What we publish, and what it claims.

1. Three rungs

Each item declares which it is and wears the label everywhere it appears.

Paper

A full study: method, protocol (including what was held out and why), results, limitations, and the artifacts.

Note

A finding worth publishing that doesn't warrant a paper. Usually 1,000–3,000 words.

Field note

Short, dated, specific. One thing that turned out to be true — a harness gotcha, a number that meant something other than what it looked like, a thing that broke.

There are no papers yet; the first will come out of the elicitation work, once its held-out numbers rest on enough hand-labelled pairs to quote.

2. Negative results publish, with their caveats attached

A method that didn't work is a result. A benchmark that turned out to be saturated is a result. We publish them at the rung they earned, with caveats beneath the main text when needed.

3. The method is published; the corpus never is

The elicitation work is built on one person's private transcripts. Every study says what was withheld and why. Harnesses, prompts, per-item results, and the random seed and salt values used to run and score each evaluation are released where they can be; the personal data is not, ever, and a post whose artifacts can't be released says so instead of leaving the block empty.

4. Disclosure: agents draft, a human approves

Agents may draft posts from the lab's records, including run reports, benchmark write-ups and design documents. Our publication policy requires a human to review the rendered draft, make any necessary corrections and explicitly approve publication.

5. Corrections are appended and dated

When something turns out to be wrong, the correction is added beneath the post with its date. The original text is not silently edited. The corrections block appears only when a post has corrections.

6. What we don't do

• No newsletter, no comments, no analytics that phone home.
• No invented dates, stakeholders or numbers. When a figure is a working estimate it says so.
• No publication-count vanity. If there are eight things, the index shows eight things.