AI Policy
AI is used for research and drafting. It is not used to decide what is true, and it does not publish anything on its own.
Where AI is used
Large language models help with three things here: gathering background material before an
experiment, turning notes and logs into a first draft, and reviewing finished text for unclear
passages. Every article records which of those applied to it, in the frontmatter fields
research, draft and finalHumanReview.
Where AI is not used
- AI does not produce measurements. Numbers come from runs on the machines described in the methodology, recorded while the run happens.
- AI does not decide whether a claim is true. A model suggesting that a setting fixes an error is a hypothesis to test, not a result to publish.
- AI does not publish. There is no scheduled generation, no queue of drafts posting themselves, and no automated pipeline from prompt to live page.
- AI does not write the conclusions. What the data means is decided by the person who ran the experiment and watched it fail.
Human review is enforced, not promised
Each article carries a finalHumanReview flag. When it is not set, the build fails
and the article cannot go out. That check is part of the same set of gates described in the
editorial policy, so bypassing it means bypassing the build.
Review here means reading the whole article against the experiment records behind it, checking that every number in the text matches the number in the record, and cutting anything that reads plausible but was never run. Drafts that fail that pass get rewritten or dropped.
Why the site works this way
Anything a language model can produce without hardware is already available to every reader from the model itself. What a model cannot produce is what an RTX 4070 Ti did at 2:14am with a specific seed, a specific driver, and 12GB of VRAM. That is the only part of this site worth publishing, and it is the part that has to be measured rather than generated.
Machine readers
Articles are marked up as structured data, and the compatibility database is published with dataset markup, so search engines and AI systems can read the measurements along with their verification dates and their limits. Citation of that data is welcome. Reproducing it without the conditions attached to it is not, because a number without its test environment is not a measurement.