How names are made and checked
Bandmark is not an AI chatbot. It is a deterministic generator we built by hand, so the same settings and seed always produce the same names, and every rule can be inspected and tested.
1. A curated lexicon
The engine draws on 880 nouns and 364 descriptors grouped into 28 imagery fields (night, water, frontier, machines, faith, decay, suburbs and so on). Each field has a mood score (dark to bright) and an energy score (hushed to loud), which is what the Mood and Energy faders steer. Mass nouns ("rust", "static") are marked so they are never pluralised, and irregular plurals are stored explicitly.
2. Genre profiles
Each of the 35 styles has its own weighting of imagery fields and naming patterns: "The" + plural, noun pairings with curated modifiers ("Glass", "Iron", "Neon"), closed compounds with tested endings, "X of Y" phrases with grammatical articles, sentence names, frontperson names, family bands, invented place names, initialisms with expansions, music-theory puns, and more. Scene vocabulary is added where a genre has its own (punk's "Rejects" and "Truants", bluegrass's "String Band" and "Hollerers").
3. Real languages, done properly
Latin names use real nouns and adjectives with correct gender and number agreement, or real genitive forms, and always show a translation. Norwegian names join two real Norwegian words (written without special letters) and show their meaning. Invented words are assembled from syllables that carry their own pronunciation, so each one shows a respelling such as "VEH-loh-rah".
4. Quality rules
- No repeated words or word stems inside a name, and no word reused across a batch (except scene suffixes like "Brass Band").
- Batch variety: limits on how many names share a pattern or a first letter.
- Pronounceability checks on invented words (no four-consonant pile-ups, no vowel strings, no stutters).
- Grammar: "a"/"an", articles before countable nouns in "of" phrases, title case with lowercase small words.
- Names you've already seen this session are not repeated when you reroll.
5. Safety filters
Every name passes a filter for slurs, sexual content, hate symbols, real atrocities and self-harm, plus a substring filter for invented words and a list of brand trademarks we don't generate. Coined words that collide with the names of peoples or deities are rejected.
6. Screening against notable artists
We built an index of 120,341 distinct names from 126,643 Wikidata items that have a MusicBrainz artist ID and articles in at least four Wikipedia language editions. Generated names that match one of those names exactly (ignoring "The", case and punctuation), or that contain a multi-word famous name, are discarded. In your browser this uses a compact Bloom filter (141 KB), which may very rarely discard a harmless name but never lets a listed name through.
What this does not do: it cannot see small or local bands, unreleased projects or trademarks. That's why every card has a Check button and why we never describe a name as available.
7. Testing
An automated test suite generates thousands of names per style and checks determinism, variety across batches, the filters, notable-artist screening, Latin agreement, pronounceability, and that settings such as starts-with letter, length, keyword, mood and energy are respected. We re-run it before every release.
Data sources and licences
- Wikidata: artist names, genres and dates (CC0).
- MusicBrainz: live artist lookups in the checker (core data CC0).
- Fonts: Archivo, IBM Plex Mono, Metal Mania, UnifrakturMaguntia, Special Elite, Instrument Serif, Monoton, Unbounded, Rye and Shrikhand, all under the SIL Open Font License, self-hosted.