https://devencelab.com/articles Field notes from Devence Lab on provable autonomy, safety engineering, and deploying AI systems where failure is not an option. en-us https://devencelab.com/articles/evals-are-not-a-safety-case https://devencelab.com/articles/evals-are-not-a-safety-case Mon, 24 Aug 2026 00:00:00 GMT Rami Jaouadi 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. https://devencelab.com/articles/temperature-zero-is-not-determinism https://devencelab.com/articles/temperature-zero-is-not-determinism Tue, 09 Jun 2026 00:00:00 GMT Sara Ben Youssef 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. https://devencelab.com/articles/provisional-is-not-law https://devencelab.com/articles/provisional-is-not-law Tue, 19 May 2026 00:00:00 GMT Rami Jaouadi 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. https://devencelab.com/articles/you-cannot-debug-an-agent-from-a-latency-graph https://devencelab.com/articles/you-cannot-debug-an-agent-from-a-latency-graph Tue, 14 Apr 2026 00:00:00 GMT Sara Ben Youssef 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. https://devencelab.com/articles/agent-networks-fail-in-the-handoff https://devencelab.com/articles/agent-networks-fail-in-the-handoff Tue, 17 Mar 2026 00:00:00 GMT Rami Jaouadi 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.