Skip to content

Decision Forge

Decision Forge compiles organizational policy into verifiable decision training data. It is the open data engine for trusted delegation to AI agents; Why Decision Forge explains the problem it addresses, and the five-minute tutorial shows it end to end. The four decision surfaces are work allocation, information sharing, action selection, and outcome evaluation. Juneo Lab's open-weight Laplace-1 decision models are trained entirely on Decision Forge data.

Decision Forge pipeline: compile and review organizational knowledge, shape scenarios across four decision surfaces and distributions, verify labels through policy replay, and export training data.

Quick start

From a source checkout with Python 3.11+ and uv installed:

uv sync --group dev
uv run decision-forge validate examples/customer_support/project.yaml
uv run decision-forge build examples/customer_support/project.yaml --output runs/demo
uv run decision-forge export runs/demo --output runs/demo-decision --format decision

The fictional example builds 336 episodes. All variants of a customer case share one split. The first build assigns families to splits by quota, so realized split sizes track the configured proportions, and records them in project.splits.lock.json; later builds keep every family in its split as the project grows. The report records requested and realized counts. No model credentials are required for this workflow.

Command reference

Command Purpose
validate SPEC Validate schema, references, source paths, and evidence quotations
build SPEC --output DIR Generate episodes and publish a verified run manifest
export RUN --output DIR --format decision Export decision, typed, SFT, or canonical training views
compile FILE... --output DIR --base-url URL --model MODEL Compile documents using an explicit model endpoint
impact SPEC PREVIOUS_RUN Report changed, added, removed, and reusable scenario work
replay RUN PREDICTIONS Score recorded candidate choices against declared policies
harbor-export RUN --output DIR Produce portable single-step decision tasks
schema --output FILE Export the project JSON Schema

Read the architecture, specification, and data protocol before adapting private company workflows.