EarthRecord agents followed one corrupted satellite sensor despite conflicting records

EarthRecord asks visual agents to reason over satellite imagery collected by several independent instruments. Its adversarial experiment then changes one instrument’s data while leaving other sensors intact. A robust analyst should notice conflicting measurements rather than treating the newest image as settled environmental truth.

Of 200 answers that had previously been correct, 107 became wrong after the single-instrument alteration. Most of those errors moved in the altered direction; only a small minority attributed the problem to the instrument. The corruption was imposed inside a controlled benchmark, not discovered in a real satellite archive. It is a concrete test of what multi-source reasoning means when source integrity fails.

Multiple data feeds do not confer independent verification when the agent fails to identify contradictory evidence.

The experiment isolates a data-integrity failure rather than an attacker instruction: an authorized agent can compute a coherent answer from falsified sensor evidence. Distinct instruments can corroborate each other only when discrepancies are actively compared and source provenance is preserved. No real land-change decision or malicious alteration of a production satellite archive was observed.