Topic
5 articles
A language model will not give the same answer twice, and no prompt changes that. What can be made deterministic are the properties that matter: what never happens, what always happens, in which order and shape, and what cannot ship unverified. A close reading of Claude Code shows where instructions end, where guarantees begin, and where people belong.
Every AI option buys three things at once: a capability, a dependency and an option on the future. A decision method with three gates, a three-year cost model that includes leaving, explicit lock-in and sovereignty scores — and why the honest answer is usually a hybrid.
A portfolio method for AI use cases: four value mechanisms and how much of each reaches the profit and loss, a business case built on measured baselines and an honest adoption assumption, a value–feasibility map with the AI Act drawn on it — and the business-development half most strategies forget.
A translation of the Digital Operational Resilience Act from legal text into engineering requirements — ICT risk, incident reporting, resilience testing, third-party risk — with what it means for teams that build and run AI and platform systems.
Why most production failures of LLM systems are context failures, how tokenization shapes cost and quality, and a practical method for designing context windows as deliberately as you design a database schema.