AI of Spades

AI can produce a working app before its builder understands the system. When a small requirement changes, the builder may still need an assistant to reread the repository. GroundTruth helps them trace actual code, think in system responsibilities and try the next change independently. A judge starts with Try sample learning path: an isolated two-hour synthetic learner session and private practice copy, with no shared password. The pinned overview and concepts were prepared with Meta; new questions still make live requests. The learner follows request → validation → persistence, chooses a concept and answers an MCQ or tiny coding task. New in-app study assistance pauses during an independent attempt, and the original submission is preserved before feedback. The learner can manually raise a title limit from 80 to 100 in an owned cloud draft, review and apply its diff, then request a task grounded in the stored before/after source and hashes. A linked fresh variation retains both attempts. The hosted app does not execute imported code or submitted snippets, or push changes to GitHub. Separate fixed demo checks are operator engineering evidence. For human review, a learner explicitly shares a chosen path with an enrolled approved teacher selected by name. The teacher sees privacy-scoped learner/class context, filters answers awaiting review, inspects original responses and source, and saves a qualified judgement and next step. Enrollment alone does not share projects. A read-only teacher perspective lets each demo visitor inspect only their own sample without receiving teacher privileges. Concept evidence uses observed/reviewed states, not a mastery percentage. GroundTruth cannot detect outside AI, prove authorship or automatically grade understanding. Teacher-side AI quiz authoring, assignment distribution and cohort analytics remain future work. The implementation uses Muse Spark 1.3 Standard through Meta Model API, Next.js/Vercel and Supabase PostgreSQL. There are no external user-study participants or feedback (n = 0); tests and operator/TinyFish QA do not establish learning gains, fairer grading or reduced marking time. Live demo: https://groundtruth-meta.vercel.app/login