atlas.
A learning platform that will not pass you until you have earned it. Strict grading, real mastery gates, and a tutor that only teaches what it was actually taught.
01 · THE PROBLEM
Most self-teaching tools reward motion. You watch, you click next, you feel productive, and nothing checks whether the idea actually stuck. I wanted the opposite: a loop that will not let me move on until I can prove I learned the thing, and a tutor that will not bluff an answer it was never taught.
02 · HOW IT WORKS
atlas teaches one subject, how large language models actually work, as a ten-topic spine that runs from tokens and embeddings up to retrieval. Each topic is a lesson, a quiz, and a mastery gate, and the topics open in order. Multiple-choice is graded by plain code; your written answer goes to a strict AI grader that scores it against a rubric. The model never decides whether you passed. Score 80 percent or better and deterministic code opens the next topic and schedules the material for review. Miss it and the topic stays shut.
03 · ENGINEERING DECISIONS
The grader sees the question, the rubric, and your answer, and nothing about who you are, your streak, or your history. It is a pure function with exactly that one shape, so it cannot soften to be nice because it never learns whose work it is. Its scores are clamped in code as a second line of defense, so a misbehaving model can never over-credit.
atlas grades on a cheap model, not a frontier one, and the risk in that trade is a cheap model that over-credits a vague but plausible answer. So a golden set of twelve labeled answers, spanning clearly correct, partial, vague traps, wrong, off-topic, and empty, runs against the real model, and the eval fails if it gives credit to anything it should have denied. That eval is the gate for ever swapping the grading model. Strictness is tested, not hoped for.
Multiple-choice grading, the mastery math, the rules that open the next topic, the spaced-repetition schedule, and placement are all pure functions with no model in them, unit tested with no database and no key. The model is called only where real judgment is needed: grading written answers and answering tutor questions. Everything mechanical stays mechanical, and testable.
Ask a follow-up and the tutor answers only from chunks of atlas's own lessons, pulled by vector search and cited. When nothing relevant comes back it says the course does not cover that, without ever calling the model, so an empty shelf can never turn into a confident guess. It reads its own teaching text and nothing else.
04 · WHERE IT STANDS
The platform is built and ready to deploy, running on Next.js 16 with Neon Postgres behind it, wired for a single user by design. Seventy-nine tests cover the loop, the grader, and the retrieval, none of them needing a database or a key to run. The point was never a user count. It was proving a learning tool can hold a hard line: grade honestly, teach only what it was taught, and open the next lesson on real mastery instead of good intentions. What is not wired yet is a stronger grader behind the cheap one; the seam is built, the provider is not.
05 · STACK
TRY IT — TOKENS & EMBEDDINGS
atlas's first lesson teaches how an LLM turns text into numbers. Here's a real question from that lesson, graded the same strict way atlas grades everything: by plain code, not vibes.
Q1. A model's vocabulary has 50,000 entries. What best describes what one token actually is?
A. Always exactly one whole English word.
B. A chunk of text, often a sub-word piece, mapped to one ID in the vocabulary.
C. A slot of GPU memory reserved for one API call.
D. A confidence score attached to the model's next guess.
Correct: B. Tokenizers split text into sub-word pieces so a fixed vocabulary can still cover almost any input — "strawberry" might cost 2–3 tokens instead of one. Each ID then looks up a learned embedding vector, which is the number the model actually computes on.
Turn on JavaScript to answer this one yourself and see atlas's real per-option feedback.