--- title: README sdk: static pinned: false --- # Semblen Semblen is a second read on the news. Give it an article, a political ad or a video, and it marks the rhetorical techniques in the wording (loaded language, name calling, appeals to fear and others), quoting the exact words each one rests on. It also checks factual claims against named sources. It doesn't take sides or rate outlets. It's live at [semblen.com](https://semblen.com), including an index of 2026 Senate and governor race ads, a daily feed of analyzed news, and a free classroom lesson. ## Models Our technique models are fine-tuned DeBERTa-v3-large multi-label classifiers: 16 technique heads, plus one that estimates whether the charged language belongs to a quoted speaker rather than the outlet. They're the first stage of the pipeline; a language model then explains each flag in plain words. The models are private. Researchers can request access: email [hello@semblen.com](mailto:hello@semblen.com). ## How we work - Every flag shows the words it rests on and a one-line reason. - We audit flags by hand. In our last blind audit, about 1 in 10 was wrong. - A flag is about how something is worded, not whether it's true or who's right. Semblen LLC · Austin, Texas · [hello@semblen.com](mailto:hello@semblen.com)