YZY-stack/DF40

DF40 is a defensible research artifact with meaningful adoption (331 stars, active community, NeurIPS 2024 acceptance) that establishes a new benchmark standard in deepfake detection. Key defensibility factors: (1) dataset gravity—once widely adopted as a benchmark, switching costs are high; (2) comprehensive coverage of 40 techniques including recent SoTAs creates a moving target that requires continuous maintenance and updates; (3) infrastructure-grade artifact that becomes embedded in research workflows and pipeline evaluation; (4) domain authority—NeurIPS acceptance and institutional backing (YZY-stack) signal legitimacy. However, frontier_risk is medium because: OpenAI/Anthropic/Google could build competing deepfake detection frameworks as part of broader safety/moderation efforts, and the dataset itself is a collection/curation task rather than a novel algorithmic breakthrough. The novelty is novel_combination (assembling 40 techniques into unified benchmark with evaluation framework) rather than breakthrough—the techniques are existing, but the comprehensive multi-method benchmark is new. Velocity is 0.0/hr (mature, post-release phase typical for datasets), which is consistent with a published benchmark. The 23 forks suggest moderate developer adoption but not viral ecosystem effects. As a reference implementation and evaluation framework, it functions as infrastructure for the deepfake detection community but lacks the network effects of a platform or the irreplaceability of a foundational algorithm. Score reflects strong traction and research legitimacy (7+) tempered by replaceability and frontier lab risk (not quite 8-9 tier).