Style Examples

Everything available for formatting pages in this site.

---

Badges

Inline status indicators:

Validated Proposed Disconfirmed Triangulated Underdetermined Default

Badges also work in sidebar navigation via frontmatter (see the Verdicts section).

---

Cards

Causal Instruments

Pearl SCM, Rubin CATE, Woodward interventionism, mediation analysis, and more.

Structural Instruments

SVD spectral analysis, OV/QK decomposition, weight alignment, effective rank.

Information Instruments

Transfer entropy, PID, mutual information, NOTEARS discovery.

Behavioral Instruments

Logit diff recovery, KL divergence, generalization gap, MDL compression.

---

Link Cards

Read the Framework Start with the taxonomy overview to understand the six layers.

Quickstart Guide Audit your first circuit claim in 10 minutes.

---

Tabs

Activation Patching

Replace activations at a specific position with those from a counterfactual input. Measures necessity of the component for the behavior.

Causal Scrubbing

Recursively verify that each node in a computational graph contributes only the information attributed to it by the hypothesis.

IIA

Interchange Intervention Accuracy — swap a single representation between inputs and measure whether behavior follows the swapped variable.

---

Admonitions

Note

This is a note — useful for supplementary context that doesn’t interrupt the main flow.

Tip

This is a tip — practical advice or best practices.

Caution

This is a caution — something that could go wrong if you’re not careful.

Danger

This is a danger — a known failure mode or critical error to avoid.

---

Mermaid Diagrams

---

Code Blocks (Expressive Code)

Basic syntax highlighting:

def compute_iia(model, clean_input, patch_input, target_idx):
    """Interchange intervention accuracy."""
    clean_acts = model.run_with_cache(clean_input)
    patch_acts = model.run_with_cache(patch_input)

    patched_logits = model.run_with_hooks(
        clean_input,
        fwd_hooks=[(hook_point, lambda act, hook: patch_acts[hook.name])]
    )
    return (patched_logits.argmax(-1) == target_idx).float().mean()

With line highlighting and title:

lib/analysis/stages/stage_10_factor_decode.py

def bidirectional_logit_lens(factor_bank, tokenizer):
    W_U = model.W_U  # (d_model, d_vocab)
    factor_logits = factor_bank.factors @ W_U
    top_tokens = factor_logits.topk(10, dim=-1)
    bottom_tokens = factor_logits.topk(10, dim=-1, largest=False)
    return top_tokens, bottom_tokens

Diff format:

# Before: uncalibrated
score = compute_iia(model, clean, patch, target)
score = compute_iia(model, clean, patch, target) - baseline_iia

---

Tables

Verdict	Tier	Requirements
Proposed	A	Construct defined, one instrument chosen
Causally suggestive	B	One causal instrument with baseline separation
Mechanistically supported	C	Necessity + sufficiency + measurement calibration
Triangulated	D	Three evidence families converge
Validated	E	All five validity types pass threshold

---

Footnotes

The framework draws on Bradford Hill’s criteria for causal inference¹, adapted for the computational setting where controlled intervention is possible².

---

Asides / Custom markup

Key insight: A circuit claim is not a binary — it’s a pattern across five dimensions. The framework makes the pattern visible.

---

Math (if needed)

Math rendering requires remark-math + rehype-katex plugins. Once installed, inline math uses single dollar signs and display math uses double:

Inline: $IIA = E[1(f(x) = y)]$

Display:
$$
\text{IIA} = \mathbb{E}[\mathbf{1}(f(x_{\text{patched}}) = y_{\text{target}})]
$$

---

Images

Images can be placed in src/assets/ and referenced with standard markdown:

![Alt text](../../assets/my-diagram.png)

---

What else could we add?

Potential additional plugins/features to explore:

• starlight-links-validator — checks for broken internal links at build time
• starlight-typedoc — auto-generates API docs from TypeScript
• Custom sidebar icons — per-section icons in the nav
• View transitions — smooth page-to-page animations (Astro built-in)
• Obsidian-style wikilinks — [[page-name]] syntax that resolves to proper links
• Graph view — D3-based visualization of page connections (heavier lift)
• PDF export — compile sections into a printable document
• Version selector — if the framework evolves, show v1/v2 side by side

Footnotes

1. Hill, A.B. (1965). The environment and disease: association or causation? ↩

2. Unlike epidemiology, we can perform true interventions rather than relying on observational associations. ↩