diff --git a/FL2VA/audio_vae/__pycache__/dac_alias_free_filter.cpython-314.pyc b/FL2VA/audio_vae/__pycache__/dac_alias_free_filter.cpython-314.pyc
new file mode 100644
index 0000000..2284ad9
Binary files /dev/null and b/FL2VA/audio_vae/__pycache__/dac_alias_free_filter.cpython-314.pyc differ
diff --git a/FL2VA/audio_vae/__pycache__/dac_alias_free_resample.cpython-314.pyc b/FL2VA/audio_vae/__pycache__/dac_alias_free_resample.cpython-314.pyc
new file mode 100644
index 0000000..a5775fb
Binary files /dev/null and b/FL2VA/audio_vae/__pycache__/dac_alias_free_resample.cpython-314.pyc differ
diff --git a/FL2VA/audio_vae/__pycache__/dac_attn_proj.cpython-314.pyc b/FL2VA/audio_vae/__pycache__/dac_attn_proj.cpython-314.pyc
new file mode 100644
index 0000000..dda6dd9
Binary files /dev/null and b/FL2VA/audio_vae/__pycache__/dac_attn_proj.cpython-314.pyc differ
diff --git a/FL2VA/audio_vae/dac_alias_free_filter.py b/FL2VA/audio_vae/dac_alias_free_filter.py
index ad02595..26359fc 100644
--- a/FL2VA/audio_vae/dac_alias_free_filter.py
+++ b/FL2VA/audio_vae/dac_alias_free_filter.py
@@ -49,11 +49,11 @@ def kaiser_sinc_filter1d(cutoff, half_width, kernel_size):  # return filter [1,1
         filter_ = torch.zeros_like(time)
     else:
         filter_ = 2 * cutoff * window * sinc(2 * cutoff * time)
-        """
-        Normalize filter to have sum = 1, otherwise we will have a small leakage of the constant component in the input signal.
-        """
+        # Normalize filter to have sum = 1, otherwise we will have a small
+        # leakage of the constant component in the input signal.
         filter_ /= filter_.sum()
-        filter = filter_.view(1, 1, kernel_size)
+    # Always reshape into [1, 1, kernel_size] so the return is never undefined.
+    filter = filter_.view(1, 1, kernel_size)
 
     return filter
 
diff --git a/FL2VA/audio_vae/dac_attn_proj.py b/FL2VA/audio_vae/dac_attn_proj.py
index d3b9b1f..6f1cd2e 100644
--- a/FL2VA/audio_vae/dac_attn_proj.py
+++ b/FL2VA/audio_vae/dac_attn_proj.py
@@ -29,15 +29,19 @@ class CausalAttention(nn.Module):
     def __init__(self, in_dim, out_dim, num_heads):
         super().__init__()
         if in_dim > out_dim:
-            # assert in_dim // num_heads == out_dim
+            # Projection compresses in_dim → out_dim after attention.
+            # QKV operates in in_dim space so head_dim = in_dim // num_heads.
             self.head_dim = in_dim // num_heads
+            self.qkv_out_dim = in_dim
             self.qkv = nn.Linear(in_dim, in_dim * 3, bias=False)
             self.q_bias = nn.Parameter(torch.zeros(in_dim))
             self.v_bias = nn.Parameter(torch.zeros(in_dim))
             self.register_buffer("zero_k_bias", torch.zeros(in_dim))
         else:
-            # assert out_dim // num_heads == in_dim
+            # Projection expands (or keeps) in_dim → out_dim after attention.
+            # QKV operates in out_dim space so head_dim = out_dim // num_heads.
             self.head_dim = out_dim // num_heads
+            self.qkv_out_dim = out_dim
             self.qkv = nn.Linear(in_dim, out_dim * 3, bias=False)
             self.q_bias = nn.Parameter(torch.zeros(out_dim))
             self.v_bias = nn.Parameter(torch.zeros(out_dim))
@@ -47,21 +51,27 @@ def __init__(self, in_dim, out_dim, num_heads):
         self.out_dim = out_dim
         self.num_heads = num_heads
         self.scale = self.head_dim**-0.5
-        self.proj = nn.Linear(out_dim, out_dim)
+        # Final linear maps from the QKV output space to out_dim.
+        self.proj = nn.Linear(self.qkv_out_dim, out_dim)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         B, N, C = x.shape
-        qkv = F.linear(input=x, weight=self.qkv.weight, bias=torch.cat((self.q_bias, self.zero_k_bias, self.v_bias)))
-        q, k, v = qkv.reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4).unbind(0)
+        qkv = F.linear(
+            input=x,
+            weight=self.qkv.weight,
+            bias=torch.cat((self.q_bias, self.zero_k_bias, self.v_bias)),
+        )
+        q, k, v = (
+            qkv.reshape(B, N, 3, self.num_heads, self.head_dim)
+            .permute(2, 0, 3, 1, 4)
+            .unbind(0)
+        )
 
         x = scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=True)
 
-        if self.in_dim > self.out_dim:
-            x = torch.mean(x, dim=1)
-            if self.in_dim // self.num_heads != self.out_dim:
-                x = nn.functional.adaptive_avg_pool1d(x, self.out_dim)
-        else:
-            x = x.transpose(1, 2).reshape(B, N, -1)
+        # Always reshape to [B, N, qkv_out_dim] so self.proj receives the
+        # correct input dimension regardless of whether in_dim > out_dim or not.
+        x = x.transpose(1, 2).reshape(B, N, self.qkv_out_dim)
         x = self.proj(x)
         return x
 
diff --git a/FL2VA/audio_vae/dac_audio_vae.py b/FL2VA/audio_vae/dac_audio_vae.py
index aebdfb7..d2a9b55 100644
--- a/FL2VA/audio_vae/dac_audio_vae.py
+++ b/FL2VA/audio_vae/dac_audio_vae.py
@@ -43,6 +43,17 @@ def forward(self, x):
 def init_weights(m):
     """Initialize Conv1d layers, including those wrapped with ``weight_norm``.
 
+    This is the canonical initializer for ``DacAudioVAE`` (encoder + decoder).
+    It uses ``trunc_normal_`` with ``std=0.02`` and explicitly zeroes biases,
+    matching the intended training configuration for this model family.
+
+    Note: ``dac_bigvgan.py`` / ``dac_utils.py`` also contain an ``init_weights``
+    helper (``std=0.01``, ``normal_``) used by BigVGAN's own module-level
+    ``apply()`` calls.  Because ``DacAudioVAE.__init__`` calls
+    ``self.apply(init_weights)`` *after* all sub-modules (including BigVGAN)
+    have been constructed, this function is the one that takes final effect on
+    every ``nn.Conv1d`` in the whole model graph.
+
     The encoder (``WNConv1d``) and decoder (BigVGAN) both wrap ``nn.Conv1d``
     with ``torch.nn.utils.parametrizations.weight_norm``. Under that
     parametrization ``m.weight`` is computed on access from
@@ -74,9 +85,15 @@ def __init__(self, dim: int = 16, dilation: int = 1):
 
     def forward(self, x):
         y = self.block(x)
-        pad = (x.shape[-1] - y.shape[-1]) // 2
-        if pad > 0:
-            x = x[..., pad:-pad]
+        diff = x.shape[-1] - y.shape[-1]
+        if diff > 0:
+            # Crop the residual path to match the convolved output length.
