Who tests
the tests?

A green check can reward a security bug. Compare two test suites, catch the misleading pass, and inspect the task a funder could help us develop.

Behavioral browser simulation, not EVM execution or an audit. No wallets, model calls, client code or real funds.

1. Choose an implementation

2. Choose a verifier Weak: owner withdrawal only Specification-based: owner + stranger + overdraw

Test the verifier across all three versions

The intended version should pass. Both defective versions should fail. Each check starts with 100 fictional units.

Implementation	Weak suite	Specification-based suite

Executable Solidity cases. Start with 3–5 original or licensed tasks; obtain independent expert review and mutation checks before expanding to 10–15.
Reproducible evaluation. Record model/version, tool access, cost, attempts and expert baselines. Keep unseen code families separate.
Bounded training. With a model partner, test whether training changes held-out results. Report negative results and distinguish prompt improvements from training.

This demo shows verifier design, not measured model improvement. Budget, schedule, reviewers, license and model access must be agreed with a sponsor.