This starts a distinct analytical record because no prior New Runtime record is supplied; the available material supports evaluating this particular verification design without assuming a related prior item.
Thesis
Repeated agent attempts become more useful when verification tests both disagreement and shared assumptions instead of treating consensus as proof.
Observed facts
The project page says VeriHarness uses the same model for candidate rollouts and verification, with separate investigation contexts for disputed and unanimous claims.
1The project reports a Gemini 3.5 Flash average native benchmark score increase from 47.2 to 53.4 across five benchmarks after evidence-backed revision.
1
Mechanism
Two independent investigations use workspace evidence to resolve competing claims and challenge agreed claims; a fresh adjudication context combines their evidence records to select and revise an artifact.
Why now
The project page presents a released harness and evaluation results for long-horizon work across documents, spreadsheets, code, and multi-file tasks, making verification architecture a concrete agent-design question rather than a purely conceptual one.
Implications
Agent quality can depend on evidence gathering and structured verification, not only on the base model or number of attempts.
A verification record can make revisions more inspectable, while the extra investigations impose tool-call and latency costs.
Evidence boundary
The reported results concern five benchmarks, ten rollouts per task, and two frontier models; they do not establish performance across all agents, tasks, or deployment conditions. The page notes that verification adds tool calls and latency.
Watch conditions
Whether independent evaluations reproduce the reported gains on new models and task suites.
Whether the added verification cost is acceptable for professional workflows.