- Updated: July 15, 2026
- 2 min read
Who Analyses the Analyser? Self‑Validating LLM Hazard Analysis with Constitutional Meta‑STPA – A Deep Dive
Who Analyses the Analyser? Self‑Validating LLM Hazard Analysis with Constitutional Meta‑STPA
Large language models (LLMs) are increasingly trusted to draft safety‑analysis artefacts such as losses, hazards, Unsafe Control Actions (UCAs) and safety constraints within rigorous processes like Systems‑Theoretic Process Analysis (STPA). However, the tool that performs the analysis itself—an LLM‑assisted safety‑analysis system—has been largely ignored. This paper asks the critical question: who analyses the analyser?
We introduce Constitutional Meta‑STPA, an LLM‑assisted STPA framework that turns the analysis back onto the tool. By running a meta‑STPA on AI‑assisted safety tools, the framework derives a governance constitution rather than merely asserting one. The resulting constitution comprises 21 Tool Principles and 8 Meta‑Safety Principles, each linked to a concrete code‑enforcement point.
The methodology formalises a constitution‑marginal coverage operator over a principle set P (|P| = 29) and proves a soundness lemma that isolates coverage from model and scanner effects. Empirical findings show that a frontier ensemble (claude‑opus‑4.8 + claude‑sonnet‑4) recovers 18/21 canonical and all 8 meta‑safety principles, demonstrating that the meta‑layer is limited by model capability rather than by the constitution itself.
Key contributions:
- Self‑derivation of governance principles from the tool’s own design.
- A formal coverage operator and soundness proof for constitutional analysis.
- Empirical validation across multiple LLM ensembles.
- Practical recommendations for integrating Constitutional Meta‑STPA into AI‑driven safety workflows.
For a full technical exposition, see the arXiv pre‑print. Further resources, implementation details, and downloadable assets are available on our internal portal ubos.tech.
By publishing this analysis, we aim to close the audit‑trail gap and provide a reproducible, self‑validating safety‑analysis pipeline for the next generation of AI‑augmented engineering tools.
Andrii Bidochko
CTO UBOS
Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.