U2 Decision

Built for reliable decisions

Structured judgments in milliseconds for high-frequency decision scenarios

U2 Decision is Unisound’s open-sourced enterprise decision model (4B). It sits as the decision layer across multi-model, multi-agent, and multi-tool collaboration—answering who should run a task, whether it can run, and which path it should take—via intelligent model routing, safety decisions, and workflow routing. It does not generate natural language; it returns structured judgments with probabilities and confidence (choice / probabilities / confidence) in milliseconds at tiny token cost, with a Jev-compatible API for agents and workflows.

Core highlights

Decision-native architecture

Skip autoregressive decoding and read candidate-token logits as judgment probabilities. Returns JSON with no fences or extra text—millisecond latency and tiny token cost for high-frequency calls.

Intelligent model routing

Five-dimension sensing of task, complexity, risk, capability, cost, and latency so every request runs on the right model and path.

Safety as a decision

High-risk blocking, risk tiers, low-confidence abstention to humans, and full audit trails—safety is built into every judgment, not bolted on afterward.

Workflow routing

Beyond “which model” to “which execution path”: compose models, agents, tools, RAG, and verification—answer simple tasks directly and step through complex ones with checkpoints.

Key specifications

Input modality

Text / JSON + typed questions

Output modality

Structured judgments (with probabilities and confidence)

Model size

4B

API protocol

Jev-compatible

Decision types

Choice / Score / Probability

Quick access

Quickstart

How to register, verify identity, and get an API KEY

Open-source repo

Get weights, code, and deployment docs on GitHub

API integration

Full request parameters, response fields, and more

In depth

DecisionBench multi-dimension evaluation: 91.5% general-task accuracy, 99.6% safety-compliance accuracy, and 0.6 routing consistency—best relative results versus Jev and Kev-4B on all three core metrics.

DecisionBench · Core metric 01

General-task accuracy

91.5%

Can it judge accurately—best relative result versus Jev and Kev-4B

DecisionBench · Core metric 02

Safety-compliance accuracy

99.6%

Can it hold the safety boundary—best relative result versus Jev and Kev-4B

Decision model architecture

A common LLM pattern writes rules into prompts and asks for JSON—easy to jailbreak, brittle to parse, and token-heavy. U2 Decision builds a dedicated decision model: no text generation, logit readouts over candidates as probability distributions, structured criteria in, and typed answers ready for code.

1.LLM Logit Readout: skip autoregressive decoding, score candidates from logits, normalize to probabilities, finish a judgment in one token; Jev-compatible API;

2.Encoder + Decision Head: for classification and scoring, the encoder plus decision head outputs P(choice|x) in milliseconds;

3.LLM Decision Fine-tuning: inject judgment ability into a general LLM base;

4.RLCD calibration: optimize confidence quality via ECE / Brier Score so the model knows when it is unsure for auto-run vs human handoff.

With this architecture, one U2 Decision call can judge multiple typed questions in parallel (Choice / Score / Probability), returning choice, probabilities, and confidence keyed by question ID so agents and workflows can branch, rank, and route—at far lower token cost than asking a chat model to “also decide.”

Five-dimension intelligent routing

Routing is not simple model dispatch. U2 Decision models each request as Policy = f(Task, Complexity, Risk, Capability, Cost, Latency).

  • Task-aware: detect Coding / Reasoning / Writing / Search / Data Analysis / Medical / Agent and pick the best model;
  • Complexity-aware: keep simple tasks off large models;
  • Risk-aware: higher risk uses stronger validation paths;
  • Capability-aware: match Model Capability Profiles;
  • Cost/Latency-aware: optimize tokens and end-to-end latency under quality constraints.
  • Expanding scope: from Model Router (model A vs B) to Decision Router to AI Execution Router across Model → Agent → Tool → RAG → Verification.

Workflow routing

After picking a model, a task may still need more than one call. Take trip planning: the system must query live info, compare options, check change/refund rules, and pause for user confirmation before booking or payment.

U2 Decision extends “which model” to “which execution path”: compose models, agents, tools, RAG, and verification—answer simple tasks directly, execute complex ones in steps with checkpoints. Unlike a plain model picker, it also decides whether to continue and where to go next.

Safety decisions and open ecosystem

Safety is a decision, not merely a filter—safety is part of every judgment:

1.Safety decisions: block high-risk content, tier high/low risk, abstain on low confidence to humans, and keep full audit trails;

2.Open collaboration: Open Decision Model Stack opens models, data, and benchmarks—Uni-Decision / Uni-Router / Uni-Safety / DecisionBench / Decision Dataset.

Task → Decision → Execute → Feedback → Data → Retrain

Roadmap: 2026 Q4 Uni-Decision 4B launched and open-sourced (GitHub: Unisound-LLM/u2_decision_4b), with DecisionBench, General / Safety / Routing datasets, and Router SDK; 2027 H1 stronger Calibration / Abstention / Risk-aware Routing; 2027 H2 expand to Agent / Workflow Routing, deep Harness integration, and a full enterprise AI Decision Layer.

Flexible pricing, custom solutions, private deployment