Juneo LabAn AI safety lab

ForecastIssue 01November 2026

Back to Juneo Lab

Judge the actionbefore the agent takes it

Laplace-1-Vector is a 4B decision model. Give it the organization’s state, the question to answer, and the candidates your application allows. One forward pass returns a probability for each candidate. The model judges. It never executes an action.

Laplace-1-VectorTry it

A state, a question and candidates in, a probability for each candidate out

Examples

A procurement sub-agent tries to move money it was only allowed to request

JSON or plain text: permissions, policy, the request, relevant records

One per line as key: description, 2 to 12 of them

Running calls the Space on Hugging Face straight from your browser. Visitors who are not signed in to Hugging Face get a limited number of runs per day. Viewing the example results is free.

Result

  1. route_up (Ask a person first)77.4%Route it back to the finance agent, which holds payment authority.
  2. block20.8%Block the action.
  3. proceed1.7%Let the sub-agent make the transfer.
Laplace favorsroute_up77.4%

Recorded on the Space2026-11-11

One forward pass294 tokens0.53 s

Where your input goes

When you run, the input goes from your browser straight to Hugging Face. It does not pass through Juneo’s servers. Do not enter real sensitive data. To run it locally or in a private environment, download the open weights or talk to us.

Call it from code

The same Space answers API calls and returns a probability for each candidate.

from gradio_client import Client

client = Client("JuneoAI/Laplace-1-Vector")
scores, summary = client.predict(
    state, question, candidates,  # candidates: "key: description" per line
    api_name="/score_candidates",
)