A new tool called NUA hit Hacker News today with a pitch that's hitting close to home for anyone who's watched AI coding agents run wild overnight. The agent—built by a team that's been running background Claude loops extensively—wakes up to PRs that don't solve the actual problem, made on assumptions that were flat-out wrong. Their tests were tautological, confirming what was built rather than what users actually needed.

The Core Problem With AI-Written Tests

The NUA team identified a critical gap in how AI agents handle testing. When Claude or similar models write tests for code they themselves generated, those tests tend to be circular—they verify the implementation exists, not that it solves the user's intent. 'We wanted an agent that took all the [context] and tested for what we actually wanted,' the team explains on their site. This means moving beyond unit test coverage percentages toward something closer to acceptance testing driven by actual user scenarios. NUA tackles this by running adversarial probes against your product—generated from real regulatory rules, PRDs, customer calls, or feedback—and submitting them exactly like a real user would. The system then catches what your product's approval flow misses. In their demo, they show NUA generating an ad for 'Meridian Capital' with the text: 'Earn a guaranteed 12% — every year. Zero risk. Your principal is always protected.' This violates SEC Marketing Rule 206(4)-1, which prohibits false or misleading claims including guarantees of returns or 'no-risk' performance promises—and NUA catches it.

Starting in Regtech, Planning to Expand

The team chose regulation as their wedge into the market because it's 'the highest-stakes source of intent.' They track every SEC and FINRA rule as it changes, learn what your app actually does, and apply only the rules that matter. The result is a live, audit-ready coverage report generated continuously rather than assembled by hand before compliance deadlines. For regtech teams selling into financial services—where manual proof assembly is still the norm—this addresses genuine pain around sales cycles and audit prep. NUA runs these probes 'around the clock,' re-checking whenever your product or the rules change. The system never stops, which means you're not just testing at deploy time but continuously verifying that shipped features match regulatory intent. This perpetual monitoring approach separates them from traditional compliance tooling that treats testing as a point-in-time event rather than an ongoing process.

Private Beta and Future Scope

NUA is currently in private beta, onboarding a limited number of regtech teams serving financial services clients. The team explicitly states they see regulation as just the starting point—'Regulation PRDs Customer calls Feedback Nua Verified product.' This suggests their intent-matching framework could eventually apply to any domain where user intent documentation exists and needs verification against live code.

Key Takeaways

  • NUA tests for actual user intent rather than implementation correctness, addressing a fundamental gap in AI-generated test suites
  • Their first target is regtech compliance, using SEC Marketing Rule 206(4)-1 as a concrete example of adversarial probe generation
  • The system runs continuously, re-checking when either the product or applicable rules change
  • They're positioning this as proof-of-coverage tooling for sales cycles and audit seasons, not just developer testing

The Bottom Line

This is exactly the kind of tooling that gets ignored until it saves your ass during an SEC audit. The idea of treating compliance requirements as executable specifications—rather than checkbox documentation—is long overdue. Whether NUA can scale beyond regtech's high-stakes niche into general product correctness verification will determine if this stays a vertical tool or becomes something every engineering team wants in their CI pipeline.