> ## Documentation Index
> Fetch the complete documentation index at: https://docs.merionlabs.com/llms.txt
> Use this file to discover all available pages before exploring further.

# What is an Intelligence Factory?

> Understand the durable learning boundary that connects real work, evaluations, experiments, and release decisions.

# What is an Intelligence Factory?

An Intelligence Factory is a persistent learning boundary for one economically valuable workflow. It contains the policy, sources, work episodes, evaluation tasks, experiments, and decisions needed to improve an AI system scientifically.

It is not synonymous with a code repository. A factory may start in:

* an empty folder for a workflow that has not been automated yet;
* an existing operational directory containing documents or artifacts;
* a Git repository for a software workflow;
* a workspace that references several local sources.

## Why start from work?

A model score only matters when it represents a capability that users actually need. Real work provides the missing evidence:

* the original request and business trigger;
* the state from which the worker or agent began;
* the tools, permissions, harness, model, and constraints;
* corrections, retries, and human interventions;
* the accepted or rejected outcome;
* the recurrence and economic value of the task.

A trace is still only raw evidence. Merion does not automatically treat every trace as an evaluation. Episodes must be reviewed, grouped with related examples, made reproducible, equipped with credible verifiers, and audited before they can become release evidence.

## The durable product object

Merion ultimately connects an audited evaluation task to a versioned `ExperimentSpec`. An experiment pins the task, environment, harness, tools, verifier stack, intervention, inference configuration, repetitions, budget, and decision constraints.

That lets a team answer a concrete question:

> What should we change next, and what evidence shows that the change improved real work without creating an unacceptable cost, latency, safety, reliability, or human-intervention tradeoff?

## Harness-neutral by design

Merion can provide a first-party guided workflow, but it remains above the harness layer. Teams should be able to compare models, prompts, tools, skills, inference configurations, and agent harnesses against the same governed task distribution.