What problem does it solve? When an initial retrieval or answer attempt fails the evidence-sufficiency gate, agents need a disciplined way to decide what happens next instead of looping endlessly or hallucinating. This Skill defines exactly one bounded retry, escalation to the RLM engine, or an honest "insufficient evidence" response, plus a safe process for proposing small, reversible changes to memory entries, skill descriptions, and prompt notes. ## Core Features & Use Cases - Bounded Corrective Retrieval: Enforces exactly one query-rewrite retry after a failed evidence check, then escalates aggregation-style queries to the RLM orchestrator or returns an honest insufficient-evidence response. - Failure-Targeted Query Rewriting: Maps verification failure reasons (no evidence retrieved, low aggregation coverage, unsupported claims) to specific rewrite strategies rather than paraphrasing the same query. - Supplemental-State Refinement: Proposes minimal, evidence-linked, reversible changes to memory entries, skill descriptions, and prompt notes, with logging to events_trace. - Use Case: A RAG query returns claims the verification gate flags as unsupported. Use this Skill to rewrite the query toward the missing sub-topic, retry once, and if it still fails, escalate to recursive language model processing instead of retrying indefinitely. ## Quick Start Use the refinement skill to decide what to do after the verification gate flags my last answer as having unsupported claims.
