5.6 Preserve information provenance and handle uncertainty in multi-source synthesis
Preserve claim-to-source mappings through every summarisation and synthesis step, annotate conflicting values with their sources and dates instead of picking one, and present findings in ways that keep certainty and content type visible.
Key points
- 1
Source attribution is lost at the summarisation step whenever findings are compressed into prose without a claim-source mapping. It cannot be recovered afterwards by a "cite your sources" instruction, by appending raw outputs, or by searching the web for something that agrees with the claim.
- 2
Require subagents to output structured claim-source mappings: the claim, the source URL or document name, and the relevant excerpt. Downstream agents must preserve and merge these mappings through synthesis.
- 3
When a coordinator must condense subagent outputs to fit a budget, keep the full outputs in external storage and carry compact claim-source references (source id, excerpt locator) in the summary; run the citation pass against the stored originals. Anthropic's research system uses a dedicated citation agent that works from the source documents.
- 4
Conflicting statistics from credible sources: present both values, each attributed to its source with that source's definition and methodology, and flag the conflict. Do not average, pick by authority, or omit the figure.
- 5
Temporal data: require publication and data-collection dates in structured outputs so that a 2022 figure and a 2024 figure are read as change over time rather than a contradiction. "Larger means newer" is a guess, not a rule.
- 6
A document-analysis agent that finds conflicting values within its corpus should complete its analysis with both values included and explicitly annotated, and let the coordinator decide how to reconcile them before synthesis. Halting to ask blocks the pipeline; silently keeping the first value hides the conflict.
- 7
Structure reports with explicit sections that distinguish well-established findings from contested ones, preserving the original sources' characterisations (their hedging) and methodological context. A blanket disclaimer or a manufactured consensus does not tell the reader which claims are safe to act on.
- 8
Render content types appropriately in synthesis: financial data as tables, news as prose, technical findings as structured lists. Converting everything to one uniform format strips comparability from numbers and nuance from contested claims.
- 9
Provenance and error propagation meet in coverage annotations: a report should say which findings rest on strong sources and which areas are thin because sources were unavailable.
Read the source
Test yourself on 5.6 Preserve information provenance and handle uncertainty in multi-source synthesis
Ten questions, with the answer and explanation after each one.