How the Paper Collection Is Built
This page explains where the 1074-paper collection came from, what belongs in it, and what doesn’t. It replaces an earlier split between a 96-paper hand-taught collection and a 191-paper bibliometric corpus — those were two collections built for different jobs. There is now one collection, with views (filters) on top.
The visible core
The underlying literature map covers 1,806 candidate works, with a working set of 1,488 under active analysis. Neither of those ships here. This page describes the visible core — the subset that has cleared an inclusion bar and been fully annotated:
| Period | Years | Bar |
|---|---|---|
| Historical | ≤2018 | year-cohort citation percentile ≥ 50 or k-core ≥ 3 |
| Contemporary | 2019–2024 | that bar, or references ≥3 other core papers |
| SOTA (state of the art) | 2025–2026 | 2026: references ≥3 core papers or is cited by ≥1; 2025: both |
“References” and “is cited by” only count links to other papers in this same core — not raw global citation counts. That keeps a paper’s place in the collection tied to how connected it is to the rest of the field being taught here, not just how popular it is.
This is not a second screening of the whole field and not a ranked canon of “the best” papers. It’s the set dense enough, and connected enough, to be worth teaching from today.
Every paper has one stable identity
Each record has a uuid: its DOI (lowercased) when one exists, otherwise a stable catalog work id. That uuid is what the journal-club filters, the content-library deep-dive pages, the citation graph, and “related work” links all key off of. A paper never gets a second id because it shows up in a different view.
What a card carries
Every paper in the collection gets the same shape:
- Bibliographic identity — title, authors, year, venue, DOI/uuid
- OCAR — Opportunity, Challenge, Action, Resolution, Future Work
- Three reading levels — beginner, intermediate, advanced summaries
- Tags, and discussion prompts
- Graph place — how many core papers cite it, how many it cites, its k-core (how embedded it is in the densest part of the citation network), and its citation percentile within its publication year
- Streams — pipeline stage(s), organism, dataset, method, and charting axis
- Related work — other core papers it cites or is cited by
- Links to the paper’s DOI/landing page and, where public, a PDF
Pedagogical prose (OCAR, summaries, discussion prompts) is written from the paper itself wherever a source PDF was available, which is true for the whole visible core in this drop. Where a source file couldn’t be matched, the card falls back to the catalog abstract, and annotation_status says so — that flag exists so a mismatched or abstract-only card is never presented as more thoroughly read than it was.
Views, not corpora
A view is an ordered or grouped list of uuids — it never adds or removes papers from the collection. Shipped views:
- Highest k-core — ranked by how embedded a paper is in the citation network
- Era — historical / contemporary / SOTA
- Pipeline stage — preparation through analysis
- Organism
- Dataset — named registry volumes (FlyWire, MICrONS, H01, …)
- Method
- Charting axis — including training/outreach and health-translation threads
- Year
- Suggested reading paths — editorial sequences (historical arc, methods deep dive, analysis & interpretation)
The eleven teaching dimensions used on the hand-annotated journal paper deep-dives are themselves a view over this same collection, not a second library. Browse the full collection, filtered any of these ways, on the journal club page.