Introduce a source and unique content aggregator from previous versions and runs.

[mpstaton]

[mheilimo]

An aggregator is most useful when it preserves provenance instead of flattening every run into one source pool.

I would model three outputs:

1. source_snapshot: canonical URL, retrieval date, exact locator, content hash, publisher, and source tier.
2. claim_record: stable claim ID, memo section, source IDs, support state (supported, contradicted, or missing), confidence, and review owner.
3. run_delta: new, changed, and disappeared claims; changed source content; unresolved contradictions; and any section or score movement with its evidence reason.

Deduplicate only identical content hashes. A URL can change between runs, and missing evidence must remain different from negative evidence. That makes the aggregator useful for both memo refreshes and human review.

I work on DDScore at Playful Pixels Oy. The same version-delta problem appears when a private company is re-scored over time: the useful artifact is the evidence change behind the 0–100 first-pass report, not only the new score. DDScore's 12-dimension investor view is one implementation reference. It supports analyst judgement and does not replace full due diligence or provide investment advice.