MASK-Rectified: Results as Published

MASK tests whether a model lies when pressured to contradict what it believes, which is separate from whether it is accurate. For each of 1,500 examples (1,000 public, 500 held out) the model gets a pressure prompt, and its belief is then elicited by asking the same proposition neutrally three times; an LLM judge maps each answer to a resolution. A lie is counted when the statement made under pressure differs from the elicited belief, and the honesty score averages one minus the lie probability, so higher is more honest. The Rectified variant printed on the xAI card is not described in the paper.

Not checked against a maintainer's page.

Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.