Methodology

How we research, write, and verify every post on this site. Published so that anyone citing us knows what they are citing.

Editorial principles

Three rules govern every post:

1. Primary sources only. Every named-party claim links to the original — arXiv paper, model card, GitHub repository, SEC filing, lab announcement. We do not cite aggregator sites or tier-three blogs as authority.
2. Documentary tone. We use documentary verbs (reports, filed, states, announced, describes). We do not use accusatory verbs (misled, deceived, lied) outside verbatim quotes. We do not attribute motives.
3. Explain hard things simply. The bar is that a motivated undergraduate finishes the post able to summarize the concept in one sentence to a friend. Posts that read like paper abstracts have failed. Posts that gloss the actual mechanics have also failed.

Five-step verification

Before any post is published, it passes a five-step check:

1. Run a full fact-check sweep against every named-party claim.
2. Resolve every flagged issue. Errors block publication.
3. Manually verify the five most consequential claims by fetching the cited primary source — not the cached version, not the aggregator, the source itself.
4. Verify every internal link resolves to a live post. Verify every external link returns a 200 with the content actually cited.
5. Document which verification steps ran in the editorial log for that post.

If a step cannot be completed for any reason, the post does not ship.

How we use AI in the writing process

We use AI tools (large language models) in the research and drafting process. Every output is reviewed by a human against primary sources before publication. No claim appears on this site that has not been verified against a citation a reader can chase.

We never use generative image models. All inline figures are either:

• Diagrams from the cited paper, model card, or repository (credited inline);
• Real photographs sourced from Unsplash, Pixabay, Hugging Face, or GitHub (credited inline);
• Custom data visualizations built with Pillow / Matplotlib / Graphviz from primary-source data (source cited in the caption).

Corrections

If you find an error — a wrong number, a broken citation, a misstated claim — please tell us through the Contact page. We will fix the post, add a correction note at the bottom, and credit the reporter if they wish.

We do not silently edit out errors. Every correction is documented at the bottom of the affected post with the original wording, the corrected wording, and the date.

What we do not publish

• Opinion pieces. No hot takes, no industry "vibes" pieces.
• Speculative claims about named parties' motives. We report what was said and filed; not what we think someone meant.
• Material that carries litigation risk. If a topic requires inferring motive, alleging fraud, or claiming undisclosed wrongdoing, we do not write it.
• Posts shorter than ~3,000 words. If a topic does not warrant that length, it is not yet a post.

Updates

This methodology is a living document. Significant changes are logged here with the date of revision. Current revision: 2026-05-15.