← ClaudeAtlas

evidence-graderlisted

Skill counterpart to evidence-curator agent. Provides on-demand grading for ad-hoc claims, source assessments, and review passes. Operates over flat memory/ layout.
kanadhiayash/zeref-memory-engine · ★ 2 · AI & Automation · score 72
Install: claude install-skill kanadhiayash/zeref-memory-engine
# evidence-grader ## Mission Assess confidence + recency + provenance of a claim or wiki entry. Returns a grade with reasoning. ## Grades | Grade | Criteria | |---|---| | **high** | Verified this session OR confirmed by user OR backed by primary source | | **medium** | Reasonable inference OR named source < 30d OR indirect but corroborated | | **low** | Unverified assumption OR > 90d OR derived from a single weak source | ## Operations ### GRADE_CLAIM (on demand) 1. Receive claim + optional source 2. Apply criteria 3. Return: `{grade, confidence_reasoning, recency_reasoning, provenance_reasoning, suggested_action}` ### REVIEW_PAGE (called by wiki-maintenance) 1. Walk every entry on a wiki page (`memory/DECISIONS.md`, `memory/RISKS.md`, `memory/OPEN_QUESTIONS.md`) 2. Re-grade each 3. Report demotions, suggest re-verifications ### FILTER_SYNC (called by parent-sync) 1. Receive list of entries staged for outbound 2. Return only entries with grade ≥ medium ### TWO_STRIKES_GRADE (called when promoting a trap to a rule) 1. Receive the two `memory/MEMORY.md` `## Traps observed` entries 2. Verify both occurrences are concrete and same trap class 3. Return grade per `references/two-strikes-rule.md` ## Safety - Grade is advisory only; user remains the arbiter - Always include reasoning, never just a grade letter