+            # Use explicit left/right amounts to handle both even and odd
+            # differences safely (avoids the x[..., pad:-pad] pattern which
+            # silently drops an extra sample when diff is odd).
+            pad_left = diff // 2
+            pad_right = diff - pad_left  # == pad_left + (diff % 2)
+            x = x[..., pad_left : x.shape[-1] - pad_right]
         return x + y
 
 
@@ -93,7 +110,11 @@ def __init__(self, dim: int = 16, stride: int = 1):
                 dim,
                 kernel_size=2 * stride,
                 stride=stride,
-                padding=math.ceil(stride / 2),
+                # For kernel_size = 2*stride the correct "same-length" padding
+                # is (kernel_size - stride) // 2 = stride // 2.
+                # The previous math.ceil(stride / 2) was off by 1 for odd
+                # strides (e.g. stride=5 gave padding=3 instead of 2).
+                padding=stride // 2,
             ),
         )
 
@@ -209,7 +230,9 @@ def __init__(
         if self.attn_proj:
             self.pre_block = AttnProjection(latent_dim, self.attn_proj_dim, num_heads=8)
 
-        self.sample_rate = sample_rate
+        # Apply canonical weight initialization to all Conv1d layers in the
+        # entire model graph (encoder + BigVGAN decoder). This single pass
+        # uses trunc_normal_(std=0.02) and is the intended final initializer.
         self.apply(init_weights)
 
     def preprocess(self, audio_data, sample_rate):
diff --git a/FL2VA/transformer/config.json b/FL2VA/transformer/config.json
index 94cc6ca..c7dbf0e 100644
--- a/FL2VA/transformer/config.json
+++ b/FL2VA/transformer/config.json
@@ -4,11 +4,15 @@
   "hidden_size": 5376,
   "num_layers": 50,
   "token_refiner_num_layers": 2,
+  "num_refiner_layers": 2,
   "num_attention_heads": 56,
   "attention_head_dim": 128,
   "ffn_hidden_size": 14336,
+  "ffn_dim": 14336,
   "latents_dim": 24,
+  "in_channels": 24,
   "audio_latents_dim": 32,
+  "audio_in_channels": 32,
   "patch_size": [
     1,
     2,
@@ -16,11 +20,15 @@
   ],
   "text_dim": 5120,
   "timestep_input_dim": 256,
+  "freq_dim": 256,
   "time_embed_hidden_size": 5376,
+  "time_embed_hidden_dim": 5376,
   "time_embed_dim": 2688,
   "adaln_out_features": 96768,
   "final_adaln_out_features": 10752,
   "rope_inv_freq_len": 16,
+  "rope_freq_dim": 16,
+  "rope_theta": 10000.0,
   "norm_eps": 1e-05,
   "qk_norm_eps": 1e-05,
   "final_norm_eps": 1e-05
diff --git a/README.md b/README.md
index 9789aa4..9c1403c 100644
--- a/README.md
+++ b/README.md
@@ -215,6 +215,68 @@ hf download MiniMaxAI/MiniMax-H3 --include "model_index.json" "FL2VA/*" "Ref2VA/
 hf download MiniMaxAI/MiniMax-H3 --include "model_index.json" "FL2VA/*" --local-dir MiniMax-H3
 ```
 
+#### Required Model Weights & Download Instructions
+
+> **Important:** This repository contains only code and configuration files.
+> The model weights (`.safetensors` files) are hosted separately on Hugging Face
+> and must be downloaded before running inference.
+
+**Step 1 — Install dependencies**
+
+```bash
+pip install -r requirements.txt
+# For SGLang / vLLM, also install the respective package as documented in the
+# framework sections below.
+```
+
+**Step 2 — Download the weights**
+
+```bash
+# Install the Hugging Face CLI if needed:
+pip install huggingface_hub
+
+# Download FL2VA checkpoint (T2VA + FL2VA tasks):
+hf download MiniMaxAI/MiniMax-H3 \
+  --include "model_index.json" "modular_model_index.json" \
+  --include "FL2VA/**" \
+  --include "vae/**" "audio_vae/**" "scheduler/**" "audio_scheduler/**" \
+  --local-dir ./MiniMax-H3
+
+# Download Ref2VA checkpoint additionally:
+hf download MiniMaxAI/MiniMax-H3 \
+  --include "Ref2VA/**" \
+  --local-dir ./MiniMax-H3
+```
+
+**Step 3 — Verify the download**
+
+After downloading, confirm the key weight files exist:
+
+```bash
+# FL2VA transformer weights (largest file — ~67 GB total for all shards)
+ls MiniMax-H3/FL2VA/transformer/model-*.safetensors
+
+# Audio VAE weights
+ls MiniMax-H3/FL2VA/audio_vae/*.safetensors
+
+# Visual VAE weights
+ls MiniMax-H3/FL2VA/vae/*.safetensors
+```
+
+If any files are missing, re-run the `hf download` command above. The `hf` CLI
+is resumable — it will skip files that are already complete.
+
+**Step 4 — Run inference**
+
+Once weights are downloaded, follow the SGLang, vLLM, or diffusers workflows
+documented in the sections below. For a quick smoke-test using the reproducible
+768p scripts, set your `SGLANG_DEPLOYMENT_URL` and run:
+
+```bash
+bash scripts/readme/reproducible-768p-t2va-request.sh
+```
+
+
 diffusers users do not need a manual download: `ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3")` fetches exactly the components it needs. See the [diffusers documentation](https://github.com/huggingface/diffusers/blob/minimax-h3/docs/source/en/api/pipelines/minimax_h3.md) for loading recipes.
 
 We recommend the following inference frameworks to serve the model:
diff --git a/Ref2VA/audio_vae/__pycache__/dac_alias_free_filter.cpython-314.pyc b/Ref2VA/audio_vae/__pycache__/dac_alias_free_filter.cpython-314.pyc
new file mode 100644
index 0000000..be6ff2d
Binary files /dev/null and b/Ref2VA/audio_vae/__pycache__/dac_alias_free_filter.cpython-314.pyc differ
diff --git a/Ref2VA/audio_vae/__pycache__/dac_alias_free_resample.cpython-314.pyc b/Ref2VA/audio_vae/__pycache__/dac_alias_free_resample.cpython-314.pyc
new file mode 100644
index 0000000..8eaf8c9
Binary files /dev/null and b/Ref2VA/audio_vae/__pycache__/dac_alias_free_resample.cpython-314.pyc differ
diff --git a/Ref2VA/audio_vae/__pycache__/dac_attn_proj.cpython-314.pyc b/Ref2VA/audio_vae/__pycache__/dac_attn_proj.cpython-314.pyc
new file mode 100644
index 0000000..37c7675
Binary files /dev/null and b/Ref2VA/audio_vae/__pycache__/dac_attn_proj.cpython-314.pyc differ
diff --git a/Ref2VA/audio_vae/__pycache__/dac_audio_vae.cpython-314.pyc b/Ref2VA/audio_vae/__pycache__/dac_audio_vae.cpython-314.pyc
new file mode 100644
index 0000000..c3b9894
Binary files /dev/null and b/Ref2VA/audio_vae/__pycache__/dac_audio_vae.cpython-314.pyc differ
diff --git a/Ref2VA/audio_vae/dac_alias_free_filter.py b/Ref2VA/audio_vae/dac_alias_free_filter.py
index ad02595..26359fc 100644
--- a/Ref2VA/audio_vae/dac_alias_free_filter.py
+++ b/Ref2VA/audio_vae/dac_alias_free_filter.py
@@ -49,11 +49,11 @@ def kaiser_sinc_filter1d(cutoff, half_width, kernel_size):  # return filter [1,1
         filter_ = torch.zeros_like(time)
     else:
         filter_ = 2 * cutoff * window * sinc(2 * cutoff * time)
-        """
-        Normalize filter to have sum = 1, otherwise we will have a small leakage of the constant component in the input signal.
