How these fifty were chosen
A corpus you cannot audit is a corpus you cannot argue from. Here is every choice that produced this set, including the five things that went wrong and how they were caught.
Where they came from
| Lab | Source | Rows found |
|---|---|---|
| Anthropic | transformer-circuits.pub | 18 |
| Anthropic | alignment.anthropic.com | 32 |
| Anthropic | anthropic.com/research | 8 |
| Google DeepMind | deepmind.google/research/publications | 30 |
| OpenAI | openai.com/research/index (browser-driven — the domain refuses plain fetchers) | 69 |
157 rows pooled → 125 eligible → newest 50 taken.
What was excluded, so you can argue with it
Dropped: site categories Company, Security, Global Affairs, Product; any title containing system card or model card; Circuits Updates roundups; the Preparedness Framework. Two Anthropic cross-posts were deduped to their primary venue.
The exclusion rule is a judgment, and it is the load-bearing one
It is why OpenAI contributes ten papers and Anthropic twenty-two. OpenAI publishes far fewer paper-genre items on its research index; most of its output there is system cards and product posts. That skew is a fact about the labs, not an artifact of sampling — but it rests on this rule, so the rule is stated rather than buried.
Five contaminations, all caught by machine, none by reading
Each was found by a mechanical check — a hash comparison, a title probe, an affiliation probe — after a careful read had already passed them.
- Titles and URLs drifted out of alignment across all 81 OpenAI rows.
Two separately-extracted lists were zipped together and slipped a position. Fixed by extracting each card as a unit instead.
- One PDF was captured as two different papers.
Both had grabbed
preparedness-framework-v2.pdf— a site-wide footer link on every OpenAI page. Caught by a SHA-256 duplicate check across all fifty texts. - A third-party mathematics paper was sitting in the corpus.
OpenAI’s “Ten advances” post is a roundup; the fetcher followed it to one of the ten — a paper by four non-OpenAI authors. Caught by a title-match probe.
- An ACL Anthology landing page had been captured instead of the paper.
Eighteen kilobytes of site navigation. Caught by an affiliation probe that asked whether each text names its own lab.
- A supplementary appendix stood in for a paper.
A rewritten chain-of-thought document rather than the study. Same probe caught it; repointed to the actual chemistry paper.
Verification that passes now
50 files · 0 duplicate texts · every text contains its own lab’s name · every text matches its own title probe · 3,600,682 bytes total.
What is still imperfect
- “Empirical research paper” is a judgment call, applied by the exclusion rule above rather than by reading each item. A roundup post and a benchmark release both survive it.
- Enumeration is only as complete as each lab’s own index page. Anthropic’s research page under-returned; the gap was covered from two other Anthropic surfaces, but no source was verified exhaustive.
- Sixteen entries carry month-only dates, sorted at the 15th as a midpoint. Any claim about exact ordering within a month is soft.
- Six papers are abstract-only. Style claims resting on those rest on abstracts.
- Twelve OpenAI rows were dropped during transcription rather than at the filter step. All are dated April 2025 or earlier and the window opens 23 April 2026, so none could have entered the fifty — but the raw file is not byte-complete against the source page, and that is said here rather than left to be discovered.