LUMINA Technical Due Diligence

Comprehensive documentation of statistical methods, architectural decisions, and evaluation metrics for AI music attribution.

Version 7.0 · by Fold Artists · January 2026
512
Signature Dimensions
4.4%
Qualified Threshold (1σ)
68%
Baseline Confidence
~35s
E2E Attribution Time

LUMINA Architecture

End-to-end pipeline from audio generation to rightsholder attribution, leveraging gradient-based signatureing and dual-channel analysis.

🎵
Training Data
99 licensed songs
🧠
MusicGen
Transformer Model
📊
Gradient Capture
P/M Channels
⚖️
Attribution
Kernel Regression

Channel P & Channel M

Attribution is computed through two complementary signal pathways, each capturing different aspects of musical influence.

🎹

Channel P (Composition)

Source: Self-Attention layers (self_attn).

Captures: Melodic patterns, harmonic progressions, and structure.

Technical: Cross-entropy teacher forcing, 10s chunked processing.

🎚️

Channel M (Production)

Source: Output Linear projections (lm.linears).

Captures: Timbre, texture, and sound design.

Technical: 3 intelligent segments per song via librosa.

Significance Thresholds

Thresholds are derived from the expected cosine similarity distribution of random 512-dimensional unit vectors.

Threshold Sigma Level Confidence Meaning
< 4.4% < 1σ < 68% Indistinguishable from noise
≥ 4.4% ≥ 1σ ≥ 68% Qualified Influence
≥ 8.8% ≥ 2σ ≥ 95% High Confidence
≥ 13.2% ≥ 3σ ≥ 99.7% Definitive Proof

LUMINA Influence Potency (LIP)

LIP measures influence using Standardized TracIn Score (STS) with tanh normalization for meaningful percentage-based attribution.

📊

STS (Z-Score)

STS = (score - μ) / σ
Z-score normalized cosine similarity.

🔥

LIP (Tanh Saturation)

LIP = tanh(k × STS)
LIP% = (LIP + 1) / 2 × 100
Maps to 0-100% with 50% baseline.

Why The σ Rules Apply

In 512D space, random unit vectors are nearly orthogonal. Their dot products follow a tight Gaussian distribution around 0 with σ ≈ 4.4%. This makes outlier detection robust.

0 (Random) +1σ (4.4%) Threshold Signal!

Attribution Share System

Songs with score ≥ 1σ (4.4%) qualify. Shares are proportional to their LIP contribution. Share_i = LIP_i / Σ(LIP).

System Performance

On NVIDIA H100 SXM5 (80GB).

~24s
Generation
~80ms
Extraction
~1ms
Attribution
11GB
VRAM

Validation Measures

Rigorous safeguards implemented to ensure attribution accuracy, prevent false positives, and handle edge cases.

⚠️

Low-Energy Filter

Problem: Silent or low-volume segments (e.g., a capella breaks) can produce random high-variance gradients.

Solution: Audio segments with RMS energy below -50dB are strictly excluded from attribution.

Causal Verification

Method: "Ablation Testing". We remove the top attributed song from training and regenerate.

Pass Condition: Output similarity to the removed song must drop by at least 2σ.

🔄

Reproducibility

Guarantee: 100% Deterministic.

Fixed seed 422024 for JL projection ensures that the same audio always produces the exact same signature, essential for legal audits.

Positive-Only Policy

Rule: Negative cosine similarity is ignored.

Rationale: "Anti-influence" (doing the opposite of a song) does not constitute copyright infringement or influence.