CLAIM-2026-10-12-01

Cadence, a runtime monitor that varies its inspection schedule, resolves 42.33% of 300 SWE-bench Lite tasks on mini-swe-agent against 17.00% for the same agent with no monitor.

Why it matters

The comparator is the same agent unmonitored, so the figure isolates the monitor rather than the model. It is the number to quote when arguing for an inspection budget.

Sources

arxiv2026cadence

Note

This file was overwritten on 2026-10-12 by a second edition published the same day, which restarted claim numbering at 01. Statement and rationale are the originals, recovered from the site's ingested copy. The source list was reconstructed from the edition's single cited paper. See D51.

CLAIM-2026-10-12-01

Cadence, a runtime monitor that varies its inspection schedule, resolves 42.33% of 300 SWE-bench Lite tasks on mini-swe-agent against 17.00% for the same agent with no monitor.

Why it matters

The comparator is the same agent unmonitored, so the figure isolates the monitor rather than the model. It is the number to quote when arguing for an inspection budget.

Sources

arxiv2026cadence