Skip to main content

Monitoring & Observability

A deployed AI Product must be observable in operation. The signals required depend on its Intelligence Capability, Agency/Autonomy, Embodiment, Intelligence Effects, risk tier, and underlying intelligence mechanisms.

AIPDS therefore uses applicability-aware observability rather than assuming model accuracy and drift are universal.


Monitoring Dimensions​

1. Operational Health​

  • availability and error rates;
  • latency / deadlines / throughput;
  • resource utilization;
  • dependency health;
  • queue/session/agent state where applicable.

2. Intelligence Quality​

Signals appropriate to the Product's IC may include:

  • accuracy / precision / recall / F1;
  • groundedness or factuality;
  • recommendation quality;
  • planning success rate;
  • rule-consistency / reasoning correctness;
  • goal completion;
  • control stability;
  • domain-specific effectiveness measures.

3. Behavioral & Agency Signals​

For agentic or action-capable Products:

  • tool/action invocations;
  • approval requests and overrides;
  • policy refusals;
  • goal/subgoal progression;
  • unexpected action sequences;
  • human handoffs.

4. Fairness & Impact​

Where applicable:

  • group fairness measures;
  • outcome disparity;
  • distributional impacts;
  • policy-specific harm indicators.

5. Drift / Change​

Where learned or data-sensitive mechanisms are present:

  • data drift;
  • concept drift;
  • model performance drift.

For non-learning AI, equivalent change signals may include:

  • rule/knowledge version mismatch;
  • ontology inconsistency;
  • stale external knowledge;
  • upstream AI Product semantic/currentness change.

6. Explainability / Evidence​

Depending on mechanism:

  • feature attribution;
  • evidence traces;
  • reasoning traces;
  • rule provenance;
  • plan/action rationale;
  • cited knowledge sources.

7. Security & Abuse​

  • adversarial or malicious input detection;
  • prompt/tool injection where applicable;
  • prohibited-use attempts;
  • anomalous access/action patterns.

8. Physical Safety​

For embodied/actuating AI:

  • control-loop health;
  • sensor degradation;
  • safety-envelope violations;
  • actuator constraints;
  • collision / near-miss signals;
  • emergency-stop or safe-state activation.

Observability Mechanisms​

Signals should be exposed through machine-readable metrics, events, traces, logs, evidence records, and alerts suitable to the deployment environment.

Observability evidence must be:

  • attributable to a specific Product version and deployment;
  • auditable;
  • retention-governed;
  • accessible to authorized Product and governance actors.

Principle: Observability must monitor the actual intelligence and effects of the Product, not merely the infrastructure or model server underneath it.