-        """
+        # Normalize filter to have sum = 1, otherwise we will have a small
+        # leakage of the constant component in the input signal.
         filter_ /= filter_.sum()
-        filter = filter_.view(1, 1, kernel_size)
+    # Always reshape into [1, 1, kernel_size] so the return is never undefined.
+    filter = filter_.view(1, 1, kernel_size)
 
     return filter
 
diff --git a/Ref2VA/audio_vae/dac_attn_proj.py b/Ref2VA/audio_vae/dac_attn_proj.py
index d3b9b1f..6f1cd2e 100644
--- a/Ref2VA/audio_vae/dac_attn_proj.py
+++ b/Ref2VA/audio_vae/dac_attn_proj.py
@@ -29,15 +29,19 @@ class CausalAttention(nn.Module):
     def __init__(self, in_dim, out_dim, num_heads):
         super().__init__()
         if in_dim > out_dim:
-            # assert in_dim // num_heads == out_dim
+            # Projection compresses in_dim → out_dim after attention.
+            # QKV operates in in_dim space so head_dim = in_dim // num_heads.
             self.head_dim = in_dim // num_heads
+            self.qkv_out_dim = in_dim
             self.qkv = nn.Linear(in_dim, in_dim * 3, bias=False)
             self.q_bias = nn.Parameter(torch.zeros(in_dim))
             self.v_bias = nn.Parameter(torch.zeros(in_dim))
             self.register_buffer("zero_k_bias", torch.zeros(in_dim))
         else:
-            # assert out_dim // num_heads == in_dim
+            # Projection expands (or keeps) in_dim → out_dim after attention.
+            # QKV operates in out_dim space so head_dim = out_dim // num_heads.
             self.head_dim = out_dim // num_heads
+            self.qkv_out_dim = out_dim
             self.qkv = nn.Linear(in_dim, out_dim * 3, bias=False)
             self.q_bias = nn.Parameter(torch.zeros(out_dim))
             self.v_bias = nn.Parameter(torch.zeros(out_dim))
@@ -47,21 +51,27 @@ def __init__(self, in_dim, out_dim, num_heads):
         self.out_dim = out_dim
         self.num_heads = num_heads
         self.scale = self.head_dim**-0.5
-        self.proj = nn.Linear(out_dim, out_dim)
+        # Final linear maps from the QKV output space to out_dim.
+        self.proj = nn.Linear(self.qkv_out_dim, out_dim)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         B, N, C = x.shape
-        qkv = F.linear(input=x, weight=self.qkv.weight, bias=torch.cat((self.q_bias, self.zero_k_bias, self.v_bias)))
-        q, k, v = qkv.reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4).unbind(0)
+        qkv = F.linear(
+            input=x,
+            weight=self.qkv.weight,
+            bias=torch.cat((self.q_bias, self.zero_k_bias, self.v_bias)),
+        )
+        q, k, v = (
+            qkv.reshape(B, N, 3, self.num_heads, self.head_dim)
+            .permute(2, 0, 3, 1, 4)
+            .unbind(0)
+        )
 
         x = scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=True)
 
-        if self.in_dim > self.out_dim:
-            x = torch.mean(x, dim=1)
-            if self.in_dim // self.num_heads != self.out_dim:
-                x = nn.functional.adaptive_avg_pool1d(x, self.out_dim)
-        else:
-            x = x.transpose(1, 2).reshape(B, N, -1)
+        # Always reshape to [B, N, qkv_out_dim] so self.proj receives the
+        # correct input dimension regardless of whether in_dim > out_dim or not.
+        x = x.transpose(1, 2).reshape(B, N, self.qkv_out_dim)
         x = self.proj(x)
         return x
 
diff --git a/Ref2VA/audio_vae/dac_audio_vae.py b/Ref2VA/audio_vae/dac_audio_vae.py
index aebdfb7..d2a9b55 100644
--- a/Ref2VA/audio_vae/dac_audio_vae.py
+++ b/Ref2VA/audio_vae/dac_audio_vae.py
@@ -43,6 +43,17 @@ def forward(self, x):
 def init_weights(m):
     """Initialize Conv1d layers, including those wrapped with ``weight_norm``.
 
+    This is the canonical initializer for ``DacAudioVAE`` (encoder + decoder).
+    It uses ``trunc_normal_`` with ``std=0.02`` and explicitly zeroes biases,
+    matching the intended training configuration for this model family.
+
+    Note: ``dac_bigvgan.py`` / ``dac_utils.py`` also contain an ``init_weights``
+    helper (``std=0.01``, ``normal_``) used by BigVGAN's own module-level
+    ``apply()`` calls.  Because ``DacAudioVAE.__init__`` calls
+    ``self.apply(init_weights)`` *after* all sub-modules (including BigVGAN)
+    have been constructed, this function is the one that takes final effect on
+    every ``nn.Conv1d`` in the whole model graph.
+
     The encoder (``WNConv1d``) and decoder (BigVGAN) both wrap ``nn.Conv1d``
     with ``torch.nn.utils.parametrizations.weight_norm``. Under that
     parametrization ``m.weight`` is computed on access from
@@ -74,9 +85,15 @@ def __init__(self, dim: int = 16, dilation: int = 1):
 
     def forward(self, x):
         y = self.block(x)
-        pad = (x.shape[-1] - y.shape[-1]) // 2
-        if pad > 0:
-            x = x[..., pad:-pad]
+        diff = x.shape[-1] - y.shape[-1]
+        if diff > 0:
+            # Crop the residual path to match the convolved output length.
+            # Use explicit left/right amounts to handle both even and odd
+            # differences safely (avoids the x[..., pad:-pad] pattern which
+            # silently drops an extra sample when diff is odd).
+            pad_left = diff // 2
+            pad_right = diff - pad_left  # == pad_left + (diff % 2)
+            x = x[..., pad_left : x.shape[-1] - pad_right]
         return x + y
 
 
@@ -93,7 +110,11 @@ def __init__(self, dim: int = 16, stride: int = 1):
                 dim,
                 kernel_size=2 * stride,
                 stride=stride,
-                padding=math.ceil(stride / 2),
+                # For kernel_size = 2*stride the correct "same-length" padding
+                # is (kernel_size - stride) // 2 = stride // 2.
+                # The previous math.ceil(stride / 2) was off by 1 for odd
+                # strides (e.g. stride=5 gave padding=3 instead of 2).
