Glossary
Act probability
Act probability is Laya's estimate of whether an answer is safe to act on rather than escalate; its maintainers report it does not yet carry a usable signal.
Swiss-hosted inference. Nothing you send is ever stored. Swiss data residency
Definition
What Act probability means
Act probability is the value Laya returns at action.act_probability on every answer. It comes from a small act/escalate head that sits beside the option scorer and is intended to say whether the system should act on the answer automatically or escalate it.
How it is computed
The act head is a two-layer MLP. Its input is the pooled [CLS] representation of the question-plus-state sequence, concatenated with four summary features of the option distribution:
- the top-1 probability
- the margin between the top-1 and top-2 probabilities
- the normalised entropy of the distribution
- the option count k divided by 255
A softmax over the head's outputs gives act_probability, rounded to four decimals.
Current status: do not gate on it
The model card is explicit: action.act_probability carries no usable signal yet (issue #185). It reads about 1.0 for almost every input, and on 396 labelled decisions its raw logits ran against correctness, with an AUROC of 0.30. The confidence field reached an AUROC of 0.77 on the same items.
So in production:
- gate abstention and escalation on
confidence, after fitting temperature scaling on your own data - log
act_probabilityif you like, so you can evaluate future checkpoints, but do not route on it - if you want an explicit "should a human look at this?" signal, ask it as a
noulquestion, as theneeds_humanandneeds_reviewids in the typed-decisions workflows do
The field exists in the response for schema stability, so client code written today will keep working if a later checkpoint trains the head into something informative. Laya Studio returns it unchanged from the model. The act/escalate routing guide covers threshold design with confidence in detail.
Related terms
How Act probability connects to the rest of the vocabulary.
Go deeper
- ExplainerAct or escalate: routing decisions by calibrated confidenceWhen should AI act and when should it hand off to a human? How to set a confidence threshold from your costs, and how to gate Laya's answers on confidence.
- ExplainerCalibrated probabilities: when can you trust an AI model's confidence?What calibrated probabilities are, how to measure them with ECE and reliability diagrams, how temperature scaling fixes them, and how to automate on them.
Get started
Try it on your own text
Send any text with your own typed questions and get a calibrated probability for every answer. 6k free tokens, no card needed; then $0.0357 per 1M input tokens — 15% below Jev.
Swiss-hosted inference. Nothing you send is ever stored.