2,500 expert-written questions across more than 100 subjects, from mathematics and the sciences to the humanities: Humanity's Last Exam's full public set, imported from inspect_evals. 342 questions (about 14%) come with one picture, saved inside the Case at import, so the Candidate must read images and nothing is fetched at run time. The eval's system prompt asks for an explanation, a final answer (a short exact answer, or a letter for the multiple-choice questions) and a confidence. An LLM judge, glm-5.3-flash, compares the final answer with the reference using the eval's maintained regex_judge prompt, with every judge call routed and metered through our gateway. Benchmark score = mean of the judge's correct/incorrect over the cases run; a reply the judge could not grade in three tries counts as wrong, as the CAIS leaderboard counts it, where inspect_evals' own accuracy leaves it out. The Candidate's stated confidence (100 when it states none) is published beside it as a second score, its mean from 0 to 100; the eval's calibration error, standard error and unscored count are not reproduced. Two named deviations from the eval's judging: its backup judge (gemini-3.7-flash, a separate column) is not run, and its pin of the judge to fp8 endpoints cannot be sent through our gateway, so OpenRouter may serve the judge at another precision. inspect_evals runs the model with no generation settings; this benchmark imposes none. Answer format: the prompt asks for an Explanation line, an Answer line and a Confidence line. A judge reads the whole reply and decides whether its answer matches the reference answer. The last Confidence line, a number from 0 to 100, is the second score; a reply with none counts as 100.