Numerai example-scripts leads with an http script piped to bash and a codex flag called yolo

The first thing the page asks for is an http script piped into bash

The repository bills itself as the official place to start playing the Numerai tournaments, then immediately points at an agent workflow rather than a notebook:

git clone [email protected]:numerai/example-scripts
cd example-scripts && curl -sL http://numer.ai/install-mcp.sh | bash
codex exec --yolo "find the best neural network architecture to predict target ender"

Two details in those three lines are worth pausing on. The installer is fetched over plain http, with no encryption on the hop that delivers a script straight into a shell, and the silent flag means nothing of the transfer is printed. The clone, by contrast, uses an SSH URL, so the two halves of the same instruction use different transports for different trust decisions.

The third line is the one that decides the blast radius. A coding agent is launched with approvals disabled and told to go looking for a better architecture on its own, which is a reasonable thing to want in a sandbox and a different thing to want on a machine with credentials in the environment.

The example prompt ends with the words target ender

The whole agent path rests on one example prompt, and it reads in full as find the best neural network architecture to predict target ender. Those last two words are the only place any target appears anywhere on the page, and nothing here says what a target is, which one it refers to, or what the column in the data is called. The notebook titles elsewhere suggest the vocabulary would have come from there, but the agent path deliberately skips the notebooks.

That matters for the audience the page is written for. Someone arriving to learn the tournament gets four notebooks with names that explain their own purpose. Someone taking the agent path is told to install a server, hand control to an autonomous loop, and describe a modelling goal in two words, with no key, no login step and no dataset description anywhere on the page. The only support route offered is a Discord invite at the top.

The notebooks are a dependency chain presented as four independent links

The second half of the page offers the manual route, and it reads as four separate tutorials when it is really one sequence. Hello Numerai comes first and is described as the place to start if you are new, exploring the dataset and building a first model. Feature Neutralization follows, on measuring feature risk and controlling it with feature neutralization. Target Ensemble comes next, on creating an ensemble trained on different targets. Model Upload closes the list, described in its own words as a barebones example.

The order is a real dependency chain: you have to have a model before you have something to neutralize, something to neutralize before you ensemble, and an ensemble before it is worth submitting. Each link points at a Colab notebook URL that resolves back to a file inside the numerai/ directory of this repository, so the page is a set of pointers into its own tree rather than a copy of anything.

The framing sentence above them is telling, though. It recommends the agent path highly, and offers the notebooks to anyone looking to kill some time on artisan data science, which is a much lower bar than the one the agent section sets.

Three of the six top-level directories are never explained

The root holds five directories and four files. The page explains one of the directories. numerai/ is where the four notebooks live, and everything documented on the page points into it. The other four entries get no mention at all: crypto/, signals/, cached-pickles/ and AGENTS.md, alongside a .cursorignore file.

The names imply a submission pipeline that the documentation stops short of. A directory called crypto and a directory called signals together suggest the step after uploading a model, and the last notebook ends at upload, described as building and uploading a model rather than as submitting predictions. So the documented path finishes one stage earlier than the tree suggests the project goes, and a reader is never told what the remaining two directories are for or whether they are needed to compete.

AGENTS.md is the other notable omission. The repository ships instructions written for coding agents at the same time as it tells you to install an MCP server for one, and it never mentions the file.

Link text and filenames disagree, and the anchors are malformed

Three of the four notebooks carry a filename that matches their title. The fourth does not: the entry titled Model Upload points at a file called example_model.ipynb, the same generic name the others would have if they followed the pattern, so the one notebook with a distinct purpose is also the one whose name says the least about it.

The markup around each entry is broken in the same way four times. Each title is wrapped in an opening anchor tag with a target and an href, and then the closing tag appears on its own line, separated from its opener by a blank line, after the description text. Whatever renderer processes this has to guess where the link ends, which is why the titles can appear to swallow their descriptions.

None of that blocks a reader, and none of it is load-bearing either. It is the sort of thing that tells you the page is maintained quickly rather than carefully, which is worth calibrating for when you are about to pipe a script into a shell.

A cache directory is tracked, and there is no changelog

cached-pickles/ sits at the root of the repository as a tracked directory rather than as something ignored. A cache of serialized Python objects is exactly the kind of artefact that is normally generated locally and excluded, and its presence in version control means either that the directory is committed by someone who did not think about it or that it is meant to ship pre-cached data for the notebooks to load faster. The page does not say which, and a cache directory is the sort of thing that goes stale without anyone noticing.

The licence file is named LICENSE.txt, the default branch is master, and the repository publishes no GitHub releases. The language reported for the repository is Jupyter Notebook rather than Python, which is a reasonable reading of a repository whose entry point is four notebooks, but it does mean that a language-based tool looking for Python scripts will rank this differently from the crypto/ and signals/ code.

The last push is dated 2026-09-28, and the repository carries 1,187 stars with 313 forks and no open issues.

Editorial conclusion

Two different products share this page and they are not equally documented. The notebooks are a real, well-sequenced tutorial that a newcomer can follow in a browser with no local setup. The agent path is three lines long and expects a lot: an SSH clone, a script fetched over plain http and executed by bash, and a coding agent run with confirmation prompts turned off. That combination is the decision. If you take the notebooks, you need nothing from this repository except the notebook files themselves. If you take the agent path, you should read the installer before running it rather than after, and you should decide deliberately about the flag that lets an agent act without asking. Two smaller things to check first: three of the six top-level directories are never explained on the page, so a repository clone gives you more than the documentation covers, and a cache directory is tracked in version control.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. numerai/example-scripts on GitHub
4. Project website
5. README