5.6 Provenance and uncertainty in multi-source synthesis
Attribution loss
Source attribution is lost during summarization when findings are compressed without preserving claim → source mappings — the multi-agent equivalent of losing data lineage. The defence is a structured claim-source mapping that the synthesis agent must preserve and merge:
{
"claim": "Streaming revenue grew 14% YoY.",
"excerpt": "…revenue rose 14 percent year over year…",
"source": { "name": "IFPI Global Report", "url": "https://…", "published": "2026-02-11" },
"support": "well-established"
}
Conflicting values from credible sources
| Do | Do not |
|---|---|
| Annotate the conflict with source attribution — carry both values forward | Arbitrarily select one value |
| Complete document analysis with both conflicting values included and explicitly annotated, letting the coordinator decide how to reconcile before synthesis | Have the analysis subagent silently pick a winner |
| Structure the report with explicit sections separating well-established from contested findings, preserving original source characterizations and methodological context | Flatten everything into one confident narrative |
Temporal data
Require publication or data-collection dates (ISO 8601) in structured subagent outputs. Without dates, a 2024 figure and a 2026 figure look like a contradiction rather than a trend.
Render by content type
In synthesis output, render each content type appropriately — financial data as tables, news as prose, technical findings as structured lists — rather than converting everything to one uniform format.
Quiz
- Question 1
Two credible sources give different figures for the same market: an analyst note collected February 2025 says $4.2B, and a regulatory filing covering the period to November 2024 says $3.6B. The synthesis agent currently takes the larger figure and reports a single number. What should it do instead?
- Question 2
The search subagent returns findings with source URLs attached. The analysis subagent condenses those into key points and drops the URLs as redundant detail. Synthesis then produces a report in which no claim can be traced to anything. Where does the fix belong?
- Question 3
A report states that two sources contradict each other on an adoption rate, quoting 34% and 51%. Inspection shows one figure comes from a 2022 survey and the other from a 2025 survey, and that neither date travelled through the pipeline. What is the structural fix?
- Question 4
Your extraction summary converts every content type into one uniform bulleted format: financial tables become bullet lists of numbers, narrative clauses become bullets, and payment schedules become bullets. Downstream consumers report that the financial data is now hard to read and that schedule rows occasionally appear out of order. What principle is being violated?