Run3 multi hop evidence extraction

Multi-Hop Evidence Extraction Skill

Step 1: Tier 1 - Explicit Reviewer/Approver Fields

• Locate explicit "reviewers", "approvers", "assigned_to" fields in artifacts
• Extract all names/IDs listed
• Track the source field name for each person

Step 2: Tier 2 - Substantive Feedback Contributors

• Search artifact for linked Slack conversations and meeting transcripts
• For Slack replies: Extract names of people who provided substantive feedback or comments
  • Do NOT include people who merely reacted or said "thanks"
  • Include people who asked questions, suggested changes, or provided critique
• For meeting transcripts: Extract names of people who:
  • Actually spoke in the transcript (check dialogue)
  • Provided substantive feedback, suggestions, or decisions
  • Do NOT include people listed only in a participants array without speaking

Step 3: Multi-Hop Reference Traversal

• For each extracted person/artifact, check if they reference other artifacts or people
• Follow "related_artifacts", "depends_on", "references" links
• Recursively extract reviewers from referenced artifacts
• Traverse up to 3 hops to catch indirect relationships
• Track the traversal path for validation

Step 4: Tier 3 - Other Identified Contributors

• Identify any other people who contributed (authors, editors, signers)
• Exclude people already captured in Tiers 1-2
• Only include if they have a substantive role

Step 5: Deduplicate and Return

• Remove duplicates across all three tiers
• Return as a list of unique names/IDs
• Document which tier each person came from (for debugging)