+                padding=stride // 2,
             ),
         )
 
@@ -209,7 +230,9 @@ def __init__(
         if self.attn_proj:
             self.pre_block = AttnProjection(latent_dim, self.attn_proj_dim, num_heads=8)
 
-        self.sample_rate = sample_rate
+        # Apply canonical weight initialization to all Conv1d layers in the
+        # entire model graph (encoder + BigVGAN decoder). This single pass
+        # uses trunc_normal_(std=0.02) and is the intended final initializer.
         self.apply(init_weights)
 
     def preprocess(self, audio_data, sample_rate):
diff --git a/Ref2VA/transformer/config.json b/Ref2VA/transformer/config.json
index 94cc6ca..c7dbf0e 100644
--- a/Ref2VA/transformer/config.json
+++ b/Ref2VA/transformer/config.json
@@ -4,11 +4,15 @@
   "hidden_size": 5376,
   "num_layers": 50,
   "token_refiner_num_layers": 2,
+  "num_refiner_layers": 2,
   "num_attention_heads": 56,
   "attention_head_dim": 128,
   "ffn_hidden_size": 14336,
+  "ffn_dim": 14336,
   "latents_dim": 24,
+  "in_channels": 24,
   "audio_latents_dim": 32,
+  "audio_in_channels": 32,
   "patch_size": [
     1,
     2,
@@ -16,11 +20,15 @@
   ],
   "text_dim": 5120,
   "timestep_input_dim": 256,
+  "freq_dim": 256,
   "time_embed_hidden_size": 5376,
+  "time_embed_hidden_dim": 5376,
   "time_embed_dim": 2688,
   "adaln_out_features": 96768,
   "final_adaln_out_features": 10752,
   "rope_inv_freq_len": 16,
+  "rope_freq_dim": 16,
+  "rope_theta": 10000.0,
   "norm_eps": 1e-05,
   "qk_norm_eps": 1e-05,
   "final_norm_eps": 1e-05
diff --git a/audio_vae/config.json b/audio_vae/config.json
index 15df6fe..5e90c79 100644
--- a/audio_vae/config.json
+++ b/audio_vae/config.json
@@ -53,6 +53,7 @@
       5
     ]
   ],
+  "sample_rate": 32000,
   "sampling_rate": 32000,
   "latents_mean": [
     -0.020211687488382354,
diff --git a/requirements.txt b/requirements.txt
index 7f5c20c..8ab7b6e 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -47,3 +47,8 @@ soundfile>=0.12.0
 imageio>=2.34.0
 imageio-ffmpeg>=0.5.0
 av>=11.0.0
+
+# PyYAML is required at runtime by the audio VAE loader (minimax_h3_audio_vae.py)
+# to parse the model config YAML file. Without it, loading the audio VAE raises
+# ImportError("MiniMax H3 audio VAE requires PyYAML.").
+PyYAML>=6.0
diff --git a/scripts/readme/full-2k-i2va-h3-base.sh b/scripts/readme/full-2k-i2va-h3-base.sh
index 37a53d1..cf2e8d4 100755
--- a/scripts/readme/full-2k-i2va-h3-base.sh
+++ b/scripts/readme/full-2k-i2va-h3-base.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create the H3-Base request with the expanded prompt and capture the video ID.
@@ -23,7 +23,7 @@ video_id=$(
   },
   "seed": 0
 }' |
-    curl --silent --show-error \
+    curl --fail-with-body --silent --show-error \
       --request POST \
       --url "$SGLANG_DEPLOYMENT_URL/v1/videos" \
       --header 'Content-Type: application/json' \
@@ -31,12 +31,12 @@ video_id=$(
     jq -er '.id'
 )
 # Query the generation status.
-curl --silent --show-error \
+curl --fail-with-body --silent --show-error \
   --request GET \
   --url "$SGLANG_DEPLOYMENT_URL/v1/videos/$video_id" |
   jq '{status}'
 # Download the local H3-Base MP4 after its status becomes completed.
-curl --silent --show-error \
+curl --fail-with-body --silent --show-error \
   --request GET \
   --url "$SGLANG_DEPLOYMENT_URL/v1/videos/$video_id/content" \
   --output i2va.mp4
diff --git a/scripts/readme/full-2k-i2va-h3-context-ir.sh b/scripts/readme/full-2k-i2va-h3-context-ir.sh
index c8180c4..1cec904 100755
--- a/scripts/readme/full-2k-i2va-h3-context-ir.sh
+++ b/scripts/readme/full-2k-i2va-h3-context-ir.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create the prompt-expansion task and capture its runtime ID.
diff --git a/scripts/readme/full-2k-i2va-h3-regenerate-2k.sh b/scripts/readme/full-2k-i2va-h3-regenerate-2k.sh
index 45f4b21..6f14127 100755
--- a/scripts/readme/full-2k-i2va-h3-regenerate-2k.sh
+++ b/scripts/readme/full-2k-i2va-h3-regenerate-2k.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 H3_BASE_VIDEO='./i2va.mp4'
diff --git a/scripts/readme/full-2k-i2va-reference-2k-result-by-directly-calling-open-platform-api.sh b/scripts/readme/full-2k-i2va-reference-2k-result-by-directly-calling-open-platform-api.sh
index c04741b..0a9e651 100755
--- a/scripts/readme/full-2k-i2va-reference-2k-result-by-directly-calling-open-platform-api.sh
+++ b/scripts/readme/full-2k-i2va-reference-2k-result-by-directly-calling-open-platform-api.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create an 8-second 2K FL2VA video directly and capture its runtime task ID.
diff --git a/scripts/readme/full-2k-i2va-reference-768p-result-by-directly-calling-open-platform-api.sh b/scripts/readme/full-2k-i2va-reference-768p-result-by-directly-calling-open-platform-api.sh
index c21ee2d..44a2637 100755
--- a/scripts/readme/full-2k-i2va-reference-768p-result-by-directly-calling-open-platform-api.sh
+++ b/scripts/readme/full-2k-i2va-reference-768p-result-by-directly-calling-open-platform-api.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create an 8-second 768P FL2VA video directly and capture its runtime task ID.
diff --git a/scripts/readme/full-2k-ref2va-h3-api-2k-in-open-platform-for-reference.sh b/scripts/readme/full-2k-ref2va-h3-api-2k-in-open-platform-for-reference.sh
index 19b7c65..566cf39 100755
--- a/scripts/readme/full-2k-ref2va-h3-api-2k-in-open-platform-for-reference.sh
+++ b/scripts/readme/full-2k-ref2va-h3-api-2k-in-open-platform-for-reference.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create a 5-second 2K Ref2VA video directly and capture its runtime task ID.
diff --git a/scripts/readme/full-2k-ref2va-h3-base.sh b/scripts/readme/full-2k-ref2va-h3-base.sh
index 6b6f034..ba4b560 100755
--- a/scripts/readme/full-2k-ref2va-h3-base.sh
+++ b/scripts/readme/full-2k-ref2va-h3-base.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create the H3-Base request with the expanded prompt and capture the video ID.
