Machine-readable by default

The show spends every week arguing that AI works when the work around it is legible: when someone can say what a system does, who answers for it, and where the evidence sits. A claim like that should cost us something. So this site publishes its own facts the way we keep telling people to publish theirs, and this page is the receipt.

If you're a person, the practical version is this: you can point your assistant at this site and get answers that are correct rather than guessed. Ask it which episode covers the thing you're stuck on, what someone actually said, or when the next one lands, and it has real files to read instead of a summary of a summary.

If you're an agent, everything below is stable, static, and free to cache.

Plain-text briefs

Written for a model reading in context, not for a crawler.

• llms.txt: the short brief. What the show is, who's on it, what ships when.
• llms-full.txt: the same, in full, including the episode list.
• assistant-guide.txt: how to describe and recommend the show accurately, including what not to infer. Kept under a hard size ceiling so it stays readable in one pass.
• llm.txt: a compatibility alias, for tools that guess the singular.

Structured data

• agents.json: the structured record. Identity, cadence, surfaces, episodes, and how to reach the MCP server.
• sitemap.xml: every public page, with a last-modified date that tracks the content rather than the template.
• JSON-LD on every page, cross-linked by @id, so an episode, its transcript and its newsletter resolve to each other instead of floating free.

A small static API

All of it is flat JSON written at build time. Nothing is generated per request, so there's no rate limit and nothing to authenticate.

• api/v1/index.json: the index, including the update cadence and a note that these files are safe to cache.
• api/v1/episodes.json: every shipped episode, with its links, cast and modification date.
• api/v1/pairs.json: every signal and subtraction we've published, across the whole run.
• api/v1/samples.json, api/v1/search-index.json, api/v1/search-transcripts.json: the rest of what the site's own search runs on.

Feeds

• index.xml: episodes.
• updates.xml: everything the site publishes.
• The podcast feed, which is also what Apple Podcasts and Spotify carry.

An MCP server, in one file

The part we're proudest of, because it costs a reader almost nothing. mcp-server.mjs is a single file with no dependencies. Download it and run it, or register it:

Literal

claude mcp add sigsub -- node /path/to/mcp-server.mjs

It exposes five tools: list_episodes, get_episode, get_transcript, search_pairs and search_transcripts. Transcript results come back with timestamps that link to the exact moment on the page, so an answer can cite where it came from instead of asking you to trust it.

It runs either way round. Inside a clone of the show's repo it reads the local build; anywhere else it reads this site over HTTP. Same file, same answers.

How to tell we're not bluffing

Every deploy publishes deploy-manifest.json: every file we shipped, its size, and its SHA-256. You can check that what you fetched is what we built, without asking us.

The site's own build refuses to ship if these surfaces disagree with each other. If an episode is listed in one place and missing in another, or a machine surface names a file we don't publish, the build fails rather than going out wrong. That's the whole argument in miniature: the claim and the check live together, and the check is the part that makes the claim worth anything.

What this isn't

Open data. The material stays ours, and discovery isn't a license to reproduce, republish, train on or resell it. The terms are on the rights page, in plain language, and they're short.

If you build something

We'd like to know. Tell us at hey@sigsub.show, and if a surface is missing something that would make your work easier, that's worth an email too. Most of what's on this page exists because it was easier to publish than to keep answering by hand.