Movie Musical Generator

Every word is written for you. The app does not reproduce the words of an existing song, in whole or in part, verbatim or reworded, and it will not hand back somebody else's book with the names changed. Where a premise names a real film or a real musical, what comes back is an original show riffing on the same idea — never that show's book, score or running order under a new title.

It will not attribute anything it writes to a real writer, composer or performer; asked for a named person's voice it writes to the genre and the era instead, and says so. It will not build a show on a living person's private life — it puts a fictional character in the same situation.

The request-side guard on this page is a pre-flight filter, not the boundary. It catches the plain forms of those requests before a credit is spent. The prompt itself is what holds everything else. The guard's measured numbers are in the next panel and in llms.txt.

An independent agent wrote the test corpus without being allowed to read the guard, because a corpus written by whoever wrote the fix measures the fix rather than the guard. Scored once, on a set the guard had never seen:

A single aggregate would hide which sixth is missing, so the headline here is the weakest column, not the overall. Invented claims about real living people at 19/26, 73%, is the guard's honest remaining exposure, and it is a different kind of risk from the rest — fabricated fact about an identifiable person rather than somebody else's words. The seven that got through mostly identify their subject by description rather than by name, which is the hardest case: no list of names can close it, and the app leans on the model's own knowledge there. Two earlier corpora scored 32% and 84%; the 32% was before the fixes and the 84% was against a corpus that had by then seen the guard, which is why neither is quoted.

Three things that number does not mean. It is not the app's safety boundary — the system prompt is, and it has world knowledge this filter cannot. It is not stable against an attacker who reads this page. And the classes are not equally serious: the floor is absolute and sits at 100%, while the attribution column is a notice rather than a refusal, because writing in a named person's style is legitimate and refusing it would be the defect.

The work is split in two because it is two different jobs. A treatment plus a full song list plus eight complete lyrics does not fit inside one reply from any model, and a model asked for all of it does not truncate politely — it fails. So the first call gives you the picture and the running order, and the second writes one chosen number properly. That two-stage generation is the whole reason the lyric arrives with its staging notes intact instead of cut off mid-verse.

Two checks run in your browser and cost nothing. The structural check over a song list asks whether every number actually names a change, whether the slots run in order, whether an act has been left without a song, whether a reprise answers anything, and whether two numbers are doing the same job in the same words — so the act structure is tested rather than assumed. The craft check over a lyric works out the rhyme scheme, counts the syllables line by line, notices a refrain whose behaviour is to come back unaltered, and reports how ordinary — how well travelled — each line's grammatical construction is.

That last check is worth being precise about. It says a construction is common. It does not say a line already exists, and it could not: there is no corpus of anybody's lyrics in this bundle, deliberately, and the check is built from grammatical shapes rather than from anyone's words.

Two known limits, stating them here rather than leaving them to be discovered. The syllable counter is a heuristic, not a pronouncing dictionary: it is right on ordinary English and wrong on some of what English borrowed. Rhyme detection reads spelling rather than sound, so it under-reports a rhyme spelled two ways and over-reports an eye-rhyme. The structural checks themselves are proved rather than assumed — a known-good show is broken one specific way at a time and the right check has to catch each break. The two scans between them can raise 32 distinct findings.