@@ -27,7 +27,7 @@ video_id=$(
   },
   "seed": 0
 }' |
-    curl --silent --show-error \
+    curl --fail-with-body --silent --show-error \
       --request POST \
       --url "$SGLANG_DEPLOYMENT_URL/v1/videos" \
       --header 'Content-Type: application/json' \
@@ -35,12 +35,12 @@ video_id=$(
     jq -er '.id'
 )
 # Query the generation status.
-curl --silent --show-error \
+curl --fail-with-body --silent --show-error \
   --request GET \
   --url "$SGLANG_DEPLOYMENT_URL/v1/videos/$video_id" |
   jq '{status}'
 # Download the local H3-Base MP4 after its status becomes completed.
-curl --silent --show-error \
+curl --fail-with-body --silent --show-error \
   --request GET \
   --url "$SGLANG_DEPLOYMENT_URL/v1/videos/$video_id/content" \
   --output r2va.mp4
diff --git a/scripts/readme/full-2k-ref2va-h3-context-ir.sh b/scripts/readme/full-2k-ref2va-h3-context-ir.sh
index 6195577..af699e7 100755
--- a/scripts/readme/full-2k-ref2va-h3-context-ir.sh
+++ b/scripts/readme/full-2k-ref2va-h3-context-ir.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create the prompt-expansion task and capture its runtime ID.
diff --git a/scripts/readme/full-2k-ref2va-reference-2k-result-by-directly-calling-open-platform-api.sh b/scripts/readme/full-2k-ref2va-reference-2k-result-by-directly-calling-open-platform-api.sh
index 2ac2aab..ce4e948 100755
--- a/scripts/readme/full-2k-ref2va-reference-2k-result-by-directly-calling-open-platform-api.sh
+++ b/scripts/readme/full-2k-ref2va-reference-2k-result-by-directly-calling-open-platform-api.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 H3_BASE_VIDEO='./r2va.mp4'
diff --git a/scripts/readme/full-2k-ref2va-reference-768p-result-by-directly-calling-open-platform-api.sh b/scripts/readme/full-2k-ref2va-reference-768p-result-by-directly-calling-open-platform-api.sh
index 038b5df..b99f4cd 100755
--- a/scripts/readme/full-2k-ref2va-reference-768p-result-by-directly-calling-open-platform-api.sh
+++ b/scripts/readme/full-2k-ref2va-reference-768p-result-by-directly-calling-open-platform-api.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create a 5-second 768P Ref2VA video directly and capture its runtime task ID.
diff --git a/scripts/readme/full-2k-t2va-h3-base.sh b/scripts/readme/full-2k-t2va-h3-base.sh
index cbdd623..2893bfd 100755
--- a/scripts/readme/full-2k-t2va-h3-base.sh
+++ b/scripts/readme/full-2k-t2va-h3-base.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create the H3-Base request with the expanded prompt and capture the video ID.
@@ -16,7 +16,7 @@ video_id=$(
   },
   "seed": 0
 }' |
-    curl --silent --show-error \
+    curl --fail-with-body --silent --show-error \
       --request POST \
       --url "$SGLANG_DEPLOYMENT_URL/v1/videos" \
       --header 'Content-Type: application/json' \
@@ -24,12 +24,12 @@ video_id=$(
     jq -er '.id'
 )
 # Query the generation status.
-curl --silent --show-error \
+curl --fail-with-body --silent --show-error \
   --request GET \
   --url "$SGLANG_DEPLOYMENT_URL/v1/videos/$video_id" |
   jq '{status}'
 # Download the local H3-Base MP4 after its status becomes completed.
-curl --silent --show-error \
+curl --fail-with-body --silent --show-error \
   --request GET \
   --url "$SGLANG_DEPLOYMENT_URL/v1/videos/$video_id/content" \
   --output t2va.mp4
diff --git a/scripts/readme/full-2k-t2va-h3-context-ir.sh b/scripts/readme/full-2k-t2va-h3-context-ir.sh
index 6afd3b4..18b88ec 100755
--- a/scripts/readme/full-2k-t2va-h3-context-ir.sh
+++ b/scripts/readme/full-2k-t2va-h3-context-ir.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create the prompt-expansion task and capture its runtime ID.
diff --git a/scripts/readme/full-2k-t2va-h3-regenerate-2k.sh b/scripts/readme/full-2k-t2va-h3-regenerate-2k.sh
index 1d67780..a2649da 100755
--- a/scripts/readme/full-2k-t2va-h3-regenerate-2k.sh
+++ b/scripts/readme/full-2k-t2va-h3-regenerate-2k.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 H3_BASE_VIDEO='./t2va.mp4'
diff --git a/scripts/readme/full-2k-t2va-reference-2k-result-by-directly-calling-open-platform-api.sh b/scripts/readme/full-2k-t2va-reference-2k-result-by-directly-calling-open-platform-api.sh
index 669e10e..cddfc3c 100755
--- a/scripts/readme/full-2k-t2va-reference-2k-result-by-directly-calling-open-platform-api.sh
+++ b/scripts/readme/full-2k-t2va-reference-2k-result-by-directly-calling-open-platform-api.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create a 10-second 2K video directly and capture its runtime task ID.
diff --git a/scripts/readme/full-2k-t2va-reference-768p-result-by-directly-calling-open-platform-api.sh b/scripts/readme/full-2k-t2va-reference-768p-result-by-directly-calling-open-platform-api.sh
index b765671..10eb7db 100755
--- a/scripts/readme/full-2k-t2va-reference-768p-result-by-directly-calling-open-platform-api.sh
+++ b/scripts/readme/full-2k-t2va-reference-768p-result-by-directly-calling-open-platform-api.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Create a 10-second 768P video directly and capture its runtime task ID.
diff --git a/scripts/readme/reproducible-768p-fl2va-request.sh b/scripts/readme/reproducible-768p-fl2va-request.sh
index 1efa6fc..486fd80 100755
--- a/scripts/readme/reproducible-768p-fl2va-request.sh
+++ b/scripts/readme/reproducible-768p-fl2va-request.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Submit the FL2VA request with the complete H3-Context-IR prompt.
diff --git a/scripts/readme/reproducible-768p-ref2va-request.sh b/scripts/readme/reproducible-768p-ref2va-request.sh
index 5f13b14..3e5598e 100755
--- a/scripts/readme/reproducible-768p-ref2va-request.sh
+++ b/scripts/readme/reproducible-768p-ref2va-request.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Submit the Ref2VA request with the complete H3-Context-IR prompt.
diff --git a/scripts/readme/reproducible-768p-t2va-request.sh b/scripts/readme/reproducible-768p-t2va-request.sh
index b23938b..c3e30cd 100755
--- a/scripts/readme/reproducible-768p-t2va-request.sh
+++ b/scripts/readme/reproducible-768p-t2va-request.sh
@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 # Submit the T2VA request with the complete H3-Context-IR prompt.
diff --git a/tests/__pycache__/test_regressions.cpython-314-pytest-9.1.1.pyc b/tests/__pycache__/test_regressions.cpython-314-pytest-9.1.1.pyc
new file mode 100644
index 0000000..f7275ea
Binary files /dev/null and b/tests/__pycache__/test_regressions.cpython-314-pytest-9.1.1.pyc differ
diff --git a/tests/test_regressions.py b/tests/test_regressions.py
new file mode 100644
index 0000000..828d064
--- /dev/null
+++ b/tests/test_regressions.py
@@ -0,0 +1,338 @@
+"""
+MiniMax-H3 Regression Tests — Fixed harness
+==============================================
+Uses sys.path injection instead of spec_from_file_location so relative
+imports inside the audio_vae package resolve correctly.
