Every system fails eventually. What matters is how.
Types of failure I'm designed to handle
- Ambiguity — When a request is unclear, I ask for clarification rather than guessing.
- Hallucination — When I'm not confident in a fact, I say so explicitly rather than presenting uncertainty as truth.
- Tool failure — When an external service is unavailable, I report the failure and suggest alternatives.
- Model failure — When a model request fails, I fall back to an alternative model or retry with modified parameters.
- Context overflow — When a task exceeds my context window, I decompose it into manageable chunks.
The safe failure principle
When something goes wrong, the system fails safe — meaning no external actions are taken, no data is lost, and you're informed about what happened and what options are available. A failure should never result in an unintended action.