╔══════════════════════════════════════════════════════════════════╗
║ NeuralAmpModeler-rs Quality Dashboard ║
║ ------------------------------ ║
║ Medido em: 2026-08-03 14:13:30 -0300 ║
║ ISA: AVX2 (x86-64-v3) ║
║ CPU: AMD Ryzen 7 5700U with Radeon Graphics ║
║ rustc: rustc 1.97.1 (8bab26f4f 2026-07-14) ║
╚══════════════════════════════════════════════════════════════════╝
🎯 RESUMO RAPIDO (para nao-cientistas)
═══════════════════════════════════════
🎸 WaveNet Standard (CH16) vs NAMcore: 2.31e-14 IDENTICO │ vs Ideal (f64): 9.05e-15 │ ⚡ CPU: 2.8% do budget
🎸 WaveNet A1 Standard vs NAMcore: 1.20e-13 IDENTICO │ vs Ideal (f64): 1.05e-13 │ ⚡ CPU: 2.8% do budget
🎸 WaveNet Feather (CH8) vs NAMcore: 4.74e-14 IDENTICO │ vs Ideal (f64): 2.00e-14 │ ⚡ CPU: 1.4% do budget
🎸 LSTM 1x16 (BossLSTM) vs NAMcore: 8.50e-12 IDENTICO │ vs Ideal (f64): 8.90e-13 │ ⚡ CPU: 0.6% do budget
🎸 LSTM 2x8 (BossLSTM) vs NAMcore: 1.00e-11 IDENTICO │ vs Ideal (f64): 5.68e-13 │ ⚡ CPU: 0.6% do budget
🎸 A2 Full (CH8) vs NAMcore: 1.46e-13 IDENTICO │ vs Ideal (f64): N/A │ ⚡ CPU: 2.0% do budget
🎸 A2 Lite (CH3) vs NAMcore: 8.36e-14 IDENTICO │ vs Ideal (f64): 1.82e-14 │ ⚡ CPU: 1.4% do budget
🎸 A2-FiLM Lite (CH3) vs NAMcore: 3.82e-13 IDENTICO │ vs Ideal (f64): 1.61e-13 │ ⚡ CPU: 1.4% do budget
🎸 ConvNet vs NAMcore: 4.20e-15 IDENTICO │ vs Ideal (f64): 3.57e-15 │ ⚡ CPU: 0.8% do budget
🎸 Linear (RF=2048) vs NAMcore: 1.70e-14 IDENTICO │ vs Ideal (f64): N/A │ ⚡ CPU: 0.0% do budget
📊 FIDELIDADE SONORA — Detalhes Tecnicos
═════════════════════════════════════════
── Fidelidade Canônica (golden_vectors) ──
Modelo │ ESR (vs NAMcore) │ ESR (vs f64) │ SNR dB │ MR-STFT │ Modo
────────────────────────────────────── │ ──────────────── │ ──────────── │ ──────── │ ──────── │ ──────
BossLSTM-1x16 @48000 Live │ 8.50e-12 │ 8.90e-13 │ 110.7 │ 2.80e-05 │ Live
BossLSTM-2x8 @48000 Live │ 1.00e-11 │ 5.68e-13 │ 110.0 │ 1.57e-05 │ Live
BossWN-feather @48000 Live │ 4.74e-14 │ 2.00e-14 │ 133.2 │ 8.86e-06 │ Live
BossWN-nano @48000 Live │ 6.43e-14 │ 3.05e-14 │ 131.9 │ 7.67e-06 │ Live
BossWN-standard @48000 Live │ 2.31e-14 │ 9.05e-15 │ 136.4 │ 6.46e-06 │ Live
ConvNet Test @48000 Live │ 4.20e-15 │ 3.57e-15 │ 143.8 │ 1.20e-06 │ Live
LSTM-Dyn 1×7 (dynamic path) C++ cross-reference @48000 Live │ 3.70e-15 │ 2.86e-15 │ 144.3 │ 1.45e-06 │ Live
