Claim-Source Mapping: Every Fact Needs a Return Address

A research system produces a market analysis stating “The AI market is projected to reach $500 billion by 2025.” A reviewer asks: where does this figure come from? The report has no source attribution. The research sub-agents had the data — one returned {claim: "AI market projected at $500B by 2025", source_url: "...", document: "2024 AI Market Forecast"}. But the coordinator dropped the attribution during synthesis, converting structured claim-source mappings into plain text prose.

The figure might be correct. But without provenance, nobody can verify it, nobody can check whether the source is current, and if it is wrong, nobody can trace the error back to its origin.

What Claim-Source Mapping Is

A claim-source mapping pairs every factual assertion with the specific source that supports it:

{
  "claim": "Returns accepted within 30 days of purchase",
  "source_document": "return_policy_v3.pdf",
  "source_excerpt": "Section 4.2: All products may be returned within thirty (30) calendar days...",
  "last_updated": "2025-01-15",
  "page": 12
}

This is not a bibliography. A bibliography lists sources at the end of a document. Claim-source mapping links each individual claim to its specific source — per-claim, not per-document.

The difference matters when the coordinator synthesizes information from 3 sub-agents querying different knowledge bases. A document-level “Sources” section listing all three databases does not tell the reader which database supports the return policy claim versus the warranty extension claim. A claim-source index does.

Where Attribution Gets Lost

Attribution loss is almost always a synthesis problem, not a retrieval problem. Sub-agents typically return well-structured data with source metadata. The coordinator is where things go wrong.

Three failure modes:

Narrative merging. The coordinator blends findings from multiple sources into a single paragraph without tracking which claim came from which source. Once merged, the provenance is irrecoverable.

Natural language citations. Sub-agents embed citations in text (“according to the API docs…”). The coordinator rephrases or merges during synthesis, and the embedded citations break or disappear.

Document-level attribution. The coordinator links claims to the right document but not to the specific excerpt. When a claim contradicts the source (“covers accidental damage” vs the document’s “manufacturer defects only”), a 4-hour investigation is needed to determine whether the error was in extraction, synthesis, or the source itself. Excerpt-level mapping would have made the discrepancy visible in minutes.

Audit Data: 38% Unverifiable

One team audited their research system’s attribution quality:

38% of claims could not be verified. The fix requires addressing all three failure types: validate URLs at extraction time (18% broken links), require structured claim-source mappings instead of generic phrases (15% generic), and ensure every claim has attribution (5% missing).

Generic attribution — “studies show,” “research suggests,” “according to experts” — is never acceptable in a system that claims to provide traceable information. These phrases sound authoritative but are unverifiable. They should be treated as missing attribution.

End-to-End Provenance

In a multi-stage pipeline (analysis → synthesis → scoring → report), attribution must survive every stage. If any stage drops provenance, downstream stages cannot recover it.

The anti-pattern: Add attribution only at the final report stage, having the report generator look up sources retroactively. By the time the report is generated, claims have been rephrased, merged, and restructured through synthesis and scoring. Matching them back to original sources is unreliable.

The correct approach: Each pipeline stage passes through and preserves claim-source mappings from the previous stage. Provenance accumulates from initial extraction through the final output. The report inherits complete traceability from every upstream stage.

For multi-agent systems, this means:

1. Sub-agents return structured claim-source mappings (not plain text, not agent names as sources)
2. The coordinator maintains a provenance index during synthesis, assigning each claim a unique ID linked to its source
3. The final output includes per-claim attribution that links every factual statement to its specific source

Source Metadata Must Travel With the Data

A developer proposed removing source page numbers and section references from extraction output to save storage. The current output: {value: "$2.3B", source_page: 47, section: "Revenue Summary", document: "2024_annual_report.pdf"}. The proposed output: {value: "$2.3B"}.

For financial data, this is unacceptable. Without provenance, a wrong $2.3B figure cannot be traced back to the source document for correction. Auditors cannot verify the extraction against the original. The storage savings are negligible compared to the verification cost.

Storing metadata separately (in a lookup table) creates a fragile dependency. If the table is lost, corrupted, or falls out of sync, all provenance is lost. Source metadata should travel with the extracted value as a single unit.

Provenance Enables Error Tracing

A customer support system told a customer: “Extended warranty covers accidental damage.” The source document (warranty_policy_v3.pdf) actually states: “Extended warranty covers manufacturer defects only.” Investigation took 4 hours — the team had to determine whether the error was in the source document, the extraction, or the synthesis.

With excerpt-level attribution, the discrepancy would have been immediately visible:

{
  "claim": "Extended warranty covers accidental damage",
  "source_excerpt": "Extended warranty covers manufacturer defects only",
  "source": "warranty_policy_v3.pdf, Section 7, page 4"
}

The claim contradicts the excerpt. The error is in the synthesis step. Investigation time: minutes, not hours.

This is the operational value of claim-source mapping: not just attribution for the reader, but error tracing for the team. When something goes wrong — and in production, things go wrong — provenance is the difference between a 4-hour investigation and a 4-minute fix.

One-liner: Link every claim to its specific source (document, page, excerpt) at extraction time and preserve that link through every pipeline stage — facts without provenance are unverifiable.