Chapter 12

Preserving Provenance

Claims that carry their source and date, conflicts kept rather than resolved, and format chosen by content.

D5guide part i

12.1

A claim without a source is not a finding

Summarising many sources is exactly where the claim → source link gets dropped.

Unattributed

"The AI music market is
 estimated at $3.2B."

No source, no year, no confidence. Unverifiable, and impossible to reconcile with anything.

Attributed

{
  "claim": "The AI music market is
            estimated at $3.2B.",
  "source_url": "https://…/report",
  "source_name": "Global AI Music
                  Report 2024",
  "publication_date": "2024-06-15",
  "confidence": 0.9
}

Claim, where it came from, when it was published, how sure the extractor was.

Both say the same thing. Only one can be checked, dated, or weighed against a rival number.

12.2

Two sources disagree: keep both

Picking a winner throws away the information that explains the disagreement.

Methodology and date usually are the explanation — automated classification is not a survey of 500 labels.

12.3

Dates turn a contradiction into a trend

Undated numbers from different years look like an error in the data.

Same two numbers. The date is the whole difference between "contradiction" and "+5% growth".

12.4

Render by content type

One output format for everything loses structure the content already had.

content	form	why
financial data	tables	alignment makes figures comparable at a glance
news and analysis	prose	the argument is the content
technical findings	structured lists	each item is independently actionable
time series	chronological order	sequence carries the meaning

Recall in 60 seconds

Every claim carries source name, URL, publication date and a confidence value.
Attribution is lost during multi-source summarisation unless the structure forces it through.
Conflicting values are both preserved, with methodology and date — never averaged, never picked.
conflict_detected plus a possible explanation; the coordinator arbitrates, not the extractor.
Methodology usually explains the gap — classification vs survey measures different things.
Undated figures from different periods read as contradictions instead of trends.
Render financial data as tables, analysis as prose, findings as lists, series chronologically.
Uncertainty is data: record it rather than resolving it silently.