Changes to live AI are controlled comprehensively so that they cannot quietly cause unintended behaviour, worse performance or a drift away from the intended use. Three elements. Scope: a framework for judging how significant a change is, from its nature (training data, model or system architecture, key assumptions, scope of use, knock-on effects on downstream AI) and from the materiality of the use case, with significant changes reviewed and re-approved before they go in, and compensating measures such as heightened monitoring where significant changes to high-materiality uses can occur without prior review. Change control: mechanisms, under human oversight, that block unauthorised alteration and ensure changes to code, configuration or training data do not cause outages, disruption or degraded performance, backed by version control that tracks not only model code but the associated data, parameters, hyperparameters and other artefacts, giving traceability, auditability and the ability to roll back. Dynamic updates: stronger controls for AI built to update itself, including a strict case for allowing automatic updates at all, a clear boundary on what may change automatically (retraining or hyperparameters, say, but never the core architecture), extra data quality checks and tighter performance monitoring.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.