[Superseded by #46/#159] Evaluate verification-routing policies, not just task-routing policies

[dgenio]

Context

The referenced research-agent discussion suggests a routing problem beyond model/tool selection:

When should a system route an output through extra verification, independent review, adversarial probing, or human escalation?

This fits agent-routing-eval-lab because verification itself is a routing policy with cost, latency, and trust tradeoffs.

Problem

Routing evaluation often focuses on whether a task was sent to the right model/tool. But in agent-assisted research and AI-generated code, a major decision is whether the produced artifact has cleared enough evidence to be trusted.

Bad verification routing has two failure modes:

• False negative: accept output that needed deeper verification.
• False positive: over-verify cheap/low-risk output and waste cost/latency.

Proposal

Extend the lab with a verification-routing evaluation mode.

Policy decisions could include:

• accept_without_extra_review
• route_to_independent_reviewer
• route_to_adversarial_probe
• route_to_baseline_comparator
• route_to_human
• reject_or_rework

Evaluation dimensions

• correctness / defect catch rate
• false accept rate for high-risk artifacts
• unnecessary verification rate
• cost and latency per accepted artifact
• shared-blind-spot risk
• expected value of additional verification
• calibrated confidence before vs after verification

Dataset sketch

Create synthetic logged decisions with:

• artifact risk level
• generator confidence
• historical defect outcome
• verification actions taken
• reviewer disagreement
• cost/latency of each verification route
• final accept/reject decision

Deliverables

• New synthetic dataset generator for verification-routing logs.
• Baseline policies: always accept, always verify, confidence threshold, risk-aware threshold, expected-value policy.
• Metrics for false accept, false reject, trust/cost frontier, and p95 latency.
• Example notebook/CLI comparing policies.
• README section connecting verification routing to agent-assisted research workflows.

Acceptance criteria

• Running the lab produces a comparison table of verification policies.
• At least one policy shows lower false accepts at comparable cost to a naive threshold.
• The output makes visible that agreement between agents is not equivalent to independent verification.