For years, architectural conformance in regulated firms was enforced by an unstated control: a human in code review who knew the system and wrote at human speed. AI coding agents remove both halves - they don't know your declared architecture, and they generate change faster than any review board can inspect. The control didn't get weaker; it got bypassed.
This talk argues that telling an agent about your architecture is not the same as enforcing it. Advisory context still drifts; only a deterministic, non-bypassable gate that reconciles code against a declared model holds. Using FINOS CALM as the architectural source of truth, I'll walk through how to reconcile what an agent actually built, across files, through indirection, against what the architecture permits, and why the verdict must be sound, three-valued, and honest about what it can't determine.
Attendees leave with a pattern for making architectural governance enforceable at PR time in an agent-driven SDLC, using open standards, with an audit trail a regulator accepts.