Information Provenance
Attribution dies during summarisation unless you preserve it. Learn claim-source mappings, conflict handling, temporal awareness, and format-by-content-type.
Information provenance is what makes a research system trustworthy: every claim can be traced to where it came from. The core risk is that attribution dies during summarisation — unless mappings are explicitly preserved through each stage of a multi-agent pipeline, the citations dissolve and you are left with confident-sounding but unsourced assertions.
Structured claim-source mappings
Each finding must carry five components so the claim stays traceable end to end.
- Claim — the specific assertion being made.
- Source URL — where the information was found.
- Document name — the source title.
- Relevant excerpt — the supporting passage.
- Publication date — when it was published or collected.
{
"claim": "Global renewable investment rose 12% in 2023.",
"sourceUrl": "https://example.org/energy-report-2023",
"documentName": "Global Energy Outlook 2023",
"excerpt": "Investment in renewables grew 12% year over year...",
"publicationDate": "2023-11-15"
}Resolving conflicting sources
When two credible sources report different numbers, do not silently pick one — that destroys information and presents false certainty. Instead, annotate both values with full attribution, include publication dates, and present the difference as a potential trend for the reader to interpret based on methodology or timing.
Silently pick one number (say, the newest) and present it as fact. This discards real information and manufactures false certainty.
value = newest_source.value # the other figure vanishesAnnotate both values with full attribution and dates, and frame the difference as a possible trend for the reader to interpret.
[ {"value": "8%", "date": "2022"},
{"value": "12%", "date": "2023"} ] # likely accelerationChoosing the most recent source when two credible sources conflict looks decisive but discards real information. Report both with dates and let the consumer judge.
Temporal awareness
Publication dates turn apparent contradictions into meaningful trends. Sources showing 8% growth and 12% growth from different periods are likely signalling acceleration, not conflicting data. Without dates, a reasonable trend reads as a contradiction.
Content-appropriate rendering
| Information type | Best format | Why |
|---|---|---|
| Financial data | Tables | Enables direct comparison |
| News / events | Prose | Preserves narrative context |
| Technical findings | Structured lists | Clarifies hierarchy |
Attribution across pipeline stages
Attribution must survive every stage: research subagents → analysis subagents → synthesis → report generation. Synthesis is the highest-risk step, because that is where an agent is most tempted to paraphrase and drop the mappings. Instruct synthesis agents to carry the claim-source mapping forward, never to summarise it away.
The five-part claim-source mapping (claim, URL, document, excerpt, date) must pass intact through synthesis. That stage is where attribution most often dies — guard it explicitly.
How the exam will try to trick you
The distractors below look right under time pressure — learn the tell.
- The trap
When two credible sources conflict, take the most recent value and present it as fact.
Correct answerAnnotate both values with source attribution and publication dates, and let the consumer decide.
Why: Arbitrarily selecting one value destroys information and manufactures false certainty.
- The trap
Treat different numbers from different sources as a contradiction to resolve.
Correct answerInclude publication dates so the difference reads as a trend (e.g. acceleration), not a conflict.
Why: Different collection periods usually explain the gap — without dates a real trend looks like bad data.
- The trap
Let the synthesis agent paraphrase findings freely to produce a clean report.
Correct answerRequire synthesis to carry the five-part claim-source mapping forward intact.
Why: Attribution dies during summarisation — synthesis is where citations most often dissolve.
- The trap
Render every finding in one uniform format (all prose, or all tables).
Correct answerMatch format to content: tables for financial data, prose for news, structured lists for technical findings.
Why: Flattening everything to a single format degrades readability and comprehension.
Key takeaways
- Attach a five-part mapping — claim, source URL, document name, excerpt, publication date — to every finding.
- When credible sources conflict, annotate both with attribution and dates instead of silently choosing one.
- Publication dates convert apparent contradictions into trends like acceleration — always include them.
- Render by content type: tables for financials, prose for news, structured lists for technical findings.
- Attribution must survive research → analysis → synthesis → report; synthesis is the highest failure risk.
- Never let synthesis agents paraphrase away the claim-source mappings.
Frequently asked questions
How do you handle two credible sources that report different numbers?+
Do not arbitrarily select one value — that destroys information and manufactures false certainty. Present both figures with their full attribution and publication dates, and frame the difference as a potential trend rather than a contradiction. When the sources are from different periods, the dates often reveal that the numbers reflect acceleration or change over time, letting the reader interpret them on methodological or temporal grounds.
At which pipeline stage is source attribution most likely to be lost?+
Synthesis. Attribution has to survive research subagents, analysis subagents, synthesis, and report generation, and synthesis is the highest-risk step because that is where the agent condenses many findings and is most tempted to paraphrase without carrying the claim-source mappings forward. Instruct synthesis agents explicitly to preserve the five-part mapping so provenance is not summarised away.