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

# Causal Workbench

> Structured causal dialogue grounded in Structural Causal Models and Pearl's causal ladder.

The Causal Workbench is Wu-Weism's primary research surface. Every conversation is grounded in a Structural Causal Model (SCM) loaded for your domain — so responses are not free-form language model output but causally constrained inference anchored to explicit assumptions. Questions you pose are automatically classified, an appropriate model is loaded, and every reply is tagged with a rung on Pearl's causal ladder.

Navigate to the Causal Workbench by selecting **Causal Chat** in the sidebar, or by opening `/chat` directly.

## Starting a session

<Steps>
  <Step title="Open the Causal Workbench">
    Click **Causal Chat** in the left sidebar. A new session starts automatically.
  </Step>

  <Step title="Type your causal question">
    Enter your question in the composer at the bottom of the screen and press **Enter** or click **Send**. Aim for a specific causal question rather than a general prompt — see the tip below.
  </Step>

  <Step title="Review domain classification">
    Wu-Weism reads your question and classifies it into a causal domain. The detected domain appears as a tag beneath the session title. This determines which SCM is loaded.
  </Step>

  <Step title="Read the response">
    Each response is annotated with one or more causal density levels (Rung 1, 2, or 3) indicating the epistemic type of each claim. Review these labels before drawing conclusions.
  </Step>
</Steps>

<Tip>
  Start with a specific causal question rather than a broad one. "Does peer-reviewed evidence support a causal link between sleep deprivation and impaired working memory consolidation?" gives Wu-Weism far more to work with than "Tell me about sleep and memory." Specific questions load tighter SCM constraints and produce more actionable causal density labels.
</Tip>

## Domain classification

When you send a message, Wu-Weism classifies your question into one of the recognized causal domains before loading a model. You do not need to specify a domain manually.

| Domain                                      | Examples                                                     |
| ------------------------------------------- | ------------------------------------------------------------ |
| `alignment`                                 | AI safety, value alignment, mesa-optimization                |
| `consciousness`                             | qualia, phenomenal binding, global workspace theory          |
| `neuroscience` / `theoretical_neuroscience` | neural circuits, plasticity, connectomics                    |
| `iml`                                       | interpretable machine learning, feature attribution          |
| `education`                                 | learning interventions, pedagogy, curriculum design          |
| `legal`                                     | tort, liability, regulatory causation                        |
| `ecology`                                   | trophic cascades, population dynamics, habitat interventions |
| `evolutionary_biology`                      | selection pressures, speciation, adaptation                  |
| `cognitive_psychology`                      | cognitive load, memory, attention                            |
| `scaling_laws`                              | emergent capabilities, compute-performance relationships     |
| `physics`                                   | quantum measurement, thermodynamic causation                 |
| `abstract`                                  | formal causal reasoning without a specific empirical domain  |

The classified domain is displayed as a tag in the session header and in the session ledger.

## Truth Cartridges and SCM loading

Once a domain is identified, Wu-Weism loads a **Truth Cartridge** — a pre-validated SCM encoding known causal constraints for that domain. The cartridge defines:

* The causal variables relevant to your domain
* Permitted directed edges between variables
* Constraints inherited from established theory and prior empirical evidence

Responses are generated within these constraints. Claims that would require edges not present in the loaded SCM are flagged or reformulated.

<Note>
  The loaded SCM is a constraint layer, not a complete domain ontology. It bounds what the model can assert causally; it does not enumerate every possible relationship.
</Note>

## Causal density levels

Every response labels each claim with one of three rungs from Pearl's causal hierarchy:

<Tabs>
  <Tab title="Rung 1 — Association">
    **What it means**: The claim describes a statistical association or correlation between variables. No causal direction is asserted.

    **When you see it**: Observational findings, descriptive summaries, correlational evidence from the literature.

    **Example label**: `[Rung 1: Association]`
  </Tab>

  <Tab title="Rung 2 — Intervention">
    **What it means**: The claim describes what would happen under an active intervention — a *do*-operator statement. This is causal, not merely correlational.

    **When you see it**: Claims derived from experimental evidence, randomized trials, or do-calculus reasoning.

    **Example label**: `[Rung 2: Intervention]`
  </Tab>

  <Tab title="Rung 3 — Counterfactual">
    **What it means**: The claim addresses what *would have* happened under a different history. This is the most epistemically demanding rung and requires a fully specified SCM.

    **When you see it**: Counterfactual reasoning, retrospective causal attribution, legal and policy analysis.

    **Example label**: `[Rung 3: Counterfactual]`
  </Tab>
</Tabs>

A single response may contain claims at multiple rungs. Review each label independently.

