Preserve information provenance and handle uncertainty in multi-source synthesis

Provenance is the record of where each claim came from: which source, which supporting excerpt, and when it was published or collected. In a multi-agent research pipeline (Scenario 3), information passes through several compression steps. A search agent gathers pages, an analysis agent summarizes them, and a synthesis agent merges everything into a report. Every one of those steps is a summarization boundary, and summarization is exactly where attribution silently dies.

The failure is dangerous because the output still looks polished. The analysis agent reads a page that says "According to the 2024 IFPI report, streaming revenue grew 10.2%" and helpfully condenses it to "streaming revenue grew about 10%." The number survived; the source and the date did not. By the time the synthesis agent sees it, there is no way to cite the claim or reconcile it against a competing figure. The report ends up full of confident, uncitable assertions, which is fatal for a system whose whole job is to produce cited reports.

The fix is to treat attribution as data that must be carried, not prose that can be paraphrased away. Instead of asking subagents to "summarize," require them to emit structured records in which the claim and its source travel as separate fields. Compression then acts on the claim text while the source name, URL, excerpt, and date ride along untouched.

A claim-source mapping is a structured object that binds a single claim to its origin. At minimum each record carries the claim itself, a verbatim excerpt that supports it, the source identifier (URL or document name), and the publication or collection date. Optionally it also carries a page number or a relevance score.

{
  "claim": "Global streaming revenue grew 10.2% year over year",
  "excerpt": "Streaming revenues increased 10.2% to reach US$19.3 billion",
  "source": "IFPI Global Music Report 2024",
  "url": "https://ifpi.org/...",
  "published": "2024-03-21"
}

The rule that makes this work: downstream agents must preserve and merge these mappings, never flatten them. When the synthesis agent turns twelve findings into three paragraphs, every sentence it writes should still trace to the record or records that support it. Requiring the subagent to output this structure, rather than trusting it to remember to cite, is what guarantees provenance survives synthesis.

This is why separating content from metadata matters. The claim is content; the source, URL, and date are metadata. Keeping them in distinct fields lets a compression step shorten the content without ever touching the metadata.

Credible sources routinely disagree. Two reputable research firms may report different figures for the same market in the same year. The wrong move, the one exam questions punish, is to let an agent silently pick one number and drop the other. That destroys information and hides a genuine disagreement from the human who is best placed to judge it, and the value that survives is often just the last one the model happened to read.

The correct pattern is to annotate the conflict with source attribution and carry both values forward. The document analysis subagent completes its analysis with the conflicting values included and explicitly flagged, rather than resolving them itself. It hands both to the coordinator, whose job is to decide how to reconcile, or whether to reconcile at all, before passing to synthesis. The synthesis agent then preserves the conflict in the output, attributing each figure to its source.

{
  "metric": "2024 generative-AI market size",
  "conflict_detected": true,
  "values": [
    {"value": "$25.6B", "source": "Grand View Research", "published": "2024-05"},
    {"value": "$40.1B", "source": "Bloomberg Intelligence", "published": "2024-09"}
  ]
}

Note the division of labor: analysis surfaces and annotates, the coordinator decides, synthesis renders the decision transparently. No layer arbitrarily overwrites a credible source.

Many apparent contradictions are not contradictions at all. They are the same quantity measured at different times. "Enterprise adoption was 33%" and "enterprise adoption reached 65%" look like a conflict until you learn the first figure is from 2023 and the second from 2024, at which point they are simply two points on a rising trend.

Because of this, every structured finding should carry a publication or data-collection date. Without dates, a synthesis agent (or a later reader) cannot distinguish a genuine disagreement between sources from ordinary change over time. It will either flag a false conflict or, worse, average two figures that were never meant to be compared.

Make the date field non-optional for time-sensitive metrics in the subagent's output schema. Then the synthesis step can interpret correctly: same metric plus different dates equals a time series to present as a trend; same metric plus the same period plus different values equals a real conflict to annotate with source attribution.

A trustworthy synthesis output does not present every claim with equal confidence. It separates what the sources agree on from what they dispute. The recommended structure uses explicit sections: a well-established findings section for claims that multiple credible sources corroborate, and a contested (or open-questions) section for claims where sources disagree or the evidence is thin.

Within the contested section, preserve each source's original characterization and its methodological context rather than blending them into a bland consensus. If one study surveyed 200 professionals and another analyzed 50,000 transactions, that methodological difference is precisely what a reader needs in order to weigh the claims, so it must survive into the report.

This structure pairs naturally with the coverage annotations from error propagation (task 5.3): the synthesis can mark which topic areas are well-supported and which have gaps because a source was unavailable. The reader then knows not only what was found, but how solid each part of it is.

A common degradation is flattening everything into uniform prose. Financial figures, breaking-news items, and technical specifications each have a natural representation, and forcing all of them into paragraphs destroys usability and can obscure provenance.

The skill is to render content types appropriately in the synthesis output: financial and comparative data as tables (where a conflicting-values row can give each source its own column), news and narrative developments as prose, and technical findings as structured lists. A conflict between two market-size figures is far clearer in a two-row table with a source column than buried inside a sentence.

This is not cosmetic. Tabular rendering of conflicting statistics keeps each value bound to its source and date, which is provenance preservation by layout. Choosing the representation that fits the data is part of preserving both meaning and attribution, rather than converting everything to one homogeneous format.