Method and sources
A guide about verifying claims should be explicit about how it verifies its own.
What counts as a source
Vendor documentation is the primary source for how a specific model behaves, because it is updated with the product. Peer-reviewed and preprint research is the source for general claims about how language models behave. Official legal texts are the source for regulation, never a summary of one.
Vendor blog posts and benchmark announcements are treated as marketing until the underlying documentation or an independent result supports them.
How we use forums and social media
Reddit, Hacker News, Stack Overflow and community forums are excellent for discovering what people actually struggle with, which techniques are spreading, and which failures are widespread rather than anecdotal. They are not evidence that a technique works.
So community signal sets our agenda and never our conclusions. When a widely repeated trick turns out to be folklore, saying so plainly is more useful than repeating it.
What we will not do
- Present a prompt as guaranteed to work, since results depend on model, version and task
- Quote benchmark numbers as proof of real-world performance
- Give legal advice. Regulation pages explain what the rules say, with links to the actual text
- Recommend a tool because of an affiliate arrangement