## Operator modes

Beyond normal dialogue, two operator modes modify how Wu-Weism processes and frames responses.

<AccordionGroup>
  <Accordion title="Intervene mode">
    Activating **Intervene** mode shifts every response to Rung 2 framing. The system applies do-calculus reasoning to your question — asking what would change under an active manipulation of the variable you specify, not merely what correlates with it.

    Use intervene mode when you are designing an experiment, evaluating a policy, or need to reason about the effect of a specific manipulation rather than an observed association.

    Select **Intervene** from the mode selector in the composer toolbar before sending your message.
  </Accordion>

  <Accordion title="Audit mode">
    **Audit** mode produces a structured breakdown of the causal assumptions embedded in your question and the SCM's response. Output includes an explicit listing of:

    * The edges activated in the SCM
    * Assumptions required to support each Rung 2 claim
    * Gaps between what was asserted and what the SCM can support

    Use audit mode when you need to interrogate the reasoning behind a response or produce an auditable record for a paper, report, or review.

    Select **Audit** from the mode selector in the composer toolbar before sending your message.
  </Accordion>
</AccordionGroup>

## Attaching PDFs

You can attach PDF files alongside your question. Wu-Weism extracts text and numeric evidence from the attachment and incorporates it into the causal reasoning for that message.

<Steps>
  <Step title="Click the attachment icon">
    In the composer toolbar, click the paperclip icon (or drag and drop a file onto the composer).
  </Step>

  <Step title="Select your PDF">
    Choose the PDF file from your local filesystem. Accepted format: `.pdf`.
  </Step>

  <Step title="Send with your question">
    Type your question referencing the document and send. Wu-Weism processes the attachment alongside the question within the same message context.
  </Step>
</Steps>

<Info>
  PDF analysis in Causal Chat is designed for single-document context. To synthesize across multiple documents, use the [Hybrid Synthesis](/workbench/hybrid-synthesis) or [PDF Synthesis](/workbench/pdf-synthesis) surfaces instead.
</Info>

## Session ledger

All past sessions are listed in the left sidebar under **Sessions**. For each session the ledger shows:

* **Title**: auto-generated from the first message of the session
* **Domain tags**: the causal domain(s) detected during the session
* **Recency**: relative timestamp (e.g., "2 hours ago", "Yesterday")

Click any session to restore it. Sessions persist across browser reloads.

## Model settings and provider selection

Wu-Weism supports three AI providers: **Anthropic**, **OpenAI**, and **Gemini**. You choose which provider and model to use in **Model Settings**.

<Steps>
  <Step title="Open Model Settings">
    Click **Model Settings** in the sidebar or the gear icon near the composer.
  </Step>

  <Step title="Select a provider">
    Choose **Anthropic**, **OpenAI**, or **Gemini** from the provider dropdown.
  </Step>

  <Step title="Select a model">
    Choose the specific model within that provider (e.g., `claude-opus-4-5`, `gpt-4o`, `gemini-2.5-pro`).
  </Step>

  <Step title="Save">
    Your selection is saved per session. You can switch providers mid-session by returning to Model Settings.
  </Step>
</Steps>

## Bring Your Own Key (BYOK)

You can supply your own API key per request rather than relying on a platform key. This is useful when you have negotiated rate limits, custom billing arrangements, or want to avoid shared-pool constraints.

Enter your API key in the **API Key** field in Model Settings before sending a message. The key is sent with each request via the `x-byok-api-key` header and is never stored server-side between sessions.

<Warning>
  Your BYOK key is held in memory for the duration of the session. Refresh the page or close the session and you will need to re-enter it. Do not share your session URL with others while a BYOK key is active.
</Warning>

## Related pages

<CardGroup cols={2}>
  <Card title="Hybrid Synthesis" icon="shuffle" href="/workbench/hybrid-synthesis">
    Reconcile conflicting claims across multiple documents and sources.
  </Card>

  <Card title="PDF Synthesis" icon="file-lines" href="/workbench/pdf-synthesis">
    Extract numeric evidence and causal claims from a single PDF.
  </Card>

  <Card title="Causal Ladder" icon="stairs" href="/concepts/causal-ladder">
    Understand Rung 1, 2, and 3 reasoning in depth.
  </Card>

  <Card title="AI Providers" icon="key" href="/guides/ai-providers">
    Configure your Anthropic, OpenAI, or Gemini API key.
  </Card>
</CardGroup>