Provenance & Citations

Explainability tells a person why the model decided. Provenance tells them where it got the evidence — the exact source behind every claim, one click from the number to the document that produced it. When the stakes are real, people don't act on a summary; they act on something they can open and check.

Put the source next to the claim

The instinct is to keep the interface clean: show the answer, tuck the evidence behind a "details" link. But a source one click away is a source nobody opens. Put the primary citation inline, beside the number it justifies, so the evidence and the claim are read together.

This is the difference between a system that asks for faith and one that offers collateral. A risk score that reads "72% — based on these three filings" is one a person can begin to trust; "72%" alone is a verdict they'll quietly override.

Let them open the original, unedited

A citation that can't be opened is decoration. The trail has to end in the actual artifact the model read — the filing, the ticket, the transcript — not a paraphrase of it. The moment a person can pull the source and see it themselves, the model stops being a black box and becomes a research assistant.

Heavy summarisation is the silent killer here. Compress the evidence too far and the trail goes cold; the person can't tell whether the model read the document carefully or skimmed it. Keep the path from claim to primary source short and intact.

Count the sources, and surface the dissent

Say how many sources back a number. "Three sources" reads very differently from "one source," and the person calibrates accordingly. A single-source claim should look more tentative than a well-corroborated one — the interface should carry that distinction.

Then surface disagreement honestly. If two sources support a conclusion and one contradicts it, show the outlier rather than averaging it away. The dissenting source is often where the real risk lives, and hiding it is how a confident-looking answer becomes an expensive one.

Implementation Checklist

• Render the primary source inline, next to the claim — not behind a "details" affordance.
• Make every citation openable: link to the real artifact, not a summary of it.
• Show the source count, and let a single-source claim look more tentative than a corroborated one.
• Surface the dissenting source instead of averaging it into the consensus.
• Watch a real user: can they get from the number to the document in one move?

See This Pattern In Action

• Technical Due Diligence: every score carried a clean drill from summary to the source document, so partners would stand on it.
• AI-Assisted Due Diligence: three sources sat beside every risk score, not buried behind it.

One pattern a month. The tradeoffs I paid for, plus code.