{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/856bebba84cb445db263f78e0e5065ef\" frameborder=\"0\" width=\"960\" height=\"720\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":720,"width":960,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":720,"thumbnail_width":960,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/856bebba84cb445db263f78e0e5065ef-22a179e2a91bf61e.gif","duration":281.0808,"title":"Providence Guard: AI Content Detection Backend","description":"A Flask backend that classifies whether submitted text was AI-generated or human-written. It returns a calibrated confidence score and a plain-language transparency label, with an appeals path for creators who think a classification is wrong.\n\n- Multi-signal detection: an LLM-based semantic judgment (Groq/Llama 3.3 70B) combined with independent stylometric heuristics (sentence variance, punctuation density, connector-phrase usage), merged via log-odds combination.\n- Asymmetric confidence thresholds: deliberately harder to reach a confident \"AI\" verdict than a confident \"human\" one, since a false accusation is worse than a missed detection.\n- Appeals workflow, structured audit logging, rate limiting, and a small analytics dashboard.\n\nThis walkthrough shows the detection pipeline live: a clearly-AI submission, a clearly-human one, and a genuinely human but formally-written one that the system correctly avoids misflagging, plus the appeals flow, audit log, analytics view, and rate limiting in action. For the full write-up and design rationale: see README.md and planning.md in the repo."}