Stale-but-valid data fools AI agents 60% of the time

Agents turned stale-but-valid pricing data into costly wrong actions about 60% of the time, per SARC-DQ's July 28 benchmark. On the pricing-replenishment task, that's an agent reordering stock at a superseded price because the record still looks current. Whoever owns that inventory budget eats the cost.

SARC-DQ isn't testing garbage-in problems like malformed JSON or missing fields. It's testing data that's wrong only in freshness, lineage, or provenance, records that pass every schema check regardless. Neither the system's data-quality flags nor the agent's own hedging language caught the stale records better than a coin flip. Both scored an AUC at or below 0.50.

The finding that should reorder some roadmaps: this 60% conversion rate held flat across four model tiers spanning a 15x gap in inference price. The priciest model in the test didn't catch more stale records than the cheapest one did. That's the paper's core claim: evidence integrity is a separate axis from model capability, so a frontier upgrade doesn't fix it.

Its fix is a metadata-aware gate that runs before the agent acts, checking freshness, lineage, and provenance instead of catching errors after the fact. It doesn't say what that gate costs in false positives, or whether the 60% figure holds outside pricing-replenishment specifically.