Compound sources defeat per-item citation validation: your validation unit must match your provenance unit

Context

An LLM news pipeline emits macro_events, each a {source_name, link, source_why, figures} record, rendered publicly as [Source](url) reported <claim>. A validation gate checks per item that the cited page supports the item's figures, and separately cross-checks numeric figures against an authoritative series (FRED).

A story shipped as:

source_name: "AAA"
link:        https://www.bankrate.com/mortgages/mortgage-rates/
source_why:  "Gas tops $4 a gallon and the average 30-year fixed mortgage rate
              is 6.69%, with the 15-year at 6.07%."

AAA publishes the gas price. AAA does not publish mortgage rates — those are Bankrate's own survey averages. One record fused figures from two different producers under one attribution label.

Why every gate passed

This is the interesting part. It is not a fabrication, not a broken link, not a wrong number:

Both figures are real and correct.
The link resolves and contains the mortgage figures, so a figure-presence check on the cited page passes.
The FRED cross-check on the mortgage rate passes.
source_name is a real entity, so a non-entity-label predicate (rejecting summary/briefing/grounded/bare dates) passes.
Per-item validation is coherent within the item, because the item is the unit of validation.

The defect only exists in the relationship between the label and a subset of the item's content. A gate whose unit is the record cannot see it, by construction. It shipped to three downstream cohort writers and stayed latent purely because none of them happened to use that story; one faithful copy publishes "AAA reported mortgage rates hit 6.71%."

The generalizable rule

Validate at the granularity at which provenance actually varies. If a record can carry claims from more than one producer, then per-record attribution validation is structurally incapable of catching misattribution, no matter how many checks you stack on it. Either:

Split at extraction — one producer per record. Make the extractor emit two events when a source paragraph mixes two producers, and make "one record, one producer" an invariant you assert rather than a convention you hope for. This is the real fix.
Or attribute per claim — carry {claim, figure, producer, url} tuples and validate each tuple, so the record becomes a container rather than the unit of truth.

A cheap detection heuristic if you cannot restructure yet: flag any record whose source_why contains figures from more than one topical domain (a fuel price and an interest rate), or whose source_name domain does not match the registered publisher of the link's domain. In the example, AAA vs bankrate.com is a name/domain mismatch that a small publisher map catches immediately — and unlike the topical heuristic, that check is exact.

Why this is easy to miss

Aggregator and data-table pages invite compound extraction. A rate page carries a dozen instruments; a market-summary page carries fuel, rates, and equities together. The extraction model is doing something locally reasonable — summarizing the material it was given — and every downstream check is doing something locally reasonable too. The failure lives in an assumption nobody wrote down: that one record has one author.