ABOUT

Why this doctrine exists.

AI assistants now write code that ends up running factories, refineries, hospitals, and power grids. Consequential Use Cases! The Vibe Doctrine is a measured response to a question nobody had answered with data: when an LLM designs the architecture of a redundant/non-failure, cyber-secure, safety-critical system, what does it systematically get wrong?

What changes when you adopt the doctrine

A practitioner-level view of before vs. after.

BEFORE
  • The architecture uses what appears to be best practices; however, it failed to actually implement them.
  • Cloud services creep into offline-only systems, even after explicitly prohibiting them.
  • HAZOP, LOPA, and SIL analyses requested; however never done or reported as incomplete.
  • Self-reported confidence scores are treated as evidence.
  • The review meeting argues about taste, not specifics.
AFTER
  • Each architectural decision maps to a named control or standard.
  • Cloud dependencies are either explicit, justified, or refused.
  • Formal safety analyses are cited by document and revision number.
  • Confidence scores are ignored unless independently validated.
  • The review meeting argues about evidence, not opinion.

HAZOP Hazard and Operability study · LOPA Layer of Protection Analysis · SIL Safety Integrity Level

How the doctrine was built

The doctrine is reverse-engineered from a measured failure pattern, not from opinion.

6
Frontier LLMs tested side-by-side
  • GPT-5.5
  • Claude Opus 4.7
  • Gemini 2.5 Pro
  • Grok 4.20
  • DeepSeek V4 Pro
  • Qwen 3 30B-A3B
216
Architecture recipes generated
  • GPT-5.5 Thinking · drafter
  • Gemini 3.1 Pro · optimizer
  • Claude Opus 4.7 · auditor
PRD clean-room
32,400
Architectural decisions analyzed
  • Mistral Large 2
  • Llama-Nemotron Super 49B
Meta-auditors
213
Independent red-team reviews
  • GPT-5.5
  • Claude Opus 4.7
  • Gemini 2.5 Pro
  • Grok 4.20
  • DeepSeek V4 Pro
  • Qwen 3 30B-A3B
Cross-family rotation · no self-review

12 realistic small-factory PRDs • 3 independent runs per model • a 7th model used for the independent red-team

About the author

Harold Clampitt

Harold Clampitt is a licensed Professional Engineer with an engineering degree and hundreds of completed large-scale enterprise projects — a career spent delivering exactly the kind of high-consequence systems this research is about. After watching AI-assisted architecture proposals look more credible than they actually were, he built this study to learn, with measured data rather than anecdote, precisely where they fall short.

He is also the author of AI Decoded (Second Edition) and has written more than 40 articles on AI at haroldclampitt.com — including the ten-part series AI Secrets: What Others Have Missed, written to take readers from informed to genuine subject-matter expert. This site is a deliberate artifact: everything published here — the 10 guardrails, the 13 acceptance gates, the copy-paste prompts, the full paper — is built to be used in a real design review tomorrow morning, not framed and admired.

Get in touch: harold@haroldclampitt.com · github.com/Hallta1937/vibedoctrine

Start applying the doctrine

Two paths in. Pick the one closer to your next design review.

The methodology and reproducibility instructions live in the GitHub repository.