Most music apps promote you when a number gets big enough. Ours holds your rank until you finish a song, and from yellow belt up, until you pass a test: write a song to a brief, have it graded, clear a floor. The mechanism is public in the code and it should be public in prose too, because a gate you cannot see is a gate you cannot train for.
The brief is canonical. Whatever assignment your device sends along with the submission is replaced on the server by the one belt test brief for the belt you are sitting, so the model grades against the real constraints, not a softened copy. The belt you are graded at is the higher of the belt you claim and the belt on your profile: a black belt sitting a yellow test is still read as a black belt. Claiming a lower belt buys nothing.
The floors climb. Yellow and orange tests need a C. Green and blue need a B minus. Purple, brown and red need a B. Black needs an A minus. Those floors were calibrated against what the model actually grades: a floor the grader never emits would round up silently and mean nothing, so the calibration is a test in the repository that reads the committed eval results and refuses a ladder the observed grading cannot reach.
Two rules make it a test rather than a lottery. A single pass is not a promotion: you need two passing attempts, so a lucky critique on a good day cannot carry you. And a missed constraint is an F, whatever the prose grade says. The brief says the chorus must land inside the first minute, or the second verse must move time, and the model is asked to quote the line that proves each constraint was met. If it cannot, the constraint is unmet and the attempt fails. A beautiful song that ignores the brief is a beautiful song, not a pass. After a fail there is a twenty-four-hour cooldown, enforced on the server, so the retry is a rewrite and not a refresh.
Our own harness tells us where the grading is weakest, and it is not where you might guess. Across the last committed run the grader was most accurate at white belt (90 percent of rubric checks) and least accurate at black and red (81 and 82 percent), because the deepest critiques are the longest and the hardest to keep honest. That is why every belt test gets a second verification pass: the model re-reads its own critique against the lyric and drops any claim it cannot support, and a grade that was pulled down by a claim that did not survive is raised. It costs a second call. On a belt test the grade is load-bearing, so it is worth it.
What this means for you
read the brief the way a co-writer reads a pitch sheet. Meet every constraint literally before you make the song good. Expect the grader to quote you back to yourself, and give it lines worth quoting. And write two passing songs, not one, because that is what the belt certifies.
The drill on this page is prosody match, because prosody is the concept every belt test brief leans on and the one the harness says the grader is best at hearing: does the stressed syllable land on the strong beat, does the stable rhyme sit in the settled section. Pass the drill and you have passed the part of the test that is not about your ideas.
