The receipts page.
Every public trust signal SongForgeAI publishes, consolidated in one place. No marketing copy on this page — just links to the canonical surfaces, with one-line context for what each one proves.
If a skeptical buyer asks “how do I know any of this is real?”, this is the URL to send them. Every linked surface is the source-of-truth artifact, not a derived summary.
Reproducibility
Every score SongForgeAI emits carries a cryptographic seal pinning the rubric version, eval model, temperature, and build SHA. Third parties can verify a score was actually produced by the system we claim it was.
Reproducibility seal
ed25519-signed receipt structure + the canonical public key. Every API response includes one.
Verify a seal
Paste any score's seal block + signature; the page rejects tampering and confirms authentic scores.
Model card
Pinned eval-stream model, temperature, max-tokens, and the prompt-fingerprint gate that detects rubric drift.
Open Scoring Standard
The 12-metric Lyric Scoring rubric is published as a versioned open standard under CC BY 4.0. Anyone can audit, cite, or extend it.
The Standard
Versioned rubric (v1.0). All 12 metrics + tier weights + anti-inflation rules.
Whitepaper v0.9
Methodology, calibration, why each metric exists. The document third parties cite.
Standard changelog
Every rubric version + diff + reason for the change. Open commit log for the rubric itself.
Prior art
Conservatory rubrics, MIR research, industry standards we extend. Honest positioning, not category-invention claims.
Multi-tradition calibration anchors
Five 95-band canonical works across five musical traditions (Williams · Mitchell · Piazzolla · Marley · Veloso) with full per-metric rationale.
Inter-rater reliability
How the rubric agrees with itself across runs + agreement metrics against the GPT-4o triangulation pass.
Engineering discipline
The build process is itself a trust signal. Every push to main runs CI gates, every commit cites the moat it advanced, every cadence ritual logs publicly.
The Receipts
4075 unit tests · 291 banned-term scanner · golden-eval drift gate · prompt-fingerprint gate · component-LOC ratchet · explicit-any allowlist at zero.
How we build
The engineering operating system: punch list, cadence rituals, ratchets, the rule of one-commit-one-moat.
Cadence ritual logs
Quality Council (every 3d), Trust Decay Audit (every 14d), Bet Review (every 45d), External Audit Prep (every 45d). Append-only logs.
Build log
Public commit log with the moat each build advanced. No private channel — every change ships here.
Roadmap
What's shipped, what's in flight, what's next. Auto-derived from the punch list + three-year plan.
Forge quality floor
The forge pipeline carries multiple quality floors that prevent specific failure modes. Each is auditable infrastructure, not a marketing claim.
Curated Artist Library — 504 entries
Hand-authored briefs for every artist in the library. When you ask for an artist-inspired song, the brief is the forge's anchor — not the model's training-data recall.
Per-genre scoring guides
Genre-specific calibration notes for the rubric. Country, rap, jazz, gospel, EDM each carry their own context.
Public craft-metrics report
Aggregate distributions across every forged song. Live telemetry, hourly refresh.
Public leaderboard
Top-scoring public songs across all metrics. Same rubric, same seal — directly comparable.
Honest disclosures
What the system gets wrong, what we don't yet know, where the limits are. Trust is built more by what we admit than what we claim.
Audit-yourself ledger
The maintainer's posture on system bias + known calibration gaps. Permanent record of what the rubric mis-scores.
Ethical commitments
Four explicit commitments consolidated from the Terms + Privacy + Whitepaper. Cross-doc navigation surface.
Your rights
IP ownership, data retention, deletion, export. Plain-English summary of §4 Terms + §3 Privacy.
Cited-by tracker
External implementations + citations of the Lyric Scoring Standard. Empty when no one's cited; honest when they have.