Preserve information provenance and handle uncertainty in multi-source synthesis

Summarise each agent's findings, then cite sources at the end of the report.

The mapping from claim to source is destroyed at the summarisation step, so the final citations are reattached by inference and can be attached to the wrong claim.

Carry claim-source pairs as structure through every hop; citations are not reconstructible afterwards.

Pick the most credible source when two figures disagree.

It presents a contested number as settled and hides the disagreement, which is often the most useful finding in the report.

Report both, attributed, and let the reader see the conflict.

Have the analysis agent resolve conflicts before synthesis, to keep the report clean.

The reconciliation happens where the least context exists, and the coordinator never learns there was a conflict at all.

Pass conflicting values forward explicitly annotated, and reconcile where the whole picture is visible.

Treat two different figures for the same measure as a contradiction.

They are often the same measure from different years, so a genuine time series gets reported as a data-quality problem.

Require dates in structured output so temporal difference is distinguishable from disagreement.

Render the synthesis uniformly as prose so it reads as one document.

Financial series flattened into sentences lose the comparisons that made them worth gathering, and precision is lost in the translation.

Render by content type — tables for figures, prose for narrative, structured lists for technical findings.