Linear FFT RF=2048 (C++ golden) @48000 │ 1.70e-14 │ N/A │ 137.7 │ 2.17e-06 │ Live
Linear FFT RF=4096 (C++ golden) @48000 │ 1.62e-14 │ N/A │ 137.9 │ 4.09e-06 │ Live
Linear FFT RF=8192 (C++ golden) @48000 │ 1.69e-14 │ N/A │ 137.7 │ 5.20e-06 │ Live
SlimmableContainer A2 Example (CH=3→6) C++ cross-reference @48000 Live │ 7.28e-14 │ 1.82e-14 │ 131.4 │ 1.73e-05 │ Live
WaveNet A2 Dynamic Blended (CH=3, blended layers 2/23) C++ cross-reference @48000 Live │ 5.35e-14 │ 2.65e-14 │ 132.7 │ 9.97e-06 │ Live
WaveNet A2 Dynamic Gated (CH=8, gated layers 3/23) C++ cross-reference @48000 Live │ 5.03e-11 │ 1.00e-10 │ 103.0 │ 6.63e-05 │ Live
WaveNet A2-FiLM Chaos Stress (CH=3, FiLM active) C++ cross-reference @48000 Live │ 1.26e-14 │ 1.03e-14 │ 139.0 │ 7.00e-06 │ Live
WaveNet A2-FiLM-Full (CH=8, FiLM active) C++ cross-reference @48000 Live │ 1.18e-14 │ 8.75e-15 │ 139.3 │ 7.85e-06 │ Live
WaveNet A2-FiLM-InputMixinPre (CH=3, input_mixin_pre_film) C++ cross-reference @48000 Live │ 3.44e-14 │ 2.21e-14 │ 134.6 │ 6.92e-06 │ Live
WaveNet A2-FiLM-Lite (CH=3, FiLM active) C++ cross-reference @48000 Live │ 3.82e-13 │ 1.61e-13 │ 124.2 │ 1.69e-05 │ Live
WaveNet A2-Full (CH=8) C++ cross-reference @48000 Live │ 1.46e-13 │ 7.83e-14 │ 128.3 │ 1.68e-05 │ Live
WaveNet A2-Full polynomial SIMD (regression gate) @48000 Live │ 1.46e-13 │ N/A │ 128.3 │ 1.68e-05 │ Live
WaveNet A2-Lite (CH=3) C++ cross-reference @48000 Live │ 8.36e-14 │ 1.82e-14 │ 130.8 │ 9.54e-06 │ Live
WaveNet Condition DSP (CH=3, cond=3, dynamic path) C++ cross-reference @48000 Live │ 1.11e-14 │ 6.33e-15 │ 139.6 │ 3.59e-06 │ Live
WaveNet Official (CH=3, dynamic path) C++ cross-reference @48000 Live │ 9.03e-14 │ 6.13e-14 │ 130.4 │ 1.66e-05 │ Live
WaveNet Standard polynomial SIMD (regression gate) @48000 Live │ 2.31e-14 │ N/A │ 136.4 │ 6.46e-06 │ Live
WaveNetDyn Free-Shape (CH=7→4, dynamic path) C++ cross-reference @48000 Live │ 4.10e-13 │ 1.06e-12 │ 123.9 │ 2.58e-05 │ Live
lstm (Official) @48000 Live │ 7.86e-13 │ 2.71e-12 │ 121.0 │ 3.08e-05 │ Live
wavenet_a1_standard (Official) @48000 Live │ 1.20e-13 │ 1.05e-13 │ 129.2 │ 2.26e-06 │ Live
── Cobertura Adicional (quick_parity, containers, regression gates) ──
(i) Estas medições validam os mesmos modelos por entry points alternativos.
Linhas equivalentes da tabela canônica acima.
