Evolution, Learning & Adaptation
AI Products evolve, but not every AI Product learns and not every AI Product retrains.
AIPDS therefore requires an explicit learning mode and a broader evolution strategy rather than assuming retraining as a universal lifecycle requirement.
Learning Mode
Every AIPDS must declare one learning mode for the deployed Product version:
NONE
The deployed Product does not update intelligence through learning or retraining.
Typical examples:
- symbolic reasoning Product;
- fixed planner;
- rules/knowledge-based AI Product;
- Derived AI Product that consumes upstream intelligence but owns no trainable model.
NONE does not mean the Product can never evolve. Rules, knowledge, software, policies, or upstream dependencies may still change through controlled versioning.
PERIODIC_RETRAINING
A learned model is retrained on a planned schedule.
TRIGGERED_RETRAINING
Retraining occurs when declared performance, drift, risk, or business thresholds are met.
ONLINE_LEARNING
The intelligence mechanism updates incrementally during operation or near-operation.
HUMAN_GOVERNED_LEARNING
Learning updates require explicit human curation, approval, or supervised governance.
Evolution Mechanisms
An AI Product may evolve through one or more mechanisms independent of its learning mode:
- Model Retraining
- Model Replacement / Upgrade
- Rule Update
- Knowledge / Ontology Update
- Prompt / Policy / Guardrail Update
- Tool / Action Contract Update
- Upstream AI Product Dependency Update
- Software / Runtime Update
These mechanisms must be versioned and governed according to their effect on Product semantics, AI qualification, APT, risk, interfaces, and behavior.
Required Declarations
Every AIPDS must declare:
- learning mode;
- applicable evolution mechanisms;
- trigger / approval rules where evolution occurs;
- validation and regression requirements;
- versioning behavior;
- rollback / recovery strategy where applicable;
- qualification-currentness impact where a dependency or AI Primitive changes.
Where learning mode is not NONE, the Product must additionally declare relevant training/retraining data, approval, evaluation, and promotion controls.
AI Qualification Impact
Evolution may change more than performance.
If a Product version loses its only qualifying AI Primitive — for example a learned-model dependency is replaced with ordinary deterministic software — the resulting Product version may no longer qualify as an AI Product.
AIPDS must therefore surface ontology-significant evolution to AIPROD/Core qualification handling rather than treating it as an invisible runtime implementation change.
Example — Symbolic Policy Reasoner
- Learning Mode:
NONE - Evolution Mechanisms: Rule Update + Knowledge Update
- Approval: Policy owner + AI Product Owner
- Regression: canonical decision corpus + explanation-trace comparison
- Qualification: remains AI while a qualifying symbolic reasoning primitive remains in the constitution
Example — Fraud Detection
- Learning Mode:
TRIGGERED_RETRAINING - Evolution Mechanisms: Model Retraining + Model Replacement
- Trigger: precision / false-positive / drift thresholds
- Promotion: evaluation + risk approval before deployment
Principle: Evolution is universal; retraining is conditional. AIPS must govern both learning and non-learning AI Products without forcing one intelligence mechanism onto all Products.