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

# Forge overview

> Robot datasets, augmented and verified: the Dataset Workbench and the Playground.

<div className="forge-kicker">◆ FORGE · forge.alakazam.gg</div>

The forge takes a robot dataset you already have and gives you back a bigger,
verified one. It runs the same world-model engines that power Alakazam's
interactive worlds, pointed at your episodes instead of ours.

There are two ways in, and they work on the same campaigns:

<CardGroup cols={2}>
  <Card title="Dataset Workbench" icon="factory" href="/forge-quickstart">
    The pipeline. Upload or open a dataset, check its health, sample sources,
    plan transforms, run gated renders, and deliver an augmented dataset you can
    train on. Six legible steps, every one inspectable.
  </Card>

  <Card title="Playground" icon="gamepad-2" href="/forge-playground">
    One episode at a time. Scrub a real recorded episode fullscreen, describe a
    change in words, or splice a precise counterfactual, surgical augmentation
    with a verification verdict on every version.
  </Card>
</CardGroup>

## The spine

A campaign walks six steps. You can stop, inspect, and re-enter anywhere; the
pipeline never advances on your behalf.

| Step          | What happens                                    | What you see                                |
| ------------- | ----------------------------------------------- | ------------------------------------------- |
| **Overview**  | The dataset's classes are detected and profiled | class counts, the selected target class     |
| **Source**    | Per-episode eligibility for the target class    | eligible / excluded, with the exact reason  |
| **Sample**    | You pick the source episodes                    | the selection, saved on the campaign        |
| **Transform** | You plan operations                             | readiness checklist per operation           |
| **Run**       | Candidates render and are machine-verified      | gate scores per episode, accept / reject    |
| **Deliver**   | The kept set exports (this is the billed step)  | the archive, plus a certificate per episode |

## The engines

Two model families, two different jobs:

* **New trajectories** (workbench runs, and both playground verbs) use the
  campaign's **action-conditioned render engine**: a fine-tune of the Cosmos
  world-model line that actually obeys spliced or edited action streams. The
  base model doesn't; the fine-tune is why the forge can change *what the robot
  does* and keep the video honest.
* **Object swap** and **environment restyle** (workbench transforms) use
  general **video-to-video transfer** models, change *what the scene looks
  like*, keep the demonstrated motion. Their backends connect per deployment;
  the Transform step shows exact readiness.

## Nothing is fabricated

Every rendered episode carries a machine verdict from an action-consistency
verifier (does the video's implied motion match the trajectory that was asked
for?) plus gate scores you can read. You can confirm or overrule any verdict, overrules are recorded as yours. Described changes in the playground become
real batches with real terminal states; when a backend isn't live, the item
says so instead of pretending.

## Accounts, auth, billing

One Alakazam account spans play and forge. The API authenticates with your
session's bearer token on every call: see the [quickstart](/forge-quickstart).
Billing is **pay for what you take**: uploading, checking, running, reviewing
are free; **delivery bills 1 credit per kept episode**, exactly once, ever, details in [billing](/forge-billing).