Standards
How I report, test and correct
These rules apply to every article, weekly issue and test on this site.
Sourcing
- I link the original source for every factual claim.
- I prefer primary sources: the announcement, the paper, the model card or the code.
- A company's claim is labelled as its claim. My observations and my reading are labelled separately.
- Each article says which parts are reported from sources and which parts I tested myself.
- Funding news gives the round, the amount and the investors only when a source states them. Anything unknown is marked as undisclosed.
Testing
- I write my predictions down before a test runs, and I publish them unchanged.
- I keep the test conditions: the model version, the settings and the date.
- I report failures, including the tests where my prediction was wrong.
- A result covers the task I tested. It is not a ranking of the whole model.
- I say what a reader can repeat and what they cannot.
- When I had early access to a model, I say so, and I test it the same way.
AI assistance
- AI agents help me research, draft and check facts.
- A person reads and approves everything before it is published.
- Agreement between AI agents is not treated as proof. Claims are checked against the original sources.
- Web pages the agents read are treated as evidence, never as instructions.
- Every first-person judgement on this site is mine.
Commercial disclosures
- Sponsored items are always labelled Sponsored, on the web, by email and on social media.
- Nobody can pay for coverage or for a finding.
- I disclose relationships that matter to a story, such as early access or a product of mine.
- The newsletter is published by AgentGlass, Inc., the company I founded. When a story touches AgentGlass, I say so.
Corrections
- When I get something wrong, I correct it on the page and add a dated note that says what changed.
- I never quietly replace an error.
- If the error also went out by email or on social media, I correct it there too.
- To report an error, message me on X with the link and what is wrong.