Model AI failure by cause, consequence and recovery
Refusal, safety block, timeout, rate limit, unavailable tool, malformed output, partial success and unsafe result remain distinct states.
実践ガイダンス
Preserve valid work, say what did and did not happen, provide a specific recovery and protect side effects from duplicates. Log enough context for accountability without exposing secrets or private reasoning.
推奨
- Retain valid partial results
- Offer cause-specific recovery
- Make side-effect retries explicit and idempotent
非推奨
- Collapse every failure into Try again
- Discard user input
- Reveal chain-of-thought as an explanation
検証契約
Refusal, timeout, 429, unavailable-tool, malformed, partial-success and duplicate-side-effect tests
根拠リファレンス
ref.apple.hig-machine-learning— Apple HIG — Machine Learningref.microsoft.fluent-ai-harm— Fluent 2 — AI Harmref.microsoft.human-ai-guidelines— Microsoft Guidelines for Human-AI Interactionref.nist.ai-rmf— NIST AI Risk Management Frameworkref.vercel.ai-elements— Vercel AI Elementsref.zendesk.garden-components— Zendesk Garden componentsref.zendesk.garden— Zendesk Garden