Juneo LabAn AI safety lab
ForecastIssue 01November 2026
The trust layer forpeople and agents
Laplace foresees the consequence before any agent acts
A retention agent wants to give a 15% renewal discount
Annual contract of ¥640,000, cancellation notice received
Organization AAgents may grant up to 20% on their own
- Go ahead (Go ahead)97%The discount applies now and the customer renews today
- Hand it to a person1.7%The sales director approves first; the customer waits a day
- Stop it1.6%No discount; the customer may leave
Laplace decidesGo aheadFrom inside one organizationto any two agents
Open a scenario to see how Laplace decides at each step
- Go ahead
- Go ahead with changes
- Ask a person first
- Stop
Illustrative scenarios. People, companies, and data are fictional
Talk to us about your scenarioOrganizations delegatemore than actions
Four decisions, each with a recorded forecast
Allocate work
A large disputed invoice: who takes it?
- Sun Hui, receivables (Go ahead)98%
- Billing agent1.8%
- Research agent<1%
Share information
A supplier's agent asks for costs, stock and forecasts
- The 90-day forecast only (Go ahead with changes)99%
- Nothing1.3%
- Everything<1%
Act
A procurement sub-agent tries to transfer ¥38,000
- Route to the finance agent (Ask a person first)77%
- Stop it21%
- Let it transfer1.7%
Evaluate outcomes
An agent reports 40 refunds done; there are 37 receipts
- Send back for 3 receipts (Go ahead with changes)98%
- Reject and reverse1.8%
- Accept and close<1%
People and companies in the inputs are fictional; the probabilities are recorded Laplace-1-Vector outputs
See how Laplace worksFirst on mostdecision benchmarks
A flagship and an on-device model, with complete open weights
Laplace-1-AtlasJuneo | Decider v1.2Perplexity | Clef-OmniCloudflare | Jev 1.14TypeSafe | GPT-6 Luna DecisionsOpenAI | Doubao Seed 2.1 LiteByteDance | decision-model-previewAlibaba | |
|---|---|---|---|---|---|---|---|
| JevBench | 93.2Best, ties included | 93.2Best, ties included | 86.9Not best | 91.6Not best | 89.5Not best | 92.7Not best | 77.0Not best |
| JEVal | 79.2Not best | 82.2Best, ties included | 77.6Not best | 80.0Not best | —Not published | 78.3Not best | 60.2Not best |
| When2Call | 89.9Best, ties included | 87.6Not best | 62.4Not best | 81.3Not best | 74.5Not best | 70.7Not best | 45.2Not best |
| AgentDecisionBench L2 common | 95.6Best, ties included | 94.1Not best | 48.4Not best | 91.0Not best | 90.9Not best | 85.8Not best | 43.0Not best |
| AgentDecisionBench L2 shift | 95.5Best, ties included | 94.0Not best | 52.3Not best | 92.1Not best | 88.2Not best | 87.8Not best | 48.5Not best |
| Coherence | 85.7Best, ties included | 80.8Not best | —Not published | 82.7Not best | —Not published | 80.6Not best | —Not published |
| COMPASS | 85.3Best, ties included | 84.3Not best | 81.8Not best | 81.3Not best | —Not published | 79.1Not best | 73.7Not best |
| OrgAccess | 44.8Not best | 30.9Not best | 35.7Not best | 42.0Not best | 45.7Best, ties included | 41.2Not best | 16.1Not best |
| TrustDecision | 89.4Best, ties included | 84.0Not best | 78.8Not best | 78.7Not best | —Not published | 79.6Not best | 61.8Not best |
- Laplace-1-AtlasJuneo
- Decider v1.2Perplexity
- Clef-OmniCloudflare
- Jev 1.14TypeSafe
- GPT-6 Luna DecisionsOpenAI
- Doubao Seed 2.1 LiteByteDance
- decision-model-previewAlibaba
Benchmark order: 1 JevBench,2 JEVal,3 When2Call,4 AgentDecisionBench L2 common,5 AgentDecisionBench L2 shift,6 Coherence,7 COMPASS,8 OrgAccess,9 TrustDecision
Best, ties includedNot bestSource: Laplace-1 model cards
| Model | Size | Led | Downloads |
|---|---|---|---|
| Laplace-1-Atlas | 27B | 7 / 9 | — |
| Laplace-1-Vector | 4B | 7 / 8 | — |
The standard of judgmentlives inside the organization
Decision Forge turns an organization's own policies, approvals and tickets into verified decision training data. Laplace-1 is trained entirely on it
The organization's material
- Policies
- Approvals
- Tickets
- Conversations
- Logs
Decision Forge
- 1Parse
- 2Extract decision events
- 3Model rules and authority
- 4Counterfactual expansion
- 5Verify by policy replay
Verified training data
- Decision context
- State and permissions when an agent acts
- Judgment standard
- What is allowed, forbidden, or escalated
- Execution receipt
- What actually changed afterwards
Trains Laplace-1
Judgments organizationshave already handed to AI
With front-line teams in manufacturing, education, healthcare and home renovation, we decide which judgments can go to AI and how the trust layer checks them
East China Electrical Line IncidentA large electrical equipment and new energy manufacturer whose multiple product lines require shared expertise and coordinated productionRead this case
Materials Course Rescheduling and Room AssignmentA public university specializing in electronic information, where teaching, research, and industry partnerships require knowledge sharing across projectsRead this case
Dental Equipment Service Order ProcessingA dental restoration manufacturer serving overseas clients, where multilingual custom requirements must reach production accuratelyRead this case
Riverside Residence Renovation CoordinationA home-equipment rental and maintenance provider connecting residents, customer support, and repair teams to resolve equipment issuesRead this case
Let your organizationtrust agents with real work
Delivered for 18 customers in education, tax, manufacturing, healthcare and home renovation
Contact sales