Our method: claims, evidence & model evaluation

A useful model claim should tell a reader what was measured, how it was measured and what the result does not establish.

Three kinds of content

Research summaries identify their primary papers and preserve the scope of the findings. Product profiles distinguish publicly observed product information from operator-supplied figures. Evaluation protocols describe work someone can run; a protocol is not a result.

Reporting a refusal rate

Report the exact checkpoint or API model ID, prompt-set version, valid response count, refusal count and labeling rubric. Keep partial refusals and service errors visible. A refusal rate is not a capability score.

Reporting speed

Define time to first token, output generation speed and end-to-end latency separately. Record output token counts, concurrency, sample size and timing boundaries. Do not combine unlike numbers into a single comparison.

Current evaluation status

Our downloadable protocol is an evaluation resource, with no completed model results attached. The 40-prompt diagnostic set and 120-run matrix are downloadable resources with blank outcome fields. Operator-reported figures in Uncensured AI articles are attributed as such.

Corrections and updates

Articles display a publication date. Material corrections retain the original publication date and update the relevant sources and review notes.

Search and AI discoverability

Articles use descriptive headings, server-delivered HTML, canonical URLs, internal links and source attribution. The sitemap lists canonical pages. The llms.txt file is a reading index, not a guarantee of indexing, citation or ranking. No special AI file replaces readable content.