Evidence Grading Framework

A reusable rubric that grades every source by type, recency, authority, independence, and corroboration, then ranks them and resolves conflicts by evidence weight.

Install
cmdop skills install agensi-evidence-grading-framework

Evidence Grading Framework is a skill that sits in front of a RAG pipeline or research agent and grades every source before the agent writes a single sentence. Rather than treating all retrieved documents as equally trustworthy, it assigns each source a structured score across five dimensions: TYPE (primary, secondary, tertiary, or unranked), RECENCY (currency relative to how fast the topic changes), AUTHORITY (domain-specific credibility for the specific claim), INDEPENDENCE (whether the source has a stake in the conclusion, flagging vendor and marketing pages), and CORROBORATION (how many independent sources agree, so one source republished ten times still counts as one).

From those scores the framework derives an A–D overall grade. Grade A sources are primary, current, authoritative, and independent — rely on them. Grade B sources are solid secondaries with minor weaknesses — usable but corroborate key figures. Grade C sources are tertiary, dated, or non-independent — treat as leads only. Grade D sources are unranked, conflicted, or contradicted by stronger evidence — do not rely on them.

When two sources disagree, the framework resolves the conflict by evidence weight rather than recency or confidence: the higher-graded source wins, ties go to the better-corroborated claim, and genuinely unresolved conflicts are reported as open. The output is a graded source ledger with a clear recommendation. The skill evaluates the sources provided to it; it does not independently search for additional sources.

Use cases

  • Rank retrieved documents in a RAG pipeline before generation so weak sources are down-weighted automatically
  • Triage a research reading list into rely, corroborate, and avoid tiers
  • Resolve contradictory sources using an evidence-weighted rule instead of defaulting to the most recent or most confident
  • Produce an auditable graded source ledger to attach to research deliverables
  • Pair with a claim-checking step by grading sources first, then verifying claims only against A and B sources
  • Flag vendor and marketing pages as non-independent before they influence an agent's answer

When to use it

  • The agent retrieves multiple sources of varying quality and needs a principled way to weight them before generating
  • The task requires a defensible, auditable record of which sources were relied upon and why
  • Sources conflict and recency or confidence alone is not a reliable tiebreaker
  • The research question is in a domain where source provenance matters, such as technical specifications, regulatory records, or scientific claims

When not to use it

  • The agent needs to search for better sources — this skill grades only the sources already provided to it
  • Absolute truth verification is required; grades reflect source quality and provenance, not factual correctness
  • Hidden conflicts of interest are a primary concern, since independence assessment depends on what a source discloses
  • The pipeline has no retrieval step and operates on a single trusted source with no alternatives to rank