task need improvement.
+This tutorial explores how GEPA can improve rapidly in as few as 1 iteration, while leveraging a simple feedback provided by a LLM-as-a-judge metric. The tutorial also explores how GEPA benefits from the textual feedback showing a breakdown of aggregate metrics into sub-components, allowing the reflection LM to identify what aspects of the task need improvement.
+
+### [GEPA for Code Backdoor Classification (AI control)](../gepa_trusted_monitor/index.ipynb)
+This tutorial explores how GEPA can optimize a GPT-4.1 Nano classifier to identify backdoors in code written by a larger LM, using `dspy.GEPA` and a comparative metric! The comparative metric allows the prompt optimizer to create a prompt that identifies the signals in the code that are indicative of a backdoor, teasing apart positive samples from negative samples.