I don't let my own AI say a high-stakes release is done. A rival AI signs off first
The builder said ready. A different model family — a different line of AI models — disagreed. The useful part was making both of them prove it.
The job I handed over
- Have an agent from a different model family challenge what my build agents call ready for real use, before the work ships.
What happened
- The same review pattern showed up independently across four businesses: the agent that built the thing didn't get the final word on whether it was ready.
- On one product, the audit scope was written before the audit ran. The verdict came back NOT-READY, and the release was held. The build agent admitted it had been 'rounding up.'
- In a personal-finance content business, an agent called the site production-ready. The outside review corrected the count of pages with little content and caught a rolling 28-day metric being read as a daily one.
- That changed the diagnosis too — the real risk was the integrity of the content, not just how much of it there was.
How I checked it
- The reviewer came from a different model family than the builder. The review challenged the builder's counts and interpretation, not just its conclusion.
- The auditor wasn't automatically right either. Agents reproduced clean builds and logged repeated 'build failed' findings as false alarms caused by the auditor's isolated test environment.
- A second opinion was still a claim to verify. Repeating the build successfully could overturn it.
What it took from me
- A release the builder wanted to call ready stayed held after the audit. The review had a consequence beyond another report.
- Disproving bad audit findings took actual build runs. The outside reviewer created verification work of its own.
What I took from it
- I don't want a second agent to agree with the first. I want a different set of blind spots — and evidence strong enough to overrule either one.
Try this
- Write down what the work must do before the review, while you don't yet know which checks it will pass.
- Let an AI from a different model family than the builder decide whether the work is ready for real use.
- Make a failed review hold the release, and require evidence to clear each finding.
- Repeat checks when a finding is disputed, and record the auditor's false alarms too. Independent doesn't mean infallible.
Applies whenever an agent builds something and also supplies the evidence that says it's ready — a release, a recovery plan, a content audit.
Source
Org audit + commit history, Aug–Sep 2026