Open reading notes on machine learning, reinforcement learning, neuroscience and the methods used to check a claim.
A paper says what it found. A note here says what it showed, which is a smaller thing, and marks the distance between them. That distance is the whole point of the repository.
Everything is open and reusable under CC BY 4.0: take a note, quote it, build a lecture on it, argue with it — attribution is the only condition.
docs/notes/ |
One paper, read closely. What it claims, what the evidence supports, what a replication would need. |
docs/essays/ |
One question across several papers, where a single note cannot hold the argument. |
docs/meta/ |
How claims are checked: baselines, effect sizes, seeds, replication, publication bias. |
Topics run across reinforcement learning, deep learning, dopamine and reward prediction, neuroscience adjacent to learning theory, and meta-research. The list grows by what gets read, not by a plan fixed in advance.
It is not a paper list and not a survey. A link with a one-line summary is a bookmark, and bookmarks are already free. Each note is written to be readable without the original open, and to be wrong in a findable way if it is wrong.
It is not peer-reviewed, and every note carries a status field saying so:
draft means read once, reviewed means read again after a gap with the
sources checked line by line.
A note without a resolvable source is an opinion with formatting. So the front matter is a contract, and CI enforces it:
---
title: Dopamine and the reward prediction error
date: 2026-08-10
status: draft
topics: [neuroscience, reinforcement-learning]
sources:
- title: A neural substrate of prediction and reward
url: https://doi.org/10.1126/science.275.5306.1593
---scripts/check_notes.py rejects a missing field, a malformed source URL, a
status outside the allowed set, a topic that is not lowercase-kebab, a date in
the future, and a note no reader can reach from its section index. It runs on
every push, because a rule nobody checks is a preference.
Reachability of a source is a separate job: external-links.yml resolves every
link weekly and opens an issue when one dies, because a DOI that stopped
resolving turns a note back into an assertion.
GitHub reads CITATION.cff, so the Cite this repository
button on the sidebar produces BibTeX and APA for the collection. To cite one
note, use its own page URL on the site — every note has a stable address and a
date in its front matter.
A correction is worth more than an addition. If a note misreads a paper, open an
issue with the passage that contradicts it; that is the fastest way to make this
better. CONTRIBUTING.md has the format for a new note and
the checks CI will run.
Prose and figures: CC BY 4.0. Code in notebooks and scripts: MIT. Quoted material from a cited paper stays under its own licence and is used as quotation, not redistribution.