Blog
Field notes from Devence Lab engagements — safety engineering, formal verification, and what it actually takes to deploy autonomous systems where failure is not an option.
Assurance
Half of enterprises have shipped an agent that passed its evals and still failed in front of a customer. The problem is not test coverage — it is that a score answers a different question than the one a deployment decision asks.
Reliability
Most teams believe an LLM’s unpredictability comes from sampling, and that setting temperature to zero removes it. It does not. The deeper source is arithmetic, and it depends on how many other people were using the server.
Regulation
Brussels has spent six months signalling that the AI Act’s high-risk deadline will move. Nothing has been adopted. Treating the delay as settled is a scheduling decision that cannot be defended if it turns out to be wrong.
Observability
Agents fail while returning HTTP 200. Every span is green, every latency is normal, and the answer is wrong. Conventional monitoring is structurally blind to the only failure mode that matters.
Architecture
When multi-agent systems fail in production, the overwhelming majority of failures trace to coordination and specification— not to any limitation of the underlying models. Upgrading the model does not fix a missing contract.
Collaborate