Evidence and data confidence: every number opens

The evidence drawer

Any score, median or AI-written sentence opens a drawer containing the offers and completed sales behind it: source name, retrieval timestamp, match tier and confidence, condition grade, and the raw record as captured.

Nothing is summarized away. If the underlying sample is three sales from one source, the drawer makes that obvious in about a second.

How data confidence is computed

Data confidence is a single composite between 0 and 1, displayed on every variant and folded into both role scores.

Identity confidence

The match tier of the underlying records. Tier 1 identifiers score highest.

Freshness

Time since retrieval, decayed continuously and bucketed into fresh, aging and stale for display.

Sample depth

Number of usable comps after outlier trimming, relative to the category's typical density.

Source rank

Whether the record came from a registered source with stable schema and stated refresh policy.

Rules the system will not break

Missing data is labelled missing. Estimated values are labelled estimated and never rendered in the same typographic weight as observed values. Stale records are visibly stale rather than quietly reused. An AI-written summary that cannot cite a record does not get written.

The analyst is constrained, not creative

The research assistant receives only the retrieved evidence rows for the variant in question and is required to attach the record it used to each numeric claim. Its output is persisted alongside the evidence set, so a report you read next month can still be checked against the data that produced it.