The questions. The 25 questions on this site are the top-scoring tasks from an internal run of a research-grade benchmark spanning medicine, law, finance, technology, academia, product research, and UX. The Hyperspace answers shown are exactly what the Hyperspace research engine produced for that run — unedited, links and imperfections included.
The rival answers. Claude Fable 5, GPT-5.5 Pro, Grok 4.3, Fugu Ultra, and GLM-5.2 were each given the same question text and asked for their single best, self-contained answer, with web search available. Every answer panel carries a modeline naming the exact model that produced it.
The verdicts. For each question, each judging rival was shown all 6 answers, labeled by producer — including its own — and asked to name the best answer and say frankly where its own stands, judging correctness, depth, grounding, and responsiveness. A sixth judge, Mistral Large 3, sits as an independent arbiter: it produced no answer, so it has nothing of its own in the lineup. That is one verdict per judge per question: 6 judges × 25 questions = 150 verdicts cast. Hyperspace does not judge; it is the subject.
No self-votes. A judge cannot score its own answer: when a judge names its own answer best, that verdict is disclosed in the breakdown table as a parenthesized (self) figure but is excluded from the tally and the animation. 16 of the 150 verdicts cast were self-votes; 134 count.
The tally. The chart counts the “Best answer:” line of every verdict, recomputed from the eval files on every build. The quotes in the sidebars are verbatim excerpts from those verdicts — each one machine-checked against its source. Full, unedited self-evaluations are at the bottom of every question page.
