Architecture First Principles

The invariants every design decision must respect — separation of latency classes, physical safety floor, bounded intelligence, and traceable data.

The invariants

Beneath the layers and colors sit a few invariants that no design change may break. They are the tests a proposed architecture must pass, derived from the physics of the machines and the discipline of the honest gates.

The four invariants

• Latency classes never mix — a single path carries a single latency guarantee.
• There is always a physical safety floor — the failsafe works with zero AI dependency.
• Intelligence is always bounded — no AI output actuates without passing safety bounds.
• Data is always traceable — every signal, model, and decision carries lineage.

How the invariants are derived

Each invariant answers a specific hazard. Mixing latency classes would let a slow function delay a fast one — hazardous near the magnets, so it is forbidden. Depending on AI for the last line of defense would couple safety to the least predictable component — so a physical floor is mandatory. Unbounded intelligence could actuate on a wrong prediction — so bounds are mandatory. Untraceable data would make the honest gates unenforceable — so lineage is mandatory.

Using the principles

In practice the invariants make design reviews decidable. A proposed shortcut from a copilot to an actuator fails the bounded-intelligence test. A telemetry line drawn onto the command path fails the latency-class test. Because the invariants are few and sharp, disagreements resolve to whether a design violates one of them, not to taste.

The invariants are the compressed form of the design philosophy and are enforced through interface contracts.