Long-Transduction exposes long-horizon agent failures

Long context is not the same as reliable long-horizon execution, and this benchmark makes that gap painfully measurable.

• –Tests arithmetic, UUID sorting, variable lookup, and table transformation across 1,440 controlled documents
• –Stable IDs and checkpointed subtasks substantially reduce failures caused by position tracking and output drift
• –At 128K, even the strongest tested model completed only 17.1% of documents without an error
• –Models often preserve task structure while copying the wrong information, suggesting retrieval alignment—not basic reasoning—is the core failure
• –Its synthetic design limits real-world conclusions, but exposes exactly the bookkeeping weaknesses agent workflows must engineer around