--- description: architectural decision boundaries and Architect-Conductor escalation alwaysApply: true --- # Governance: Architect-Conductor Boundaries Agents may implement within an approved architecture. They must not silently invent or replace the architecture that governs the work. ## Autonomous Implementation Boundary - `GOV-14-ARCH-001` Routine implementation and internal refactoring may proceed autonomously when they remain inside the approved requirements, public contracts, trust boundaries, persistence model, and permitted scope. - `GOV-14-ARCH-002` The smallest coherent implementation that fits established repository patterns should be preferred over a new abstraction or framework. - `GOV-14-ARCH-003` Lack of a human-authored diagram is not itself a blocker when existing code, specs, decisions, and tests make the architectural boundary clear. ## Decisions Requiring Escalation - `GOV-14-ARCH-004` Stop and raise a blueprint decision before creating or replacing a system-wide boundary. - `GOV-14-ARCH-005` Blueprint review is required before introducing shared global state, changing persistence semantics, altering a public contract, moving a security or trust boundary, creating an irreversible migration, or materially replacing an approved architecture. - `GOV-14-ARCH-006` The escalation must state the decision required, current constraints, viable options, trade-offs, affected contracts, migration or rollback implications, and the evidence needed to proceed. - `GOV-14-ARCH-007` The Architect-Conductor may resolve the boundary through an approved architecture decision record, updated spec, interface contract, migration plan, or equivalent durable artifact. ## Evidence and Completion - `GOV-14-ARCH-008` Local mocks, implementation-authored tests, compilation, and presentation quality cannot by themselves prove a system-level architectural claim. - `GOV-14-ARCH-009` Completion requires the repository's declared verification gates and direct evidence appropriate to the affected boundary. - `GOV-14-ARCH-010` When the runtime cannot execute a required gate, report the limitation and classify the claim as blocked, partial, or awaiting review rather than complete. ## Anti-Patterns Avoid these failure modes: - treating every internal abstraction as a human approval event - inventing architecture implicitly through accumulating patches - using a longer prompt as a substitute for a durable decision artifact - reporting system-level completion from isolated mocks - continuing mutation while the governing boundary is unresolved