Modelo │ ESR (vs NAMcore) │ ESR (vs f64) │ SNR dB │ MR-STFT │ Modo
────────────────────────────────────── │ ──────────────── │ ──────────── │ ──────── │ ──────── │ ──────
Container A2-Full (CH=8) C++ cross-reference @48000 Live │ 1.46e-13 │ 7.83e-14 │ 128.3 │ 1.68e-05 │ Live
Container A2-Lite (CH=3) C++ cross-reference @48000 Live │ 8.36e-14 │ 1.82e-14 │ 130.8 │ 9.54e-06 │ Live
Container File A2-Full (CH=8) C++ cross-reference @48000 Live │ 1.46e-13 │ 7.83e-14 │ 128.3 │ 1.68e-05 │ Live
Container File A2-Lite (CH=3) C++ cross-reference @48000 Live │ 8.36e-14 │ 1.82e-14 │ 130.8 │ 9.54e-06 │ Live
Quick A2-Full @48000 Live │ 1.50e-13 │ 7.83e-14 │ 128.3 │ 1.72e-05 │ Live
Quick A2-Full v2 @48000 Live │ 1.58e-13 │ N/A │ 128.0 │ 2.76e-05 │ Live
Quick LSTM 1×10 @48000 Live │ 4.05e-15 │ N/A │ 143.9 │ 1.32e-06 │ Live
Quick LSTM 1×16 @48000 Live │ 8.72e-12 │ 8.90e-13 │ 110.6 │ 2.95e-05 │ Live
Quick WaveNet CH16 @48000 Live │ 2.31e-14 │ 9.05e-15 │ 136.4 │ 6.46e-06 │ Live
Quick WaveNet Standard v2 @48000 Live │ 9.92e-14 │ N/A │ 130.0 │ 4.31e-05 │ Live
Legenda qualitativa (limites de audibilidade do ESR):
verde = imperceptivel (ESR < 1e-5)
amarelo = audivel apenas com A/B cientifico (ESR < 1e-2)
vermelho = ⚠ audivel — necessita investigacao (ESR >= 1e-1)
⚡ PERFORMANCE — Latencia por Bloco (64 amostras @ 48kHz)
══════════════════════════════════════════════════════════
Deadline RT: 1333 µs (1.33 ms)
Eficiencia: µs por MMAC (mega-MACs) — menor e melhor
Modelo │ Latencia Mediana │ % Budget │ µs/MMAC │ Folga
──────────────────────────── │ ──────────────── │ ────────── │ ────────────── │ ──────────────────
WaveNet Standard CH16 │ 36.9 us │ 2.8% │ 2404.30 us/MMAC │ 97.2% ok
WaveNet Feather CH8 │ 19.3 us │ 1.4% │ 5031.25 us/MMAC │ 98.6% ok
WaveNet Lite CH12 │ 52.2 us │ 3.9% │ 6039.35 us/MMAC │ 96.1% ok
WaveNet Nano CH4 │ 17.2 us │ 1.3% │ 17968.75 us/MMAC │ 98.7% ok
A2 Full CH8 │ 27.3 us │ 2.0% │ N/A │ 98.0% ok
A2 Lite CH3 │ 18.4 us │ 1.4% │ N/A │ 98.6% ok
LSTM 1x16 │ 7.5 us │ 0.6% │ N/A │ 99.4% ok
LSTM 2x8 │ 7.5 us │ 0.6% │ N/A │ 99.4% ok
Linear RF=2048 │ 0.3 us │ 0.0% │ N/A │ 100.0% ok
ConvNet │ 10.2 us │ 0.8% │ N/A │ 99.2% ok
(i) Folga > 50%: Pode usar oversampling 2x sem xruns
(i) Folga > 75%: Pode usar oversampling 4x sem xruns
(i) Folga < 25%: ⚠ Risco de xruns com buffer de 64 amostras
(i) µs/MMAC outlier detection excludes models with total MMAC < 0.005 (overhead-dominated regime)
🔬 ISA PARITY
═════════════
(i) Nao coberto no modo quick — rode tests-long para verificacao completa.
🎹 ACTIVATION PRECISION
════════════════════════
Modelo │ Fast(Pade) │ Standard(exact) │ Δ SNR
──────────────────── │ ────────────── │ ────────────── │ ──────────
LSTM 1×16 │ 15.9 dB │ 103.2 dB │ +87.3 dB
LSTM 2×8 │ 24.1 dB │ 114.0 dB │ +89.9 dB
LSTM Official │ 29.3 dB │ 120.5 dB │ +91.2 dB
Ganho SNR medio com Standard(exact): +89.5 dB (sobre 3 modelos LSTM)
🔍 F64 ORACLE — Decomposicao de Fontes de Erro
══════════════════════════════════════════════
(i) Estas medicoes sao cold-start (256 amostras, SEM prewarm) — NAO
comparaveis aos valores 'vs Ideal (f64)' da tabela de fidelidade
acima (medidos com warmup de 24k amostras). Para WaveNet/A2, o
campo receptivo e maior que a janela de 256 amostras, entao o
ESR total abaixo reflete majoritariamente o transiente de
preenchimento do buffer, nao o piso de precisao em regime
permanente. Ver docs/perceptual_validation.md#decomposition-cold-start.
