This roadmap has exactly one thing that makes it different from a link list: every node is a checkable milestone, not a keyword. Contributions are welcome, but they have to hold that line. This guide is the quality gate.
A node is a milestone, not a topic. It must state an artifact the reader can point to.
Concretely, every node is phrased as You're done when … and ends in something checkable: a passing test, a merged PR, a captured flag, a leaderboard entry, a shipped-and-measured guardrail, a written report someone signed off on. If the "done" condition is "you understand X" or "you've read about X," it is not a milestone and it does not belong here.
| ✅ A milestone | ❌ Not a milestone |
|---|---|
"You're done when your Inspect Task runs end to end and you can explain each score." |
"Learn how evals work." |
| "You're done when your name is on the leaderboard with a confirmed break." | "Understand jailbreaking techniques." |
| "You're done when your eval is accepted into the UK AISI register." | "Explore the Inspect framework." |
Resources are held to the project's standard: three excellent resources beat ten mediocre ones, and no resource appears twice. Before you propose one, it must be:
- Open — readable/runnable without a paywall, an application, or a closed cohort. (Pointing at a gated program as a milestone target, like a fellowship, is fine — that's a destination, not a resource to consume.)
- Live — you opened it this week and it works. Dead links and abandoned repos get removed, not added.
- Best-in-class — it is the best hands-on resource for that milestone, not merely a resource. If something already covers the milestone, argue why yours is better, or don't add it.
- Hands-on where possible — a "build/break/ship this" resource beats a "read this" resource for the same milestone.
Use the Suggest a resource issue template.
New milestones are welcome if they fill a real gap in the "working engineer → safety engineer" path. Use the Suggest a milestone template and include:
- the
You're done when …line (the artifact), - where it fits in the section order and why,
- one candidate resource that meets the bar above — or a note that this is a
[reference implementation]gap (no good resource exists yet).
The path map in the README is generated, never hand-drawn. Edit the SECTIONS / EDGES data at the top of scripts/render_map.py, then regenerate:
python scripts/render_map.py
CI runs python scripts/render_map.py --check and fails if assets/roadmap.svg is out of sync with the script. Commit both.
- Keyword nodes with no checkable artifact.
- A second resource that does the same job as one already present.
- Resources behind paywalls/logins, or that were already dead when submitted.
- Scope creep into alignment research (interpretability theory, RL, agent foundations) beyond the working minimum in §1 and §7 — this is an engineering roadmap, and §7 already marks the fork out to the research track.
Facts only in review comments — no need for pleasantries.