Production AI architecture: identity, control and recovery
Production AI architecture requirements: exact identity, observability, controls, refusal, rollback and requalification.
Production starts when every output can be bound to an exact version of the realization, its data and its policies.
A strong prototype becomes fragile when it lacks input limits, timeouts, traces, refusal behavior and rollback procedures.
Production authority must attach to an identified release, never to a moving model alias.
What the architecture must make explicit
The engineering object is the complete realization. A model is only one possible component among data, memory, routing, tools, rules, verifiers and runtime.
Move from a keyword to an inspectable engineering decision
Observe
Measure the baseline and bind every behavior to the exact realization that produced it.
Intervene
Test one change at a time to attribute gains and regressions.
Compose
Verify interfaces, shared state and interactions before promotion.
Qualify
Apply hard gates to the frozen realization within a declared scope.
A claim is only as strong as its binding
The gate must evaluate the exact output produced by the exact realization. Integrity, scientific validity and authority are related, but never interchangeable.
Questions that change the design
What does production ai architecture: identity, control and recovery require?
Production starts when every output can be bound to an exact version of the realization, its data and its policies. The engineering object is the complete realization. A model is only one possible component among data, memory, routing, tools, rules, verifiers and runtime.
Which evidence should be retained?
At minimum: exact realization identity, inputs and outputs, protocol version, measurements, failure cases and the decision rule. Production authority must attach to an identified release, never to a moving model alias.
Does AMF prescribe a particular model?
No. AMF treats the model, runtime, data, tools and deterministic components as candidate materials. Their retention depends on measured contribution and declared constraints.