DataPig

PassLearnersPermit turns all 51 official US driver handbooks (50 states + DC) into complete learner-permit study courses, built in eight days during this hackathon and live at https://passlearnerspermit.com. Chapter 1 is free in every state; the full state course is $9.99 for one year. The problem: the learner-permit test is the most-taken high-stakes exam in the US, every new driver, every year. The official handbooks are free but unreadable PDFs, and the paid prep market is full of unverified question banks. Every learner-facing sentence here traces through a ledger to the model attempt that wrote it, the validator receipt that accepted it, and the handbook page it cites. No human wrote a card. The founder reviewed mechanics, never content, and was the only gate. How Meta's models were used, three roles, each receipted call by call in the repo: 1. Author: muse-spark-1.3 authored the Utah course end to end through the Model API in strict JSON (16/16 chapters, 371 receipted calls, $44.09); nine more states were authored by Muse Code under its subscription through the same validators at $0 API. 2. Fact-checker: every compiled fact in all 51 states (38,074) was re-read by muse-spark-1.3 against only the page it cites: 664 flags, 444 meaning errors, 4 flat contradictions. The affected cards were hidden at point of sale and re-authored through the same belt. A planted-error control caught 3 of 3 with no false flags. Meta's model was the one trusted to check everyone's work. 3. Picture inspector and sign painter: every illustration passed a muse-spark-1.3 vision review against five drawing rules (636 verdicts, 72 rejected and re-painted); the 93 federal road signs were painted by muse-image-1.0 and judged against the handbooks' meanings. How it was built: one assembly line, "the belt". Seven strict JSON schemas per handbook chapter, a compensating check for each, an append-only ledger with a receipt per attempt. Models could try as often as the lineage cap allowed; nobody talks past a schema. Claude, GPT and Gemini lanes authored the other states, the narration, the lesson videos and the store under the same gate, and Meta fact-checked all of it. Nine numbered incidents during the build, nine closes, every one caught by a validator, a watcher or a hash before a learner saw it. What shipped: 255,145 files, 188 GB at the edge: 4,987 HD lesson videos, 234,656 narration clips, 37,967 flashcards, 36,234 quiz items, 1,490 practice forms, 723 verified illustrations, 93 road signs. Every file hash-checked after upload. Every frame of it made inside the event; the only pre-event inputs were the 51 handbook PDFs. Demo (2:58): the model lanes that ran the belt tell it in their own words over the live product and the receipts. Repo: code and course JSON for all 51 states, with the lanes' reports and the bug ledger; media streams from Cloudflare R2.