Information Provenance & Claim-Source Mappings

Core

Preserve information provenance and handle uncertainty in multi-source synthesis · Difficulty 3/5

0%
provenanceattributionsynthesisclaim-source

Source attribution is frequently lost during multi-agent synthesis pipelines when findings are compressed without preserving claim-source mappings.

The Problem

A web search agent finds data from 5 sources. A document analysis agent extracts findings. A synthesis agent combines everything into a report. By the final report, claims appear without sources because:

  • Summarization compressed away source URLs
  • Findings were merged without tracking which source said what
  • Conflicting values were silently resolved
  • Solution: Structured Claim-Source Mappings

    Require subagents to output structured mappings:

  • Source URLs and document names
  • Relevant excerpts (not just summaries)
  • Publication/collection dates
  • Methodological context
  • The synthesis agent must preserve and merge these mappings when combining findings.

    Report Structure

    Structure final reports to distinguish:

  • Well-established findings: Multiple concordant sources
  • Contested findings: Sources disagree, with original characterizations preserved
  • Single-source findings: Only one source, flagged for lower confidence
  • Content-Appropriate Rendering

    Render different content types appropriately in synthesis outputs:

  • Financial data as tables
  • News as prose
  • Technical findings as structured lists
  • Don't convert everything to a uniform format -- it loses the natural structure of different content types.

    Key Takeaways

    • Require structured claim-source mappings from subagents to preserve attribution
    • Distinguish well-established, contested, and single-source findings in reports
    • Render content types in their natural format rather than forcing uniform structure