Same evidence, repeated: LQET stops models from treating relays as new corroboration
On a controlled binary benchmark with supplied report lineage, LQET makes models update beliefs by evidence provenance rather than report count, cutting new-evidence update RMSE by 41-44% and log-odds drift after eight relay steps by 89-91% relative to task-only training.