Traditional software documentation methods like arc42 are insufficient for AI systems, requiring a new architectural extension to bridge the gap.
The LARS loan scoring system illustrates how a lack of architectural documentation leads to an inability to answer regulator questions about model decisions.
AI systems consistently suffer from seven architectural gaps, including hidden data dependencies, model drift, poor explainability, and fragmented documentation.
AI systems are fundamentally different due to their probabilistic outputs, deep data dependencies, and continuous lifecycle of retraining and evolution.
The arc-ai-42 extension acts as a missing architectural connectivity layer, separating specifications from evidence to make implicit processes explicit and auditable.
The framework is structured around five key views: data context, model explainability, lifecycle, deployment architecture, and governance.
Successful adoption requires using a maturity ladder, separating static and living documentation, and treating governance decisions as architectural decision records (ADRs).
arc-ai-42 serves as the central layer connecting the AI system and tooling layer below with the organizational and governance layer above.
Unexplainable, un-trainable, or ungoverned AI systems are not just risky but are fundamentally unarchitected and undocumented.
Jobs that call for the skills explored in this talk.