HEART ai.model-drift H-07005
AI model drift. The AI model's answers have changed in quality or behaviour.
| Code | HEART ai.model-drift; inside a message "code": "ai.model-drift" |
|---|---|
| H-number | H-07005: area 07, code 005 |
| Area | ai: AI providers, models, budgets |
| Typical severity | warning. A product may set another for its own case. |
| Applies to | Products that use an AI model |
Likely causes
- The provider changed or updated the model behind the same name.
- The product's data has changed, so the model sees things it was not tested on.
- A new model was adopted that behaves differently.
How to fix or prevent it
- Compare recent answers against the product's own checks and test cases.
- Pin a model version if the provider allows it.
- Roll back to the model that worked, and test the new one before adopting it.
What the HEART does by itself
It raises a warning when the quality or behaviour of answers changes, measured against its own checks. It may stop using the model for actions while the alert is open.
When to call service
If the model takes part in decisions that matter, and the drift cannot be explained.
Related codes
ai.output-rejectedH-07006: AI output rejectedai.answer-rejectedH-07009: Brain answer rejectedai.disagreementH-07010: Brain disagrees with the system's checks
Inside a message a code is written bare, "code": "storage.full". In logs, tickets, documents and searches it is written
HEART storage.full. Things that simply happen, such as system.started, are event types, not faults.
The codes are defined in Part 3, Section 5.4, and their H-numbers in Section 5.7.
All of it, as data: fault-codes.json.
All fault codes, by area · All fault codes, by H-number · Problem types (API errors)