Write the note at the length the README promises

[DenisDrobyshev]

The first note ran to about a screen — 606 words. The README promises a note is readable without the original open, and a screen does not do that. What it produces is the abstract-with-links shape: a claim, a gesture at the evidence, a conclusion. A bookmark with paragraph breaks.

Now 1,787 words in English and 1,526 in Russian, and longer in specific places rather than padded.

What got added

The algorithm is stated, not assumed. δ = r + γV(s′) − V(s), what each term is, and the three properties of the error — positive on surprise, zero once predicted, moving backwards in time — that the recordings are then matched against one by one. The old note expected the reader to already have this and jumped straight to the match.

The causal evidence is in, and it was the biggest gap. The old note said the 1997 design was correlational and stopped there. True in 1997; not true since 2013. Steinberg's blocking experiment is what upgrades the claim from looks like an error signal to acts like one: activating dopamine neurons at the moment the theory says the error should be zero produced learning about a cue that blocking should have prevented. An artificial δ inserted where none existed did what a real one would have.

Each complication gets a subsection with its mechanism, instead of a paragraph apiece:

• heterogeneous populations — value-coding versus salience-like, and that they are anatomically separable, so an experiment sampling without regard to site averages the two;
• the distributional code — why an asymmetrically-weighted neuron converges on a quantile rather than a mean, which is the step that makes the result mean anything;
• striatal ramping — a measurement still in tension with the simple story, taken with a different method on a different timescale.

"What would change this note" names concrete results rather than gesturing at future work.

Sources

Two added, both checked against Crossref before use: Sutton 1988 for the algorithm, Howe 2013 for ramping — alongside Steinberg 2013.

All six DOIs resolve. Five return 200; 10.1126/science.275.5306.1593 returns 403 to a datacentre request, which is Cloudflare rather than a dead link — the metadata was confirmed through Crossref, and external-links.yml already accepts 403 for exactly this case.

Written down, not remembered

CONTRIBUTING.md gains How long a note is — the four things a note has to do to keep the README's promise, and explicitly no word count, because a target invites padding. The test given is whether a reader who has not opened the paper can follow the argument and find where it would break.

templates/note.md grows the sections that produce that shape: the background the claim needs, what the design cannot establish, one subsection per complication, and a falsification section that has to name a result.

Verified

check_notes.py passes — both notes keep the front-matter contract, the translation carries the original's date, and neither is orphaned. mkdocs build --strict passes with no warnings.