paper-ranker

name	paper-ranker
description	Apply a conservative screening rubric to rank papers and separate selected, ambiguous, and rejected items with reasons and evidence references. Use when a candidate set already exists and final ranking or grouping is needed.

paper-ranker

Name

paper-ranker

Description

Apply a conservative screening rubric to separate selected, ambiguous, and rejected papers with reasons and evidence references.

Use When

• the user wants final selection or ranking
• a candidate set already exists
• abstract and optional full-text evidence need to be turned into decisions
• the workflow must produce reproducible, structured outputs

Inputs

• normalized paper records
• evidence entries
• inclusion criteria
• exclusion criteria
• rubric dimensions and thresholds
• require_fulltext_for_selection

Outputs

• relevance score
• decision: selected, ambiguous, or rejected
• reasons
• evidence references
• grouped result sets

Workflow

1. Start from the screening_candidates set.
2. Review criteria and available evidence.
3. Score each paper on:
   • topic relevance
   • method match
   • dataset match
   • supervision or annotation match
   • evidence confidence
4. Check exclusion criteria and missing required evidence.
5. Apply conservative decision rules.
6. Return grouped outputs and recommended follow-up actions.

Guardrails

• Apply the rubric conservatively.
• Separate selected, ambiguous, and rejected.
• Do not convert uncertainty into selection.
• If full text is required for selection and unavailable, keep the paper ambiguous.
• Provide reasons and evidence references for every decision.

Online And Offline Behavior

• works in either mode because ranking depends on normalized evidence, not on network access
• weak evidence in offline-only metadata should reduce confidence rather than inflate score

Example

Decision pattern:

• selected: strong topic and supervision match with sufficient evidence
• ambiguous: promising paper but missing method, dataset, or full-text confirmation
• rejected: exclusion triggered or relevance too weak