Complete standalone final deliverable and unaligned Q2 results

final/output/q1/README.md

@@ -0,0 +1,34 @@
+# Q1 输出包
+
+本目录对应题目 Q1 的 100 条附件 1 视频样本。原生特征、主视图和物理时间证据均保留；不复制原始视频和标签表。
+
+| 内容 | 说明 |
+|---|---|
+| `features_v2/*.npz` | 100 个样本的文本、音频、视觉原生特征、源区间、质量/观测掩码、对齐视图和查询映射 |
+| `features_v2/sample_summary.csv` | 样本 ID、视频组、源时长、有效词数和状态 |
+| `features_v2/modality_summary.csv` | 300 行模态汇总，含覆盖率、可用质量和缺失情况 |
+| `features_v2/manifest_q1.jsonl` | 逐样本路径、源/特征 SHA-256、配置哈希和状态 |
+| `features_v2/feature_manifest.json` | 特征定义、工具/模型信息、处理配置和包体审计 |
+| `typical_sample_correspondence.csv`、`typical_sample_frames.jpg` | 典型样本的词、物理区间、音频来源行与视频帧对应示例 |
+| `alignment_query_example.png` | 文本词区间到原始音频/视频位置的查询权重图 |
+| `package_audit.json` | 包含文件数、状态和 SHA-256 的完整性审计 |
+| `model_comparison/`、`model_comparison_v2/` | 已整理的对照表与结果文件（以 `REPORTS.md` 中汇总为准） |
+
+对齐视图是固定 0.1 秒网格。文本采用固定转写 CTC 单调路径和保留的边界区间；音频为 74 维，视觉为 OpenFace 2.2.0 的 35 维。四条样本为部分状态：OpenFace 的置信度/成功门控未通过，视觉特征以缺失掩码保留，没有用零值伪装成有效观测。
+
+## 读取
+
+在项目上级目录运行：
+
+```python
+from pathlib import Path
+from final.adapter import Q1AlignmentAdapter
+
+sample = Q1AlignmentAdapter().from_q1_sample(
+    "-iRBcNs9oI8/8",
+    feature_dir=Path("final/output/q1/features_v2"),
+)
+features, mask = sample.q2_arrays()
+```
+
+完整重建命令、依赖和 OpenFace 准备步骤见 [项目 README](../../README.md)。Q1 当前没有人工词边界真值，因此不报告边界误差或概率校准率。

final/output/q1/alignment_query_example.json

@@ -0,0 +1,48 @@
+{
+  "sample_id": "-iRBcNs9oI8/8",
+  "selection": "manually selected example with valid audio and vision query maps",
+  "duration_s": 8.093000411987305,
+  "word_count": 21,
+  "words": [
+    "I've",
+    "drawn",
+    "this",
+    "man",
+    "here,",
+    "and",
+    "if",
+    "you",
+    "can",
+    "tell,",
+    "he",
+    "has",
+    "a",
+    "lot",
+    "of",
+    "big",
+    "muscles,",
+    "he",
+    "looks",
+    "very",
+    "strong"
+  ],
+  "audio_query_shape": [
+    21,
+    803
+  ],
+  "audio_query_nnz": 495,
+  "audio_query_mapped_words": 21,
+  "vision_query_shape": [
+    21,
+    241
+  ],
+  "vision_query_nnz": 160,
+  "vision_query_mapped_words": 21,
+  "audio_query_feature_validity_channel": "log_energy (audio feature dimension 66)",
+  "vision_query_feature_validity_channel": "OpenFace AU12 intensity (vision feature dimension 10)",
+  "alignment_semantics": "per-word normalized overlap between stored CTC word intervals and native source intervals; no learned content similarity",
+  "coordinate_axes_shown": false,
+  "color_scale": "shared square-root intensity transform for visibility; original overlap weights retained",
+  "word_label_positions": "actual transcript labels placed at their CTC word-interval centers on the source-time axis",
+  "image": "alignment_query_example.png"
+}