Learn · Context management & reliability · CCAR-F task 5.6

Provenance and uncertainty

In multi-source synthesis, keep every claim tied to its source, surface conflicts, and let the model say "I don't know".

Key ideas

Keep claim-to-source links all the way through: subagents return findings with URL, document and page, and the final report cites them.
Surface conflicts instead of silently picking one value; explain the likely reason (different year, scope or method).
Explicitly allow "I don't know" or "not found"; it reduces hallucination.
For long documents, extract quotes first, answer only from them, and retract any claim with no supporting quote.
Use the API's citations feature for document grounding.
Treat a confidence label as something you state and check, not proof; it still needs sources behind it.

One finding with its source, quote, page, a confidence label and a recorded conflict. The URLs are placeholders; the structure is the point.

A synthesis step that drops source links, so final claims can't be traced back.
Averaging or silently choosing between conflicting figures instead of reporting both with context.
Forbidding "I don't know", which pushes the model to fill gaps with invented facts.
Treating a stated confidence label as evidence on its own.