Provenance and uncertainty in multi-source synthesis

Keep each claim connected to its source

If a summary removes the claim-to-source mapping, the claims can no longer be traced to their evidence. A later agent can't restore a source reference that was not preserved.

A box of old family photographs with nothing written on the back is not damaged; every picture is sharp. What is missing was never in the image — who, and when. Nobody can look harder at a photo and recover it.

Require subagents to output structured claim-source mappings that include source URLs, document names, and page or section locations. The synthesis agent should preserve these mappings while combining the findings.

A record of how the subagent reasoned isn't one of those. Provenance points outward, at evidence a reader can open; a reasoning chain points inward, at the path the model took to the claim. Those chains are what upstream agents should be trimming out, not forwarding as citations.

Conflicting figures are information

When two credible sources give different statistics, selecting one arbitrarily throws away the most interesting fact in the material.

Annotate the conflict with attribution: this source reports X, that one reports Y. The reader can then judge, which is exactly what they can't do if the synthesis silently picked a side.

Two near-misses sit either side of this. Averaging the figures produces a number no source reports and nothing can be attributed to. Noting that the sources disagree without giving both figures discloses the conflict and withholds the data a reader needs to weigh it.

Dates prevent false contradictions

Two figures that disagree may not conflict at all — they may describe different points in time.

Require the measurement or reference period for each figure. Keep publication date and collection date as separate metadata because a report published in Q4 may describe Q2. Market share measured in Q1 and Q3 are two facts, not a disagreement, and the reference periods are what make that visible.

Here is one claim-source mapping using fields defined by the application:

{
  "claim": "Vendor A held 34% of the market.",
  "source": "https://example.com/market-report-2026.pdf",
  "locator": "page 12, table 3",
  "published": "2026-11-04",
  "measurement_period": "2026-Q1"
}

published records when the report was released. measurement_period records when the reported data was measured. A second source reporting 41% for 2026-Q3 doesn't contradict a figure measured in 2026-Q1. The locator field tells a reader where to verify the claim in the source.

Methodology explains many divergent statistics and isn't a default explanation. Attaching it to a difference the reference periods already account for tells the reader something the sources never said. State the cause you have evidence for, and otherwise state the figures and their periods.