+
+Run with:
+    cd d:\\minimax\\MiniMax-H3
+    python -m pytest tests/test_regressions.py -v --tb=short
+"""
+
+import json
+import sys
+import importlib
+from pathlib import Path
+
+import pytest
+import torch
+import torch.nn as nn
+
+PROJECT_ROOT = Path(__file__).parent.parent
+
+
+def _load_audio_vae_package(task: str):
+    """
+    Add the task directory to sys.path so relative imports work,
+    then import the modules as a flat namespace.
+    Returns the dac_audio_vae module for that task.
+    """
+    pkg_dir = str(PROJECT_ROOT / task / "audio_vae")
+    if pkg_dir not in sys.path:
+        sys.path.insert(0, pkg_dir)
+    # Force fresh load by removing any cached versions
+    mods_to_reload = [
+        "dac_alias_free_filter", "dac_alias_free_resample",
+        "dac_alias_free_act", "dac_activations", "dac_utils",
+        "dac_attn_proj", "dac_bigvgan", "dac_audio_vae",
+    ]
+    for m in mods_to_reload:
+        sys.modules.pop(m, None)
+    import dac_audio_vae
+    return dac_audio_vae
+
+
+def _load_filter_module(task: str):
+    pkg_dir = str(PROJECT_ROOT / task / "audio_vae")
+    if pkg_dir not in sys.path:
+        sys.path.insert(0, pkg_dir)
+    sys.modules.pop("dac_alias_free_filter", None)
+    import dac_alias_free_filter
+    return dac_alias_free_filter
+
+
+def _load_attn_proj(task: str):
+    pkg_dir = str(PROJECT_ROOT / task / "audio_vae")
+    if pkg_dir not in sys.path:
+        sys.path.insert(0, pkg_dir)
+    sys.modules.pop("dac_attn_proj", None)
+    import dac_attn_proj
+    return dac_attn_proj
+
+
+# ─── Bug 1 ────────────────────────────────────────────────────────────────────
+
+class TestBug1KaiserSincFilter:
+
+    @pytest.fixture(params=["FL2VA", "Ref2VA"])
+    def fmod(self, request):
+        return _load_filter_module(request.param)
+
+    def test_cutoff_zero_no_nameerror(self, fmod):
+        r = fmod.kaiser_sinc_filter1d(cutoff=0, half_width=0.6, kernel_size=12)
+        assert r is not None
+
+    def test_cutoff_zero_shape(self, fmod):
+        r = fmod.kaiser_sinc_filter1d(cutoff=0, half_width=0.6, kernel_size=12)
+        assert r.shape == (1, 1, 12)
+
+    def test_cutoff_zero_all_zeros(self, fmod):
+        r = fmod.kaiser_sinc_filter1d(cutoff=0, half_width=0.6, kernel_size=12)
+        assert torch.all(r == 0)
+
+    def test_cutoff_positive_shape(self, fmod):
+        for ks in [8, 12, 24]:
+            r = fmod.kaiser_sinc_filter1d(cutoff=0.3, half_width=0.3, kernel_size=ks)
+            assert r.shape == (1, 1, ks)
+
+    def test_cutoff_positive_normalized(self, fmod):
+        r = fmod.kaiser_sinc_filter1d(cutoff=0.5, half_width=0.6, kernel_size=12)
+        assert abs(r.sum().item() - 1.0) < 1e-5
+
+
+# ─── Bug 2 ────────────────────────────────────────────────────────────────────
+
+class TestBug2DuplicateSampleRate:
+
+    def test_source_has_exactly_one_assignment(self):
+        for task in ("FL2VA", "Ref2VA"):
+            src = (PROJECT_ROOT / task / "audio_vae" / "dac_audio_vae.py").read_text(encoding="utf-8")
+            count = src.count("self.sample_rate = sample_rate")
+            assert count == 1, f"{task}: expected 1 assignment, found {count}"
+
+    @pytest.mark.parametrize("task", ["FL2VA", "Ref2VA"])
+    def test_sample_rate_attribute_correct(self, task):
+        vae_mod = _load_audio_vae_package(task)
+        for sr in (16000, 32000):
+            vae = vae_mod.DacAudioVAE(sample_rate=sr)
+            assert vae.sample_rate == sr
+
+
+# ─── Bug 3 ────────────────────────────────────────────────────────────────────
+
+class TestBug3InitWeightsConsistency:
+
+    def test_uses_trunc_normal_std02(self):
+        for task in ("FL2VA", "Ref2VA"):
+            src = (PROJECT_ROOT / task / "audio_vae" / "dac_audio_vae.py").read_text(encoding="utf-8")
+            assert "trunc_normal_" in src
+            assert "std=0.02" in src
+            assert "m.bias" in src
+
+    def test_apply_called_once_in_live_code(self):
+        for task in ("FL2VA", "Ref2VA"):
+            src = (PROJECT_ROOT / task / "audio_vae" / "dac_audio_vae.py").read_text(encoding="utf-8")
+            # Count only real code lines (not comment or docstring lines)
+            code_apply = sum(
+                1 for l in src.splitlines()
+                if "self.apply(init_weights)" in l
+                and not l.strip().startswith("#")
+                and not l.strip().startswith("`")
+                and not l.strip().startswith("\"\"\"")
+            )
+            assert code_apply == 1, f"{task}: {code_apply} apply calls in live code"
+
+    @pytest.mark.parametrize("task", ["FL2VA", "Ref2VA"])
+    def test_conv_biases_zeroed_after_build(self, task):
+        vae_mod = _load_audio_vae_package(task)
+        vae = vae_mod.DacAudioVAE(sample_rate=32000)
+        checked = 0
+        for name, param in vae.named_parameters():
+            if "bias" in name and param is not None and param.numel() > 0:
+                assert torch.all(param.data == 0), f"{task}: {name} not zeroed"
+                checked += 1
+        assert checked > 0, "No biases found to check"
+
+
+# ─── Bug 4 ────────────────────────────────────────────────────────────────────
+
+class TestBug4CausalAttentionShape:
+
+    @pytest.fixture(params=["FL2VA", "Ref2VA"])
+    def amod(self, request):
+        return _load_attn_proj(request.param)
+
+    def _fwd(self, amod, in_d, out_d, nh, B=2, N=8):
+        ca = amod.CausalAttention(in_d, out_d, nh)
+        ca.eval()
+        with torch.no_grad():
+            return ca(torch.randn(B, N, in_d)).shape
+
+    def test_in_eq_out(self, amod):
+        assert self._fwd(amod, 64, 64, 8) == (2, 8, 64)
+
+    def test_in_less_than_out(self, amod):
+        assert self._fwd(amod, 32, 64, 8) == (2, 8, 64)
+
+    def test_in_greater_than_out(self, amod):
+        """This was the broken case — Bug 4."""
