Michael Ellis https://mlellis.com/blog/ Writing on causal inference, Bayesian methods, machine learning, generative AI, and the engineering required to run them in production. en-us Wed, 12 Aug 2026 00:00:00 GMT Post formatting reference https://mlellis.com/blog/formatting-reference/ https://mlellis.com/blog/formatting-reference/ Wed, 12 Aug 2026 00:00:00 GMT A scaffold post showing the Markdown, code, math, and table formatting available in posts on this site. Delete or replace it before publishing for real. meta This post exists to exercise the rendering pipeline. It is safe to delete once you have written something real — nothing else references it.

Every post is an MDX file in content/posts. The filename becomes the URL, so content/posts/formatting-reference.mdx is served at /blog/formatting-reference/.

Frontmatter

Each file opens with a YAML block. title, description, and date are required, and the build fails loudly if any is missing:

---
title: 'Post formatting reference'
description: 'One or two sentences. Used for meta tags, cards, and RSS.'
date: '2026-08-12'
tags: ['causal-inference', 'bayesian']
substackUrl: 'https://yourpub.substack.com/p/slug' # optional, added after syndicating
draft: true # optional, hides the post from builds, feeds, and sitemaps
---

Setting draft: true keeps a post visible in npm run dev while excluding it from the production build, the RSS feed, and the sitemap. That is how you work on something over several sessions without publishing it early.

Code

Fenced blocks are highlighted at build time, so no JavaScript ships to the browser for syntax colouring:

import jax.numpy as jnp
from jax import random


def bootstrap_filter(key, observations, n_particles=1_000):
    """Minimal bootstrap particle filter."""
    particles = random.normal(key, (n_particles,))
    log_weights = jnp.zeros(n_particles)

    for y in observations:
        particles = particles + random.normal(key, particles.shape)
        log_weights += -0.5 * (y - particles) ** 2

    return particles, jnp.exp(log_weights - jnp.max(log_weights))

Inline code such as getAllPosts() uses backticks.

Math

Math renders with KaTeX, also at build time. Inline math like p(θ∣y)p(\theta \mid y) sits in single dollar signs, and display math uses double:

p(θ∣y)=p(y∣θ) p(θ)∫p(y∣θ) p(θ) dθp(\theta \mid y) = \frac{p(y \mid \theta)\,p(\theta)}{\int p(y \mid \theta)\,p(\theta)\,d\theta}

Tables and quotes

Method Handles confounding Needs an instrument
Randomised trial Yes No
Difference-in-diff Under assumptions No
Two-stage least sq Yes Yes

Blockquotes are useful for pulling out a claim you intend to argue with.

Lists work as expected:

  1. Write the post here, in the repo.
  2. Deploy, so the canonical copy is indexed first.
  3. Syndicate to Substack a few days later.

That ordering is the whole point of the setup — see the README for why.

Originally published at https://mlellis.com/blog/formatting-reference/

]]>