Email capture preview

// subscribe

The AI Validation Boundary: Extended Edition

The articles stay open. The full worked examples come here: complete validation walkthroughs for probabilistic AI systems in regulated drug development, too long for a post. Free. No fixed schedule; you’ll hear from me when one is ready.

Free on Substack. Unsubscribe anytime.

// start here

Later pieces in The AI Validation Boundary build on earlier ones. Pick the question closest to yours and read that topic in order. Terms are defined in the glossary.

New here? Start with these three

The architecture, the unit of validation, and why context comes first. Read them in order.

  1. Why AI Governance Fails Without a Control Layer (House of AI Trust) Mar 2026
  2. Context is King Jun 2026
  3. The Validation Boundary Inside the House House Boundary series Sept 2026

Key terms: House of AI Trust · Validation boundary · Context of use

What exactly are we validating?

Why the model alone is the wrong unit, and what the boundary has to include.

  1. Context is King Jun 2026
  2. AI as Autonomous Vehicles: Can We Detect When They Fail? Jul 2026
  3. Beyond the Weights Jul 2026
  4. The Validation Boundary Inside the House House Boundary series Sept 2026

Key terms: Probabilistic system · Harness · Harness surfaces

Qualifying an AI vendor

What supplier evidence covers, what it can't, and what the sponsor still has to generate.

  1. When Your AI Vendor Passes SOC 2 and Still Fails GxP May 2026
  2. When the Check Runs Somewhere Else Check trio Sept 2026
  3. The Inheritance Boundary Inside the House House Boundary series Oct 2026
  4. Harness Evidence Worksheet for AI Supplier Assessment Tool Sept 2026
  5. VALID Trust Framework Framework

Key terms: VALID Trust · Inheritance boundary · Trust boundary

Investigating AI failures

Classifying probabilistic failures so deviations and CAPA have something to work with.

  1. Bounding the Uncertainty: An Error Taxonomy for Non-Deterministic AI May 2026
  2. You Can't CAPA What You Can't Classify (Pharmaceutical Online) Sept 2026
  3. Probabilistic Failure Mode Taxonomy Framework Framework

Key terms: PFMT · Origination point · Error nature

Human oversight

Why the reviewer is a variable to evidence, not a constant to assume.

  1. Human-in-the-Loop, Liability Still in Play Jan 2026
  2. Untangling the Web: HOTL for Agentic PV Jan 2026
  3. Balancing the Equation: Why HITL Is a Variable, Not a Constant Jul 2026

Key terms: HITL · HOTL · Correlated failure · HPQ

Evidence and the audit trail

The deterministic record underneath a probabilistic system, and what it has to capture.

  1. The Last Deterministic Thing: The Unified Capture Layer Check trio Aug 2026
  2. When a Check Doesn't Run Check trio Aug 2026
  3. When the Check Runs Somewhere Else Check trio Sept 2026

Key terms: UCL · Did not run · Replay

Data, drift and change

How the system moves after validation, including changes nobody made on purpose.

  1. Data Drift: A Risk-Based and GAMP-Aligned Approach Nov 2025
  2. Beyond the Weights Jul 2026
  3. The Version Bump Nobody Change-Controlled Sept 2026
  4. The Data Boundary Inside the House House Boundary series Oct 2026

Key terms: Drift · Harness drift · Data boundary

Security

Why security of the environment is part of whether a probabilistic system is valid.

  1. Your Frozen Architecture May Have a Back Door Mar 2026

Key terms: Security thread

Agentic AI

Oversight and validation when the system acts, not just answers.

  1. FDA's Agentic AI Announcement Signals a New Era for Scientific Computing Dec 2025
  2. Untangling the Web: HOTL for Agentic PV Jan 2026
  3. The Safety Paradox: Why Frozen Models Aren't Always Safer Than Agentic Ones Apr 2026

Key terms: Agentic AI governance · Agent orchestration

The regulatory picture

Where FDA, EMA and GAMP guidance stand, and where the gaps are.

  1. FDA & EMA Just Released AI Guiding Principles for Drug Development Jan 2026
  2. The Guidance Says What. The Next 12 Articles Show How Feb 2026
  3. Why AI Validation Architecture Belongs in the FDA's Clinical Trial Pilot Jun 2026

Key terms: Static vs dynamic · Annex 22 “deterministic”

The business case

What it costs to treat AI as a model problem instead of a system problem.

  1. The Capability Paradox Dec 2025
  2. The $100M Question Your AI Can't Answer (Yet) Apr 2026
  3. Coherence ≠ Compliance: The Capability Paradox, 10 Months Later Oct 2026

Key terms: Capability paradox · Layer 5

Mini-series

Some pieces were written to be read together.

House Boundary series

Three pieces that carry the validation boundary through the House of AI Trust.

  1. The Validation Boundary Inside the House
  2. The Inheritance Boundary Inside the House
  3. The Data Boundary Inside the House

Check trio

What has to be recorded for a check to count, including checks that silently didn't run or ran on the vendor's side.

  1. The Last Deterministic Thing: The Unified Capture Layer
  2. When a Check Doesn't Run
  3. When the Check Runs Somewhere Else

New pieces publish every two weeks. Get them early, with full worked examples, in The AI Validation Boundary: Extended Edition.