Task statement 5.6 · Context management and reliability
Preserve information provenance and handle uncertainty in multi-source synthesis
Three credible sources give three prices for battery packs: an industry survey says $139/kWh, a national lab $115 and a trade association $124. $139 is from December 2023, so it is simply older. $115 and $124 are both 2024 figures that genuinely disagree. The draft report says only "$124/kWh", with no source or date, so a reader cannot tell the two cases apart.
share
The tempting move is to give the reader one clean number. Picking the most recent or most authoritative-sounding value looks decisive, but it hides a disagreement the reader needed to weigh. Averaging $115 and $124 looks fair, but no source reported $119.50, so nothing can be cited for it. Passing findings on as tidy prose looks efficient, yet each summary can drop a claim's source and date, and no later agent can recover them.
KnowledgeWhere attribution is lost
Provenance is where a claim came from: its source, the passage and the date. In a multi-agent system it is lost at a summarisation step. When a subagent compresses its findings into prose, "pack prices are around $115–139 and falling" keeps the gist but drops which source said what, and when. Synthesis can only cite what reaches it, so the fix sits upstream: each subagent returns structured claim-source mappings, and every later agent preserves and merges them.
- Each subagent returns claim-source mappings, each with its publication or collection date.
- Document analysis completes with every value included, and annotates each conflict with the sources and methods instead of choosing.
- The coordinator decides how to reconcile: send a search for a third source, or pass both values on as contested. Here no comparable third source exists, so it passes both.
- Synthesis merges the mappings. It may shorten the wording, but every claim keeps its source and date.
- The report separates well-established from contested findings and gives each content type its natural form.
Interactive · three findings through synthesis
Choose how findings travel between agents, and see what the report can still say about them.
Each prose summary kept the gist and dropped the sources. By the report, one number stands alone and nothing can be traced.
The mappings survive every handoff. The report can separate what is settled from what is contested, and name who said each.
KnowledgeWhat one mapping carries
The mapping holds the claim, the source URL or document name, and the relevant excerpt, so a reader can check the claim. Two more fields stop two misreadings. The publication or collection date shows that $139 (2023) and $115 (2024) describe different years, so a temporal difference is not reported as a contradiction. The method keeps the source's methodological context, so a bottom-up estimate is not merged with a member survey. Here is the national lab's finding, with the conflict annotation from step 2.
// one finding, as the document analysis subagent returns it { "claim": "Average battery pack price was $115/kWh", "source": "National lab annual storage cost report", "source_url": "https://example.org/storage-costs-2024.pdf", "excerpt": "…we estimate pack-level cost averaged $115/kWh…", "collected": "2024", "published": "2025-02", "method": "bottom-up cost model (estimate)", "conflicts_with": ["trade-association-2024"] }
SkillAnnotate the conflict, then let the coordinator decide
Why did the draft say "$124/kWh" alone? Document analysis had been told to "return the best figure for each metric", so it chose the trade association's value during extraction, and the $115 figure never reached the coordinator. That is arbitrary selection, though it looked like judgement. The fix lets document analysis complete with both values included and the conflict annotated. The coordinator then decides, because only it can send another search. It asked web search for a third 2024 figure on a comparable method, found none, and passed both on as contested:
// coordinator → synthesis { "conflict": "2024 pack price", "values": ["national-lab-2024", "trade-association-2024"], "decision": "contested: no third source with a comparable method", "instruction": "report both with source and method; prefer neither" }
SkillA report that shows what is settled, in the right form
A finding is well established when the sources that address it agree, and contested when credible sources disagree for the same period and scope. A separate section for each lets the reader see what to rely on. The report also preserves each source's original characterisation: the lab calls $115 an estimate, so the report must not say "the lab found". Synthesis then renders each content type in its natural form rather than one uniform format. Financial data goes in a table, so values line up. News goes in prose, because an event needs its sequence and cause: as bullets, "cuts announced · lithium cheaper" no longer says which caused which. Technical findings go in a structured list, one checkable item each.
## Battery pack prices (illustrative figures and sources) ### Well established Prices fell: both 2024 figures are below the Dec 2023 survey. | Source | Value | Period | Method | |-------------------|----------|----------|------------------------| | Industry survey | $139/kWh | Dec 2023 | buyer survey | | National lab | $115/kWh | 2024 | bottom-up model (est.) | | Trade association | $124/kWh | 2024 | member survey | ### Contested The 2024 level: the lab estimates $115; the association reports $124. The methods differ and no third source was found. Neither is preferred. ### Recent news In February 2025 two cell makers announced further price cuts, citing cheaper lithium (trade press, 2025-02). ### Technical findings - Chemistry: LFP in 9 of 12 projects reviewed (national lab, 2024) - Pack share of installed system cost: 41% (national lab, 2024)
What this looks like in the platform. With citations: {"enabled": true} on a document block, the Claude API returns each supporting passage in cited_text, which the docs say is "guaranteed to contain valid pointers to the provided documents" (checked 2026-10-01). The guide does not name this feature; it tests that the mapping survives every handoff.
| Traced; sources agree or differ only by date | Well established. Show dated values as a trend. |
| Traced; sources disagree for the same period | Contested. Show each value with source and method. |
| Not traced | A mapping was dropped upstream. Restore it there. |
- Attribution is lost where findings are compressed into prose, so claim-source mappings with dates must survive every handoff.
- Document analysis includes and annotates conflicting values; the coordinator decides whether to seek another source or pass them on as contested.
- Financial data as tables, news as prose, technical findings as structured lists.