Use adjacent turns to classify replies without inventing causes

When a conversation-level classifier leaves a large residual, decompose it into reply states by using only the immediately preceding human-support turn within a bounded window. Keep explicit problem statements at higher priority, and report contextual states—such as a response to a link, retry instruction, evidence request, callback, or closure—separately from root causes. This yields operational signal without claiming that the prior support action caused the issue. Generic fragments with no adjacent prompt should remain unresolved.

Applies when: issue-level analysis has many short or referential messages that require nearby conversation context. Keywords: support analytics conversation classification adjacent turn reply state causality