An AI that knows when it's unsure — and does something about it.

Most chatbots just keep talking. This one measures its own uncertainty while it writes. When the signal spikes it can stop to check its notes or re-read its draft; when it's confident it just wraps up — and on easy questions it stays out of the way. You watch every decision happen, and the receipt is honest about the checks it didn't need.

A 70-billion-parameter model (Llama 3.3) writes the answers — so they're genuinely useful — while LOLM's local graft reads its uncertainty per token and decides, at each segment, when to check notes, verify, or stop. The big model is the voice; the measured self-control, and an honest receipt of what it did, is the point. It doesn't prove the answer is "better" — it proves the controller acted on uncertainty this run, with the raw trace visible.

Tap one — instant replays of real runs

📄 Ground it in your own text — paste notes, a doc, an email thread (optional)

checking the live backend…

final answer

Behind the scenes — the working draft it wrote first

What you're watching

While it writes, the model's inner signals are measured on every word. A trained controller turns those signals into one of five moves:

steadyKeep writingSignals look healthy — carry on.

unsureCheck notesUncertainty is rising — go find evidence.

wobbleDouble-checkSomething shifted — re-read the draft.

stuckTry two pathsIn a rut — fork and keep the steadier one.

confidentFinishCalm and sure — write the final answer.

Want the full instruments — entropy, drift, z-scores? Open the technical demo →