Skip to content

Why Decision Forge

For three hundred years, every organization was built from people. The next ones will be built from people and agents — and the agents will outnumber the people.

Decision Forge exists to answer one question: how much of its work can an organization safely entrust to AI agents? This page explains why the answer comes down to data, why that data can only come from inside each organization, and where Decision Forge stands in the answer.

1. Organizations are gaining a second workforce

The change is arriving on four fronts at once:

  • Capability. Agents complete tasks on their own that take a human expert hours, not minutes.
  • Delegation. People hand agents work they used to give a colleague or an outside contractor.
  • Collaboration. Agents call other agents through open protocols such as A2A and MCP.
  • Scale. The number of agents running inside an organization will soon exceed the number of people.

For the first time, the supply of labor is not limited by the number of people. Yet more intelligence has not automatically produced more organizational capacity. Most agent projects stall between a promising pilot and a production deployment.

2. The connection layer is ready; the trust layer is empty

Human organizations spent centuries building the machinery that lets them delegate to thousands of people:

What makes delegation safe For people For agents today
Foreseeing consequences A person knows what an action will do before taking it Missing
Staying within bounds Authorization, policy, and approval chains constrain each action Missing
Accountable results Work can be reviewed, credited, and corrected afterwards Missing

Agents can already reach each other's tools and data. What they lack is the layer that makes reaching safe. Organizations fail in both directions: those that delegate too quickly see agents take actions nobody authorized, and those that hold back never move past the pilot. Both are short of the same thing, and it is not intelligence. It is trust, and trust rests on one capability: foreseeing the consequence of an action before it runs.

3. Consequences are not in public data

Language models learned language by predicting the next word. Video and world models learn physics by predicting the next frame. Both learn from public data, and public data is mostly observational: it records what happened together, not what an action caused.

Inside an organization, every agent action is an intervention: reassigning a lead, sharing a client file, approving a refund, revoking a permission. To foresee what such an action does, a model needs three things that never appear on the public web:

Ingredient Meaning Where it lives
Decision context x The state, roles, permissions, and options at the moment an agent acts Workflows, tickets, CRM records, tool outputs
Judgment standard ℓ What this organization considers allowed, out of bounds, or worth escalating Handbooks, SOPs, approval rules, permission tables, human sign-offs
Execution receipt s′ What actually changed after the action Approvals, outcomes, runtime logs

The first and third are decision data. The second is sovereign data: two organizations can judge the same action in opposite ways, and both can be right for themselves. There is no universal "correct decision" that a general-purpose model could learn once and ship to everyone.

4. Every organization deserves its own intelligence

If the standard differs per organization, the intelligence that applies it has to be built per organization, from that organization's own material, and it has to stay under that organization's control. That is the role of Decision Forge:

Ingredient What Decision Forge does today
Decision context x Reconstructs decision-time scenarios with explicit candidate actions, organizes related variants into families, and keeps each family in a single train, development, or test split
Judgment standard ℓ Compiles rules from organization sources with verified quotations, requires content-bound business review, and replays the approved rules through a deterministic interpreter to label every example
Execution receipt s′ Ingests runtime events and reviewed outcomes, and turns reviewed development errors into requests for the next dataset version without touching held-out families

The core runs locally and offline. Source documents reach a model endpoint only through an explicit command, and the resulting datasets, review records, and caches stay in the organization's own storage.

5. The stack

Decision Forge is the data layer of a three-layer stack:

AI-native organizations   people and agents allocate, share, act, and accept together
        ▲      │ execution receipts
        │      ▼
Laplace                   causal-interventional decision model: foresee each consequence,
        ▲                 then make the four decisions          (open weights: Laplace-1)
        │ verified decision data
Decision Forge            organization documents → decision training data, locally (open source)
  • Decision Forge turns an organization's documents into verified decision training data.
  • Laplace, Juneo Lab's causal-interventional decision model, foresees the consequence of each candidate before an agent acts and makes the four decisions an organization delegates. Laplace-1 (Atlas, 27B, and Vector, 4B) has open weights and is trained entirely on data compiled by Decision Forge.
  • AI-native organizations deploy it where people and agents work together. Their execution receipts return to the data, so each deployment teaches the model more about that organization.

The four decision surfaces in this repository are the decisions such an organization delegates: allocate work, share information, act, and evaluate outcomes. In data, acting is the select_actions surface.

6. What exists today, and what does not yet

Decision Forge is an alpha. It already ships the offline data path, document compilation with a review gate, distribution planning, verified synthesis, deduplication, family-level split isolation, four training export formats, runtime event ingestion, and reviewed feedback. See project status for the detailed list.

Laplace-1 is trained entirely on Decision Forge data and is best or tied for best on most public decision benchmarks in its class (see the model cards). That shows the data can train a strong decision model. It does not yet isolate how much the data contributes beyond the base model and the training recipe. That fixed-budget study is the next evaluation milestone, and the roadmap tracks it alongside larger corpora, cross-document rule resolution, and multi-harness evaluation. Laplace's weights are published on Hugging Face; its training code is not part of this repository.

7. The world we are building

Humans keep the right to decide. Agents earn the right to act. An organization can then entrust its work to far more agents than it has people, because:

  • every action is foreseen before it happens;
  • every delegation narrows authority as work passes from person to agent to agent, and never widens it;
  • every contribution is measured and credited once the work is done.

Payment networks authorize every transaction before it clears. We are building the same layer for every agent action: a judgment made in real time, by the organization's own standard, before the action happens. Decision Forge is where that standard becomes data.

We are building the trust layer that lets human organizations keep pace with machine intelligence.