No neural network.
No generative model.
Not one song stored inside.

Not the way an AI model is. Music theory is what centuries of musicians distilled from music: tested, taught, corrected, written down as principles, with the actual songs deliberately left behind. A dictionary is made by reading millions of texts, yet it contains words and grammar, not books. Nobody accuses a dictionary of plagiarizing novels. A generative model is the opposite: it keeps the surface. Its parameters are large enough that reproducing fragments of training data has been documented in research. Melognite ships its musical knowledge in readable files you can open: chord progressions stored as scale degrees, rhythm figures stored as strike patterns, and melodic contours stored as abstract interval shapes. Rhythm and contour live in separate pools, so no complete melody is stored anywhere; the actual pitch is computed while the song plays, from the chord underneath. These vocabularies and their weights were measured from music analysis and then reviewed and fixed by hand: calibration, not training. Every number in a pack is a range, a weight or a proportion, never a tune. Same musical conclusions, sometimes. Fundamentally different custody of other people's work.

Rules define a space, not a point. The rules of chess are fixed and finite; no two games are alike. Within everything the theory allows, Melognite makes seeded choices: which motif, which cadence approach, which rhythm. Change the seed, you get a different valid piece. Keep the seed and your settings, you get the identical piece back, for as long as that seed is available (see Terms 4.5) and the engine version is unchanged.

Two things, both inspectable. One: rules as code: scales, functional harmony, cadences, song forms. Two: a vocabulary of abstractions: chord progressions as scale degrees, rhythm figures as strike patterns, melodic contours as interval shapes, kept in separate pools and recombined by the rules; the pitch is computed at play time from the chord underneath. No complete melodies. No recordings. No neural network. No black box. (The packs do contain frequency weights, meaning how often a pattern occurs in a style. That is calibration data, not model weights.)

You need nothing but your browser. One: open the developer tools (F12), pick the Sources tab and read the app's files: the engine and the style packs are plain, readable JavaScript, neither compiled nor obfuscated. Inside the packs you will find scale degrees, strike patterns, interval shapes and weight tables; look for a melody, there is none. Two: switch to the Network tab and play a song: every request goes to our own server, for samples and scripts. There are no calls to any model API, because there is no model to call. Three: copy a share link into a second browser and forge: the identical song appears, note for note. Melognite can make that promise because its musical decisions come from readable, deterministic rules and fixed calibration tables — not from a trained generative model.

Not the way generative models have been shown to. Melognite has no song library to copy from, so it cannot reproduce a recording it was trained on, because it was never trained on any. Like any newly written music, an accidental resemblance to one of the millions of existing melodies can never be ruled out entirely. The same is true for every human composer. When it happens, it is the resemblance of shared grammar (two writers using a V-I cadence), not of a shared source.

Same seed and settings, same song: the identical notes, chords and rhythms, reproducible on the same engine version. (Exported files also carry technical metadata such as an export timestamp, so two exports of the same idea are musically identical without being byte-identical files.) Your instrument, style and slider choices are part of the recipe, not of the seed. A share link carries both. It matters for ownership and for editing: because every note follows from a rule you can see, you change the decision instead of re-rolling a slot machine. A generative model typically offers neither: its output is a sample from a learned distribution.

Depends on the dictionary. Computer science has called rule-based composition "algorithmic" since the 1950s, long before the internet era. What "AI" commonly means today, models trained on existing works whose outputs cannot be traced to readable rules, is precisely what Melognite is not. No training run inside the product, no network, no model. The product ships no recordings and no dataset of recordings. The rules were written down on purpose, by a person; the numbers that calibrate them (how often, how wide, how dense per style) were measured from music analysis and then reviewed and fixed by hand. You can go read both.

Musicians, not metrics. Rules can guarantee that a progression is valid and that a phrase closes properly; only a trained ear can tell whether a chorus actually lands. So before launch, every genre goes through extensive listening tests with experienced musicians, and each one is fine-tuned in close collaboration with them, genre by genre, style by style.