Rap subgenre system

The 5 subgenre profiles

Each subgenre carries quantitative targets (BPM band, syllables-per-bar band, internal-rhyme rate band, multisyllabic chain minimum) + a banned-failure-modes list specific to that form. The forge prompt, the audit, and the critic all read from the same definitions.

The Rap Forbidden Archive — 10 canonical failures

When the critic flags a bar, it uses one of the 10 canonical failure names so eval output + critic output share vocabulary. Full definitions in docs/RAP-FORBIDDEN-ARCHIVE.md.

Forge a rap song

The standard forge endpoint accepts a genre string + prompt. When the detector resolves to a rap subgenre, the forge prompt automatically embeds:

• The subgenre's cadence + diction profile
• BPM band + syl/bar band + multisyllabic chain minimum
• Subgenre-specific banned failure modes
• The cultural authenticity rule
• A 16-bar BAR MAP with per-bar job + rhyme letter + syllable target
• Plan-then-draft-then-self-critique output discipline

curl -X POST https://songforgeai.com/api/songs/forge \
  -H "Authorization: Bearer <YOUR_SESSION>" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "boom bap song about leaving the old block",
    "genre": "boom bap",
    "vocalGender": "male"
  }'

Subgenre keywords that route automatically: boom bap · boom-bap · trap · atlanta trap · drill · uk drill · conscious rap · jazz rap · storytelling rap · political rap. Generic rap / hip hop falls through to MODE_RAP catch-all.

Programmatic critique (server-side)

For server-side integrations, the rap critic is callable directly via the internal library. Returns the 9-dimension scorecard + per-bar failure flags + rewrite priority list.

import { critiqueRapVerse } from '@songforgeai/server';

const result = await critiqueRapVerse({
  lyrics: verse,
  brief: { premise: 'a song about leaving the old block', anchors: ['water tower'] },
  subgenreId: 'boom-bap',
});

console.log(`${result.verdict} (composite ${result.composite})`);
for (const flag of result.flags) {
  console.log(`  bar ${flag.bar}: ${flag.failure}`);
}

One Haiku call per critique. ~$0.0005 + ~2-4s wall-clock. LRU-cached (500 entries, 15-min TTL) keyed on lyric + subgenre + premise.

Calibration + drift detection

Score thresholds live in one home — src/lib/claude/rap-calibration.ts — so the critic, the corpus analyzer, the dashboard, and this doc page all read the same numbers. Methodology:

• Probe — runs the audit on a curated baseline of synthetic + commissioned verses. Calibration reference.
• Corpus analyzer — runs the audit on production exports (JSONL); surfaces per-subgenre distributions + cross-subgenre fairness flags.
• Re-run cycle — every rap audit / plugin change triggers a probe + analyzer re-run; report diffs detect drift across builds.

Full methodology in docs/RAP-CALIBRATION.md. Corpus probe runbook: RAP-CORPUS-PROBE-RUNBOOK.md. Subgenre analysis runbook: RAP-CORPUS-ANALYSIS-RUNBOOK.md.

Next