The Trust Problem: Designing Quality Frameworks for AI-Generated Code

on YouTube ↗26m

" summary

AI-generated code often compiles and passes its own tests but fails intent checks — a third was confidently wrong on first pass; a four-line loyalty function applies discounts to order total instead of pre-tax subtotal and silently breaks when shipping is added. A five-pillar framework is proposed: explicit assertions from policy, least privilege, assume drift, micro-segmented review, continuous attestation. Demo shows an evidence ledger and automated gate that drops Trust Score 96 to 89, restricts blast radius, and forces re-attestation.

AI-written summary · about

" chapters (29)

1. 0:00The Trust Problem: Quality Frameworks for AI-Generated Code
2. 1:05Speaker Intro - Neelmani Verma, Industry Principal, Infosys
3. 2:07A Confession: A Third of It Was Confidently Wrong on the First Pass
4. 3:10The Four-Line Function That Passes Every Test and Is Still Wrong
5. 4:11The Silent Bug That Appears Three Weeks Later
6. 4:41The Assumption QE Has Always Run On: The Author Understood the Intent
7. 5:12The Five-Step Chain That AI Code Doesn't Break - It Undermines
8. 5:42Break 1 - Fluent but Not Understanding: The Currency Conversion Case
9. 6:14Break 2 - No Fixed Author: A Model Change Is a Staffing Change
10. 7:18Break 3 - Trust Decays After the Merge
11. 7:51Borrowing Zero Trust: Never Trust, Always Verify
12. 8:53Pillar 1 - Verify Explicitly and the Oracle Problem
13. 9:57A Code-Generated Test Is a Mirror, Not an Oracle
14. 10:28Pillar 2 - Least Privilege and Scoping the Blast Radius by Risk Tier
15. 11:29Pillar 3 - Assume Drift: Trust With an Expiry Date
16. 12:31Pillar 4 - Micro-Segmented Review Depth
17. 13:36Pillar 5 - Continuous Attestation Instead of a Static Test Report
18. 14:37Live Demo: Day Zero, Trust Score 96, Assertion From Policy
19. 16:08Day 12: A Dependency Auto-Merge Drops the Score With No Code Change
20. 16:38Day 28: A New Field Enters the Order Object Upstream
21. 17:09Day 45: Production Data Pushes It Into Drifting
22. 17:39The Automated Gate Response: Blast Radius Auto-Restricted
23. 18:10Re-Attestation Returns 89, Not 96 - Trust Accumulates History
24. 19:42What This Changes for QE Teams: Metrics, Roles, Tooling
25. 21:13Three Takeaways: Syntax Is Not Semantics, Coverage Isn't Enough
26. 22:13Trust Is a Credential You Keep Renewing
27. 22:43Q&A: Should AI Code Face a Higher Bar Than Human Code?
28. 23:13Q&A: How Often Should a Framework Re-Evaluate for Silent Drift
29. 24:14Q&A: Do You Track Which Code Was AI-Generated

" related talks

• 31mMeasurement of Trust in AIShashank Shekhar · PyData Global 2022
• 16mTrust by Design: Building Reliable AI for Sensitive EnvironmentsKim Nguyen, Prof Dr Gesche Joost · AI & Data Summit and Quantum Summit 2025
• 51mNever Trust a Monkey: The Chasm, the Craft, and the Chain of AI-Assisted CodeBaruch Sadogursky · Jfokus 2026
• 48mNever Trust a Monkey: The Chasm, the Craft & the Chain of AI-Assisted CodeBaruch Sadogursky · JCON EUROPE 2026
• 34mTrust in the Age of AI Agents | Building Trustworthy AIStephen Gray · AI Agents Conference 2026
• 32mCan We Trust AI-Generated Code?Baruch Sadogursky, Leonid Igolnik · AI By the Bay 2025
• 39mBack to Basics: Cutting Through the AI-Generated Code NoiseArchana Vaidheeswaran · PyCon Italia 2025
• 34mBuilding Guardrails for AI Coding SystemsCole Medin, JOHN CLIFTON · Sonar Summit 2026