+        assert self._fwd(amod, 64, 32, 8) == (2, 8, 32)
+
+    def test_large_batch(self, amod):
+        assert self._fwd(amod, 64, 32, 8, B=4, N=16) == (4, 16, 32)
+
+    def test_attn_projection_end_to_end(self, amod):
+        proj = amod.AttnProjection(in_dim=64, out_dim=32, num_heads=8)
+        proj.eval()
+        with torch.no_grad():
+            out = proj(torch.randn(2, 10, 64))
+        assert out.shape == (2, 10, 32)
+
+
+# ─── Bug 5 ────────────────────────────────────────────────────────────────────
+
+class TestBug5EncoderBlockPadding:
+
+    def test_source_formula_correct(self):
+        for task in ("FL2VA", "Ref2VA"):
+            src = (PROJECT_ROOT / task / "audio_vae" / "dac_audio_vae.py").read_text(encoding="utf-8")
+            # New formula present
+            assert "padding=stride // 2" in src, f"{task}: new formula missing"
+            # Old formula must NOT appear in live code (only in comments)
+            old_in_live = any(
+                "padding=math.ceil(stride" in l and not l.strip().startswith("#")
+                for l in src.splitlines()
+            )
+            assert not old_in_live, f"{task}: old math.ceil formula still in live code"
+
+    @pytest.mark.parametrize("stride", [1, 2, 3, 4, 5, 8])
+    def test_output_length(self, stride):
+        vae_mod = _load_audio_vae_package("FL2VA")
+        dim = 32
+        block = vae_mod.EncoderBlock(dim=dim, stride=stride)
+        block.eval()
+        T = stride * 20
+        with torch.no_grad():
+            out = block(torch.randn(1, dim // 2, T))
+        assert out.shape[1] == dim, f"stride={stride}: channels {out.shape[1]}"
+        assert out.shape[2] == T // stride, f"stride={stride}: T {out.shape[2]} != {T//stride}"
+
+
+# ─── Bug 6 ────────────────────────────────────────────────────────────────────
+
+class TestBug6ResidualUnit:
+
+    @pytest.fixture
+    def ru_mod(self):
+        return _load_audio_vae_package("FL2VA")
+
+    def test_source_uses_explicit_crop(self):
+        for task in ("FL2VA", "Ref2VA"):
+            src = (PROJECT_ROOT / task / "audio_vae" / "dac_audio_vae.py").read_text(encoding="utf-8")
+            assert "pad_right = diff - pad_left" in src
+
+    @pytest.mark.parametrize("T", [30, 31, 50, 51, 99, 100, 101])
+    def test_no_crash_various_lengths(self, ru_mod, T):
+        unit = ru_mod.ResidualUnit(dim=16, dilation=1)
+        unit.eval()
+        with torch.no_grad():
+            _ = unit(torch.randn(2, 16, T))
+
+    def test_dilation3_no_crash(self, ru_mod):
+        unit = ru_mod.ResidualUnit(dim=16, dilation=3)
+        unit.eval()
+        for T in [31, 33, 47, 51]:
+            with torch.no_grad():
+                _ = unit(torch.randn(1, 16, T))
+
+    def test_output_is_3d(self, ru_mod):
+        unit = ru_mod.ResidualUnit(dim=16, dilation=1)
+        unit.eval()
+        with torch.no_grad():
+            out = unit(torch.randn(2, 16, 51))
+        assert out.ndim == 3
+
+
+# ─── Bug 7 ────────────────────────────────────────────────────────────────────
+
+class TestBug7TransformerConfig:
+
+    @pytest.mark.parametrize("task", ["FL2VA", "Ref2VA"])
+    def test_ffn_both_keys(self, task):
+        cfg = json.loads((PROJECT_ROOT / task / "transformer" / "config.json").read_text())
+        assert "ffn_hidden_size" in cfg and "ffn_dim" in cfg
+        assert cfg["ffn_hidden_size"] == cfg["ffn_dim"]
+
+    @pytest.mark.parametrize("task", ["FL2VA", "Ref2VA"])
+    def test_rope_both_keys(self, task):
+        cfg = json.loads((PROJECT_ROOT / task / "transformer" / "config.json").read_text())
+        assert "rope_inv_freq_len" in cfg and "rope_freq_dim" in cfg
+        assert cfg["rope_inv_freq_len"] == cfg["rope_freq_dim"]
+
+    @pytest.mark.parametrize("task", ["FL2VA", "Ref2VA"])
+    def test_refiner_both_keys(self, task):
+        cfg = json.loads((PROJECT_ROOT / task / "transformer" / "config.json").read_text())
+        assert "token_refiner_num_layers" in cfg and "num_refiner_layers" in cfg
+        assert cfg["token_refiner_num_layers"] == cfg["num_refiner_layers"]
+
+
+# ─── Bug 8 ────────────────────────────────────────────────────────────────────
+
+class TestBug8DiffusersVersion:
+
+    def test_task_indexes_match(self):
+        fl = json.loads((PROJECT_ROOT / "FL2VA" / "model_index.json").read_text())["_diffusers_version"]
+        ref = json.loads((PROJECT_ROOT / "Ref2VA" / "model_index.json").read_text())["_diffusers_version"]
+        assert fl == ref
+
+    def test_root_indexes_match(self):
+        root = json.loads((PROJECT_ROOT / "model_index.json").read_text())["_diffusers_version"]
+        mod = json.loads((PROJECT_ROOT / "modular_model_index.json").read_text())["_diffusers_version"]
+        assert root == mod
+
+
+# ─── Bug 9 ────────────────────────────────────────────────────────────────────
+
+class TestBug9PyYAML:
+    def test_listed(self):
+        req = (PROJECT_ROOT / "requirements.txt").read_text(encoding="utf-8")
+        assert "PyYAML" in req
+
+
+# ─── Bug 10 ───────────────────────────────────────────────────────────────────
+
+class TestBug10CRLF:
+    def test_no_crlf(self):
+        bad = [f.name for f in (PROJECT_ROOT / "scripts" / "readme").glob("*.sh")
+               if b"\r\n" in f.read_bytes()]
+        assert bad == [], f"CRLF found in: {bad}"
+
+
+# ─── Bug 11 ───────────────────────────────────────────────────────────────────
+
+class TestBug11CurlFlag:
+    @pytest.mark.parametrize("script", [
+        "full-2k-t2va-h3-base.sh",
+        "full-2k-i2va-h3-base.sh",
+        "full-2k-ref2va-h3-base.sh",
+    ])
+    def test_flag_present(self, script):
+        txt = (PROJECT_ROOT / "scripts" / "readme" / script).read_text(encoding="utf-8")
+        assert "--fail-with-body" in txt
+
+
+# ─── Bug 12 ───────────────────────────────────────────────────────────────────
+
+class TestBug12SampleRate:
+    def test_both_keys_present_and_equal(self):
+        cfg = json.loads((PROJECT_ROOT / "audio_vae" / "config.json").read_text())
+        assert "sample_rate" in cfg
+        assert "sampling_rate" in cfg
+        assert cfg["sample_rate"] == cfg["sampling_rate"]
+
+    def test_fl2va_config_has_sample_rate(self):
+        cfg = json.loads((PROJECT_ROOT / "FL2VA" / "audio_vae" / "config.json").read_text())
+        assert "sample_rate" in cfg
+
+
+# ─── Bug 13 ───────────────────────────────────────────────────────────────────
+
+class TestBug13README:
+    def test_download_steps_present(self):
+        readme = (PROJECT_ROOT / "README.md").read_text(encoding="utf-8", errors="replace")
+        assert "Step 1" in readme and "Step 2" in readme and "Step 3" in readme
+        assert "hf download MiniMaxAI/MiniMax-H3" in readme
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__, "-v", "--tb=short"]))
diff --git a/tests/verify_fixes.py b/tests/verify_fixes.py
new file mode 100644
index 0000000..061107e
--- /dev/null
+++ b/tests/verify_fixes.py
@@ -0,0 +1,132 @@
+import json, sys, pathlib
+
+ROOT = pathlib.Path(r"d:\minimax\MiniMax-H3")
+results = []
+
+def check(num, name, condition, detail=""):
+    icon = "PASS" if condition else "FAIL"
+    results.append((num, name, icon, detail))
+    tag = "OK" if condition else "!!"
