A transformer that encodes and decodes base64, made by qkvc's builder: every weight placed by rule, no training. The page runs it here, live, on the real weights (the same numbers as the Python model, checked equal), and checks every token it writes against your browser's own base64 (btoa / atob).
Base64 turns every 3 bytes into 4 characters of 6 bits. The stream is written group by group, ENC H e l S G V s, so what an output needs is always the same distance back, however long the input is. Each output needs at most two input symbols. Two attention heads each fetch one of them (the query names a phase, only the token at that phase matches, the weight on it is 1). The MLP's units are AND gates that copy each fetched bit to its place in the output. The output layer gives every token a score by how many bits agree, and the highest score is the next token.
python -m qkvc verify base64 reads these weights and proves the model follows the task's rules for every valid input of any length. The rules are checked against real base64: every full group of bytes, every one- and two-byte ending, and a sample of whole inputs.