External verification improved task success and safety at the same time
September 25, 2026 by Asif Waliuddin
CORRECTION (2026-09-30, prunable after 2026-12-29): this entry previously stated the gains as "6.5 and 27.0 percentage points above the prior state of the art". The paper reports them as "surpassing the previous SOTA by 6.5% and 27.0%", and those are relative gains over the prior best (67.5% task success and 68.9% collision avoidance), not percentage points. Evidence: arXiv 2609.20822, abstract and Table 1.
SafeHarness attacks the Verification Debt thesis from the constructive side: with a checker placed outside the language model's planning loop, it reported 71.9% task success and 87.5% collision avoidance, surpassing the previous state of the art by 6.5% and 27.0% respectively (relative gains).
In a bounded robotics benchmark, the researchers placed a verifier outside the language model's planning loop. Proposed routes were checked against obstacle geometry before execution; rejected plans were replanned. Task success reached 71.9% and collision avoidance 87.5%. The paper reports these as surpassing the previous best by 6.5% and 27.0%, and as 2.3 times and 1.5 times the same agent run without the harness.
That is exactly the kind of evidence that should change a working thesis. Verification Debt does not imply that every additional unit of output requires another human reviewer. Independent machine-verifiable constraints can retire debt before an action crosses the trust boundary.
The result also sharpens what "independent" should mean. A second model saying "looks good" is different from checking the proposed action against an external state representation with a deterministic rejection condition. It is the constructive answer to OverclaimBench: if an agent's own report of what it verified cannot be trusted, the fix is an independent check with a deterministic reject condition, not another self-assessment.
What remains unproven: this is a narrow task suite with approximate obstacle geometry. It does not show that verification scales faster than generation in general, or that total verification economics are favorable in production.
Source: SafeHarness
What I would test next: compare prompt-only constraints with an external allowlist or deterministic action-boundary verifier in an isolated software-agent workflow, measuring task completion, prevented violations, latency, and total cost.
Test your own AI output against the evidence
Faultline checks AI-generated claims against live sources. Paste a claim, then inspect the grounded verdict and the gaps in its evidence. It verifies claims; it does not verify what a tool or agent actually did on your systems.
Self-serve: no installation, meeting or sales call. Subscribe online only if you need more scans.