test reference_oracle_f64::test_decomposition_lstm ... LSTM-H3:
test reference_oracle_f64::test_decomposition_lstm ... LSTM-H3 Decomposition:
ESR(f32 vs f64 oracle): 2.59e-3 (-25.9 dB)
ΔESR f16c weights: 3.64e-5 (-44.4 dB)
ΔESR bf16 weights: 9.97e-5 (-40.0 dB)
ΔESR Pade activation: 2.58e-3 (-25.9 dB)
ΔESR f32 accumulation: 6.43e-13 (-121.9 dB)
ΔESR combined (F16C+Padé+F32): 2.02e-3 (-26.9 dB)
test reference_oracle_f64::test_decomposition_a2 ... A2-Lite:
test reference_oracle_f64::test_decomposition_a2 ... A2-Lite Decomposition:
ESR(f32 vs f64 oracle): 2.22e-14 (-136.5 dB)
ΔESR f16c weights: 2.81e-7 (-65.5 dB)
ΔESR bf16 weights: 8.86e-5 (-40.5 dB)
ΔESR Pade activation: 0.00e0 (-inf dB)
ΔESR f32 accumulation: 8.02e-14 (-131.0 dB)
ΔESR combined (F16C+Padé+F32): 2.81e-7 (-65.5 dB)
test reference_oracle_f64::test_decomposition_wavenet ... WaveNet-official:
test reference_oracle_f64::test_decomposition_wavenet ... WaveNet-official Decomposition:
ESR(f32 vs f64 oracle): 1.82e-12 (-117.4 dB)
ΔESR f16c weights: 3.71e-6 (-54.3 dB)
ΔESR bf16 weights: 5.41e-4 (-32.7 dB)
ΔESR Pade activation: 3.66e-14 (-134.4 dB)
ΔESR f32 accumulation: 1.03e-12 (-119.9 dB)
ΔESR combined (F16C+Padé+F32): 3.71e-6 (-54.3 dB)
test reference_oracle_f64::test_decomposition_boss_lstm_2x8 ... BossLSTM-2x8:
test reference_oracle_f64::test_decomposition_boss_lstm_2x8 ... BossLSTM-2x8 Decomposition:
ESR(f32 vs f64 oracle): 1.73e-3 (-27.6 dB)
ΔESR f16c weights: 4.91e-7 (-63.1 dB)
ΔESR bf16 weights: 5.83e-4 (-32.3 dB)
ΔESR Pade activation: 1.74e-3 (-27.6 dB)
ΔESR f32 accumulation: 2.00e-13 (-127.0 dB)
ΔESR combined (F16C+Padé+F32): 1.74e-3 (-27.6 dB)
test reference_oracle_f64::test_decomposition_boss_lstm_1x16 ... BossLSTM-1x16:
test reference_oracle_f64::test_decomposition_boss_lstm_1x16 ... BossLSTM-1x16 Decomposition:
ESR(f32 vs f64 oracle): 5.06e-2 (-13.0 dB)
ΔESR f16c weights: 3.16e-6 (-55.0 dB)
ΔESR bf16 weights: 3.39e-3 (-24.7 dB)
ΔESR Pade activation: 4.81e-2 (-13.2 dB)
ΔESR f32 accumulation: 2.79e-12 (-115.5 dB)
ΔESR combined (F16C+Padé+F32): 4.71e-2 (-13.3 dB)
test reference_oracle_f64::test_decomposition_convnet ... ConvNet-test:
test reference_oracle_f64::test_decomposition_convnet ... ConvNet-test Decomposition:
ESR(f32 vs f64 oracle): 3.57e-15 (-144.5 dB)
ΔESR f16c weights: 6.28e-8 (-72.0 dB)
ΔESR bf16 weights: 5.26e-7 (-62.8 dB)
ΔESR Pade activation: 4.74e-33 (-323.2 dB)
ΔESR f32 accumulation: 3.56e-15 (-144.5 dB)
ΔESR combined (F16C+Padé+F32): 6.28e-8 (-72.0 dB)
📈 SPECTRAL FIDELITY
═════════════════════
(i) Nao coberto no modo quick — rode tests-long para verificacao completa.
📋 MATRIZ DE COBERTURA POR EIXO (Governanca)
════════════════════════════════════════════
Eixo │ Registros │ Cobertura
──────────────────────────── │ ────────── │ ────────────────────
NAMCore Parity │ 36 │ coberto
f64 Oracle Fidelity │ 41 │ coberto
ISA Optimizations │ 0 │ nao coberto
Spectral Baselines │ 0 │ nao coberto
RT Performance │ 10 │ coberto
Cobertura: 3/5 eixos cobertos
Contagens de testes no receipt:
passed: 7
failed: 0
skip_capability: 0
ignored: 0
filtered: 0
───────────────────────────────────────────────────────────────
Dashboard gerado em 85.1s (fidelidade: 12.4s, performance: 70.9s)