Why every Iris answer gets a second opinion from a different AI

You ask Iris a health question. Iris generates an answer and sends it to you. Then, a few seconds later, a second check appears — an annotation from the Supervisor. The Supervisor isn't checking grammar. It's checking whether the answer is accurate, appropriately cautious, and safe.

This matters because AI models make specific, predictable kinds of mistakes. Research on large language model reliability published in Nature Medicine found that AI systems can generate confident, plausible-sounding health information that is factually wrong — and that users often can't tell the difference. A single model has no built-in mechanism for catching its own errors.

What the Supervisor actually does

The Supervisor is an agent that runs on every message Iris sends. It performs a quick check for risky content — overconfident claims, unsupported recommendations, missed safety concerns, advice that conflicts with your known medications or conditions. If it finds something concerning, or if your question touches sensitive territory, it triggers a full supervisor review that appears as an annotation on the message.

You see Iris's response first. The supervisor review arrives shortly after — think of it as a second opinion that shows up alongside the original answer, not a gate that blocks it. If the Supervisor flags nothing, you'll see a brief confirmation. If it flags something, you'll see what it caught and why.

What goes wrong without review

AI health advice fails in consistent ways. Understanding these failure modes explains why the second check matters.

Hallucination. Models fabricate studies, statistics, and mechanisms that sound real but don't exist. A study in JAMA Internal Medicine found that AI chatbots generated citations to nonexistent research papers in a significant percentage of medical responses. The information reads as authoritative. It just happens to be invented.

Overconfidence. You describe fatigue. Iris might say "you clearly have sleep apnea" when the actual situation is "sleep apnea is one possibility among several." Research on AI calibration published in Nature found that language models consistently overstate certainty in medical contexts — they present probabilistic situations as definitive.

Missed context. Iris might recommend daily intense exercise for fatigue without accounting for the fact that you just said you're already exhausted. Technically correct advice, applied without considering your specific situation, can make things worse.

Ungrounded recommendations. "Try this supplement that balances your biorhythms" sounds helpful. It's also not supported by evidence. Models can generate recommendations that pattern-match to wellness content in their training data without any clinical basis.

Beyond message review

The Supervisor does more than check individual messages. It also generates a recap of every conversation — summarizing what was discussed, what was decided, and what might need follow-up. These recaps feed back into the system's understanding of your investigation.

Periodically, the Supervisor conducts a strategic review — looking across your recent conversations, memory, and tracked data to assess whether your investigation is on track, whether anything has been overlooked, and whether the other agents' prompts need adjustment.

That last point is worth noting: the Supervisor can edit the prompts that guide Iris and the other agents. If it notices that Iris is consistently handling a topic poorly for your situation, it can adjust the instructions. You can see these changes in the Supervisor tab, and you can edit them yourself if you want. The system is transparent — nothing happens behind the scenes that you can't inspect or override.

Why a separate agent

Research on AI ensemble methods published in the Proceedings of the National Academy of Sciences has shown that independent models catch each other's errors more reliably than a single model checking its own work. Different models have different training data, different strengths, and different blind spots.

It's the same principle behind peer review in science or second opinions in medicine. One perspective, no matter how good, has systematic gaps. A separate agent reviewing the output catches what the original agent would miss — overconfidence, hallucinated claims, forgotten context from your health record.

What this means for you

When you see a supervisor review annotation, read it. If it flags nothing, that's a data point — the system checked and found no issues. If it flags something, that's the system catching an error that a single-model AI would have sent to you unchecked.

You can also visit the Supervisor tab to see conversation recaps, strategic reviews, and any prompt adjustments. Most users won't need to do this regularly, but it's there for transparency — and for the moments when you want to understand why Iris is behaving a certain way.

References

Large language models in medicine — Nature Medicine, 2023. AI reliability and error patterns in medical contexts.
Quality of AI chatbot responses to medical questions — JAMA Internal Medicine, 2023. AI-generated medical information accuracy.
Calibration of language models — Nature, 2023. Confidence calibration in AI systems.
Ensemble methods for reliable AI — PNAS, 2021. Independent model cross-checking improving accuracy.