Source attribution is lost during summarization when findings get compressed without preserving the mapping from claim to source. The fix has to be upstream, not downstream: require each subagent to output structured claim-source mappings — source URLs, document names, relevant excerpts — that downstream agents can preserve and merge as they synthesize. A downstream instruction to "include citations" cannot recover information the upstream step never captured.
Handling genuine conflict between sources:
- Conflicting statistics from credible sources should be annotated with source attribution, not arbitrarily selected between, and not averaged — averaging two disagreeing numbers manufactures a third number nobody actually reported.
- Reports should have explicit sections distinguishing well-established findings from contested ones, so a reader can calibrate trust per section rather than treating the whole report as equally solid.
- A document-analysis subagent should complete its work with conflicting values explicitly annotated, leaving the reconciliation decision to the coordinator at synthesis time, rather than quietly resolving the conflict itself and passing along only one number.
- Temporal data needs publication or collection dates carried in structured outputs, so that a 2023 figure and a 2026 figure read as a trend rather than an unexplained contradiction.
- Render by content type — financial data as tables, news as prose, technical findings as structured lists — rather than flattening every source into one uniform format, which tends to erase exactly the structural cues that made the original data trustworthy.
The discriminator: provenance is preserved by structure created upstream, never reconstructed downstream. If the scenario's failure is "the final report can't say where a number came from," no downstream fix — no matter how carefully worded — recovers information that was never captured in the first place.
Traps to recognize:
- Instructing the synthesis agent to "include citations" it no longer actually holds — the information needed to cite correctly was already lost upstream.
- A bibliography listing every source touched, without mapping specific claims to specific sources — technically has "sources," but can't answer "where did this number come from."
- Re-searching each claim after the fact to recover its source — expensive, unreliable, and treats a structural problem as a lookup problem.
- Subagents silently resolving conflicts on their own rather than surfacing them — this looks like progress but actively destroys the information the coordinator needed to reconcile the conflict correctly.
- Storing the full raw text of every source directly in the coordinator's context, reasoning that nothing gets lost if nothing gets summarized — this looks maximally safe but fails three ways at once: it destroys the context budget long before synthesis is done, it still doesn't preserve a claim-to-source mapping (raw text sitting in context isn't the same as a structured link from a specific claim to where it came from), and it actively invites lost-in-the-middle failures (§1) by burying exactly the details that matter in a wall of undifferentiated source text.