Answering only when it's sure: trust through verification

What makes an assistant trustworthy is not only its good answers — it's answering confidently when it's sure, and pointing you to your expert otherwise. In the field, this discipline is precisely what lets a technician rely on the tool with their eyes closed.

What the research says

Answering where you're sure — and pointing elsewhere otherwise — has a name in machine learning: selective prediction (El-Yaniv & Wiener, 2010). The principle: a reliable system only answers the cases where it's confident enough, trading a little coverage for a lot of reliability.

Applied to language models, a landmark result shows they have some notion of what they know — calibration (Kadavath et al., 2022). A good system leans on this signal to answer when it's solid, and to hand off when it isn't.

What we do at Manent

At Manent, this discipline is a mechanism, not an intention:

• Answer when it's supported. A source in your documents backs the answer? Manent answers, with the page to back it up.
• Hand off when needed. A case falls outside what's documented? Manent points you to your expert — and the question becomes a card to validate. That moment enriches your base for next time.
• Safety, word for word. Safety instructions are quoted verbatim, as written, or escalated to a human.
• Good behaviour, measured. In our citation bench, a supported answer counts toward its reliability, and pointing to the expert is recognized as the right move — never as a failure.

References

1. El-Yaniv, R. & Wiener, Y. (2010). On the Foundations of Noise-free Selective Classification. JMLR.
2. Kadavath, S. et al. (2022). Language Models (Mostly) Know What They Know. (calibration of language models).