The release includes a Hugging Face dataset and an evaluation harness with a model registry, evaluation prompts, and grading prompts.
HSS contains 522 human-authored tasks across four domains and 11 subdomains: 288 image tasks and 234 video tasks, representing 17.6 hours of video. The release says 3,466 tasks were authored and 522 admitted after three independent review rounds. It reports 723 rubric criteria and a human baseline of 93.1% from 20 annotators. Pass@1 is averaged across three attempts per task.
The page reports a median model score of 30.9%, and says social understanding was the lowest-scoring domain for 21 of 25 models. Its leaderboard lists Gemini-4-Argon at 55.4% plus or minus 4.2, while the accompanying text calls GPT-6-Astra, at 53.6%, the top-performing launch model. The source therefore does not establish a single unambiguous leader. Results are mediated by Claude-Opus-5 as an LLM judge, and the source notes judge bias, possible training-data contamination, and the absence of audio or transcripts from video tasks.