ARCHITECTURE
Architecture
Blueprints, not opinions. Each one pins down a layer of an AI system, names the decisions it forces on you, and says where the pattern stops working.
Most published AI reference architectures are a capability diagram with a governance box drawn beside it. The capability layer is the part the industry has converged on and the part vendors will sell you. The other two are where systems actually fail in production, and they get a box.
This section splits an AI system into three layers and treats each as load-bearing: what it can do, what it is allowed to do, and how you know what it did. Blueprints are written to be read in order within a layer, and revised rather than republished.
These are engineering specifications. For the same territory at board and operating-model altitude, funding, mandate, maturity and the org structure around all of this, see the Enterprise AI Transformation playbook.
Capability
What the system can do. Models, retrieval, memory, tools, orchestration.
No blueprint yet. The capability patterns live in the pattern catalogue.
Control
What the system is allowed to do. Chokepoints, detectors, policy, authority, human gates.
- ARCH-001 The Control Plane A single-pass AI request has six places where a control can sit, and each one can see a different amount of the truth. Placement sets the accuracy ceiling before you pick a single model.
- ARCH-002 Detector Architecture Four detection techniques, a sourcing decision people mistake for a fifth, one interface contract, and the calibration problem that breaks every ensemble built by averaging scores that do not mean the same thing.
- ARCH-003 Measuring Detectors A detector with 95 percent recall and a 1 percent false alarm rate is wrong nine times out of ten when the attack is rare. Prevalence, not model quality, decides whether a guardrail is usable, and most test sets are too small to settle the question either way.
- ARCH-004 The Decision Layer Six possible actions, and the two easiest to build are the two most systems have. Which one fires belongs to the reversal cost, the exposure, and the confidence band, never to the detector.
- ARCH-008 Production Execution Detection is work you added to a request that a user is waiting on. The latency budget, not the accuracy target, is what decides how many detectors you can actually run and in what order.
Evidence
How you know what it did. Evaluation, telemetry, tracing, audit.
- ARCH-005 Ownership and Decision Rights A guardrail has five separable decisions inside it, and they do not all belong to the same team. Most programmes stall because the threshold and the alert queue ended up owned by whoever built the classifier.
- ARCH-006 The Guardrail Lifecycle A detector goes from proposal to production through eight gates, and the two that matter most are the two teams skip: shadow mode before enforcement, and a revalidation date set on the day it ships.
- ARCH-007 Telemetry, Audit and the Review Loop The control plane produces three different things people call logs, with three different retention rules and three different audiences. Merging them is how a system ends up unable to answer a regulator and full of sensitive data at the same time.