Glossary / Canonical epistemic failure modes
Silent Instruction Relaxation
Silent Instruction Relaxation is the weakening of a constraint across turns without acknowledgment. The instruction still sits in context but no longer binds behavior, and nothing flags that it has lapsed. Multi-turn agents drift away from their guardrails precisely when no one is re-checking the early instructions.
How it manifests · Caught in a framework
Caught in a real framework: Microsoft Semantic Kernel's chat-history truncation reducer silently deleted system and developer prompts as conversations grew, so the agent stopped following its own instructions with no log signal. Our merged fix (PR 13610) preserves them.
Evidence: Your framework deleted your instructions
Related terms
- Constraint Evasion · Canonical epistemic failure modes
- Silent AI failure mode · Core concepts