AI-Native

The architectural patterns that distinguish a working AI prototype from a production AI system — recognising that adding a model to an architecture changes the architecture's stability properties, cost profile, and failure modes in ways that don't appear in classical software engineering.

The engineering discipline of measuring and managing the harms an AI system can cause — recognising that "responsible AI" is not a value statement but a set of measurable properties that need to be designed in, tested for, and continuously monitored.

The signals that tell you whether an AI system is working — recognising that classical observability (latency, error rate, throughput) is necessary but radically insufficient, because AI systems can be perfectly fast, perfectly available, and producing wrong answers all day long.

The architectural pattern of grounding a language model's output in a corpus the model didn't memorise — recognising that "we'll just plug in a vector database" is the demo, and the architecture is everything between the user query and the retrieved-and-grounded answer.

The threats that exist because the system uses a foundation model — categorically distinct from classical application security, because the *input itself* and the *model itself* are now attack surfaces with no equivalent in pre-AI architectures.