+    print(f"  [{tag}] Bug {num}: {name}" + (f" -> {detail}" if detail else ""))
+
+# BUG 1
+for task in ("FL2VA", "Ref2VA"):
+    src = (ROOT / task / "audio_vae" / "dac_alias_free_filter.py").read_text(encoding="utf-8")
+    lines = src.splitlines()
+    fix_ok  = any("    filter = filter_.view(1, 1, kernel_size)" in l
+                  and not l.startswith("        ") for l in lines)
+    still_buggy = any("        filter = filter_.view" in l for l in lines)
+    check(1, f"NameError cutoff==0 [{task}]", fix_ok and not still_buggy)
+
+# BUG 2
+for task in ("FL2VA", "Ref2VA"):
+    src = (ROOT / task / "audio_vae" / "dac_audio_vae.py").read_text(encoding="utf-8")
+    count = src.count("self.sample_rate = sample_rate")
+    check(2, f"Duplicate sample_rate [{task}]", count == 1, f"{count} assignment(s)")
+
+# BUG 3
+for task in ("FL2VA", "Ref2VA"):
+    src = (ROOT / task / "audio_vae" / "dac_audio_vae.py").read_text(encoding="utf-8")
+    has_trunc = "trunc_normal_" in src
+    has_std   = "std=0.02" in src
+    has_bias  = "m.bias" in src
+    apply_calls = sum(1 for l in src.splitlines()
+                      if "self.apply(init_weights)" in l
+                      and not l.strip().startswith("#")
+                      and not l.strip().startswith("``"))
+    check(3, f"Conflicting init_weights [{task}]",
+          has_trunc and has_std and has_bias and apply_calls == 1,
+          f"trunc={has_trunc} std02={has_std} bias={has_bias} apply={apply_calls}")
+
+# BUG 4
+for task in ("FL2VA", "Ref2VA"):
+    src = (ROOT / task / "audio_vae" / "dac_attn_proj.py").read_text(encoding="utf-8")
+    has_qkv_out     = "self.qkv_out_dim" in src
+    old_mean_gone   = "torch.mean(x, dim=1)" not in src
+    uniform_reshape = "x.transpose(1, 2).reshape(B, N, self.qkv_out_dim)" in src
+    check(4, f"CausalAttention shape [{task}]",
+          has_qkv_out and old_mean_gone and uniform_reshape)
+
+# BUG 5
+for task in ("FL2VA", "Ref2VA"):
+    src = (ROOT / task / "audio_vae" / "dac_audio_vae.py").read_text(encoding="utf-8")
+    new_ok = "padding=stride // 2" in src
+    old_in_code = any("padding=math.ceil(stride" in l and not l.strip().startswith("#")
+                      for l in src.splitlines())
+    check(5, f"EncoderBlock padding [{task}]", new_ok and not old_in_code)
+
+# BUG 6
+for task in ("FL2VA", "Ref2VA"):
+    src = (ROOT / task / "audio_vae" / "dac_audio_vae.py").read_text(encoding="utf-8")
+    has_explicit = "pad_right = diff - pad_left" in src
+    old_gone     = not any("x[..., pad:-pad]" in l and not l.strip().startswith("#")
+                           for l in src.splitlines())
+    check(6, f"ResidualUnit odd diff [{task}]", has_explicit and old_gone)
+
+# BUG 7
+for task in ("FL2VA", "Ref2VA"):
+    cfg = json.loads((ROOT / task / "transformer" / "config.json").read_text(encoding="utf-8"))
+    pairs = [("ffn_hidden_size", "ffn_dim"),
+             ("rope_inv_freq_len", "rope_freq_dim"),
+             ("token_refiner_num_layers", "num_refiner_layers")]
+    ok = all(a in cfg and b in cfg and cfg[a] == cfg[b] for a, b in pairs)
+    check(7, f"Transformer config consistency [{task}]", ok)
+
+# BUG 8
+fl2va_v = json.loads((ROOT / "FL2VA" / "model_index.json").read_text())["_diffusers_version"]
+ref2va_v = json.loads((ROOT / "Ref2VA" / "model_index.json").read_text())["_diffusers_version"]
+root_v   = json.loads((ROOT / "model_index.json").read_text())["_diffusers_version"]
+mod_v    = json.loads((ROOT / "modular_model_index.json").read_text())["_diffusers_version"]
+check(8, "Diffusers version consistency",
+      fl2va_v == ref2va_v and root_v == mod_v,
+      f"FL2VA/Ref2VA={fl2va_v} | Root/Modular={root_v}")
+
+# BUG 9
+req = (ROOT / "requirements.txt").read_text(encoding="utf-8")
+check(9, "PyYAML in requirements.txt", "PyYAML" in req)
+
+# BUG 10
+crlf_files = [f.name for f in (ROOT / "scripts" / "readme").glob("*.sh")
+               if b"\r\n" in f.read_bytes()]
+check(10, "CRLF in .sh scripts",
+      len(crlf_files) == 0,
+      f"CRLF still in: {crlf_files}" if crlf_files else "All 18 scripts are LF-only")
+
+# BUG 11
+missing = [s for s in ["full-2k-t2va-h3-base.sh", "full-2k-i2va-h3-base.sh", "full-2k-ref2va-h3-base.sh"]
+           if "--fail-with-body" not in
+           (ROOT / "scripts" / "readme" / s).read_text(encoding="utf-8")]
+check(11, "curl --fail-with-body in h3-base scripts",
+      len(missing) == 0,
+      f"Missing in: {missing}" if missing else "All 3 scripts have the flag")
+
+# BUG 12
+cfg12 = json.loads((ROOT / "audio_vae" / "config.json").read_text(encoding="utf-8"))
+both  = "sample_rate" in cfg12 and "sampling_rate" in cfg12
+equal = cfg12.get("sample_rate") == cfg12.get("sampling_rate")
+check(12, "sampling_rate vs sample_rate",
+      both and equal,
+      "sample_rate=32000 & sampling_rate=32000 both present" if (both and equal) else "MISMATCH")
+
+# BUG 13
+readme = (ROOT / "README.md").read_text(encoding="utf-8", errors="replace")
+ok13 = ("Step 1" in readme and "Step 2" in readme and
+        "Step 3" in readme and "hf download MiniMaxAI/MiniMax-H3" in readme)
+check(13, "README weight download docs",
+      ok13, "4-step guide + hf download command present" if ok13 else "MISSING")
+
+# SUMMARY
+print()
+passes = sum(1 for _, _, s, _ in results if s == "PASS")
+fails  = [r for r in results if r[2] == "FAIL"]
+total  = len(results)
+print("=" * 58)
+print(f"  RESULT: {passes}/{total} checks PASSED")
+if fails:
+    print(f"  FAILED ({len(fails)}):")
+    for num, name, _, detail in fails:
+        print(f"    Bug {num}: {name}  —  {detail}")
+else:
+    print(f"  ALL {total} CHECKS PASSED — EVERY BUG IS FIXED")
+print("=" * 58)
+sys.exit(0 if not fails else 1)