# Sadalsuud — a confidence detector

Canonical page: https://keilab.io/research/sadalsuud/

research · sadalsuud

A confidence detector, not a truth oracle: Sadalsuud is designed to sort a language model's answers into high, medium or low confidence and send low confidence to retrieval or an honest refusal.

## What it does

Sadalsuud reads how confident a language model is before its answer is used. High-confidence answers pass; low-confidence ones are sent to retrieval or turned into an honest refusal instead of a guess. It detects the model's own confidence; it does not judge whether an answer is true.

## Where it fits

Sadalsuud is the confidence gate of the Genesis line. It is built for the lab's own MAX runtime, next to the language models it watches.

cross-references

- [→ Genesis (parent)](https://keilab.io/research/genesis/)
- [→ K2K (sibling)](https://keilab.io/research/k2k/)
- [→ KeiBeta](https://keilab.io/research/keibeta/)
- [@ Request access](https://keilab.io/investors/)

Status: In development

Output: High / medium / low confidence

Low confidence: Retrieval or an honest refusal

Sensitivity: concept public · methods NDA

[@ Request theory access](https://keilab.io/investors/)

---
KeiLab is operated by KeiSei LLC, a Wyoming limited liability company. +1 307 210 8538
Site index for AI agents: https://keilab.io/llms.txt