Publish Ropedia Xperience-10M derived artifacts

ARTIFACT_GUIDE.md

@@ -21,7 +21,8 @@ The project separates these reading layers:
 6. **Data contract:** how one public Xperience-10M sample episode becomes
    aligned model windows and feature blocks.
 7. **Task evidence:** minimal and neural results for the 12 task contracts plus
-   four research-direction extension probes.
+   audio ablation, raw-audio feature replacement, and four research-direction
+   extension probes.
 8. **Reproducibility:** public commands, expected outputs, and exact-match
    evidence for the single-episode pipeline.
 9. **Public project surface:** repo, website, and Hugging Face pages,
@@ -51,6 +52,8 @@ The project separates these reading layers:
 | [`docs/data/xperience10m_dataset_card_alignment.json`](docs/data/xperience10m_dataset_card_alignment.json) | Machine-readable source-alignment summary, including gated metadata, sample license/tooling, and current project coverage. |
 | [`docs/data/source_alignment_audit.json`](docs/data/source_alignment_audit.json) | Machine-readable source metadata and HF card parity report. |
 | [`docs/data/evaluation_protocol.json`](docs/data/evaluation_protocol.json) | Machine-readable evaluation protocol generated from committed metrics. |
+| [`results/audio_ablation/AUDIO_ABLATION_SUMMARY.md`](results/audio_ablation/AUDIO_ABLATION_SUMMARY.md) | Shows measured current-audio and raw log-mel replacement deltas across the 12 task contracts. |
+| [`docs/data/audio_ablation_summary.json`](docs/data/audio_ablation_summary.json) | Machine-readable audio ablation summary for website and HF mirrors. |
 | [`docs/data/quality_gates.json`](docs/data/quality_gates.json) | Machine-readable release-check summary for website and HF mirrors. |
 | [`docs/data/public_surface_qa.json`](docs/data/public_surface_qa.json) | Machine-readable public project-surface report for website, repo, and Hugging Face pages. |
 | [`docs/data/live_publication_status.json`](docs/data/live_publication_status.json) | Last live GitHub/HF verification after upload. |
@@ -101,6 +104,7 @@ The project separates these reading layers:
 | [`results/episode_task_suite/windows.csv`](results/episode_task_suite/windows.csv) | The sample episode is converted into 1,161 aligned 20-frame windows. |
 | [`results/episode_task_suite/feature_manifest.json`](results/episode_task_suite/feature_manifest.json) | The current input vector has 8,546 dimensions with explicit feature-block boundaries, including a 168-d AAC audio block. |
 | [`results/episode_task_suite/available_modalities.json`](results/episode_task_suite/available_modalities.json) | The sample modality coverage is recorded, including the current audio-featurization status. |
+| [`results/audio_ablation/raw_logmel_fisheye_cam0_sr16000_mels64_fft512_hop160.npz`](results/audio_ablation/raw_logmel_fisheye_cam0_sr16000_mels64_fft512_hop160.npz) | Derived 588-d raw log-mel window features decoded from the local public-sample MP4 audio stream; raw audio itself is not redistributed. |
 | [`docs/data/modality_atlas.json`](docs/data/modality_atlas.json) | The responsive website modality cards and derived thumbnail assets are documented without redistributing raw data. |
 | [`docs/assets/modalities/`](docs/assets/modalities/) | Small public-sample thumbnails used by the readable modality atlas. |
 
@@ -113,6 +117,9 @@ The project separates these reading layers:
 | [`results/episode_task_suite/research_directions/`](results/episode_task_suite/research_directions/) | Mapping from the 12 tasks to the four Ropedia research directions. |
 | [`results/episode_task_suite/research_direction_extensions/`](results/episode_task_suite/research_direction_extensions/) | Four additional coded probes, one per research direction. |
 | [`results/episode_task_suite/task_walkthroughs/`](results/episode_task_suite/task_walkthroughs/) | Human-readable research names and case studies explaining input, process modules, output, metric, limitation, and the website task-player data. |
+| [`results/audio_ablation/audio_ablation_metrics.csv`](results/audio_ablation/audio_ablation_metrics.csv) | All 72 measured audio rows: 12 tasks times six variants, including no-audio, handcrafted-audio-only, raw-audio-only, raw replacement, and all-plus-raw. |
+| [`results/audio_ablation/audio_delta_summary.csv`](results/audio_ablation/audio_delta_summary.csv) | Compact per-task audio delta table for quick manual inspection. |
+| [`scripts/audio_ablation_and_raw_upgrade.py`](scripts/audio_ablation_and_raw_upgrade.py) | Regenerates current-AAC audio ablation and raw log-mel upgrade results from real task-suite artifacts plus the local public-sample MP4. |
 | [`scripts/validate_task_surface.py`](scripts/validate_task_surface.py) | Fails publication if public task cards drift back to raw artifact ids or lose their thumbnail/player wiring. |
 
 ## Reproducibility

PROJECT_README.md

@@ -92,6 +92,7 @@ multi-episode held-out model metrics:
 | Feature contract | `results/episode_task_suite/feature_manifest.json`, `available_modalities.json` | 8,546 current features, including a real AAC audio block decoded from `fisheye_cam0.mp4` |
 | Evaluation protocol | `EVALUATION_PROTOCOL.md`, `docs/data/evaluation_protocol.json`, `scripts/build_evaluation_protocol.py` | defines windowing, chronological split, leakage controls, per-task metrics, and current limitations |
 | Research takeaways | `RESEARCH_TAKEAWAYS.md`, `docs/data/research_takeaways.json`, `scripts/build_research_takeaways.py` | summarizes result interpretation from committed metrics and identifies which experiments need held-out episodes |
+| Audio ablation | `scripts/audio_ablation_and_raw_upgrade.py`, `results/audio_ablation/`, `docs/data/audio_ablation_summary.json` | measures current AAC audio contribution and a raw log-mel audio replacement across all 12 task contracts |
 | Research roadmap | `RESEARCH_ROADMAP.md`, `docs/research_roadmap.html`, `docs/data/research_roadmap.json`, `docs/data/research_roadmap_interactive.json` | stages and visualizes the path from public-sample task development to multi-episode held-out evaluation and larger omni-model extensions |
 | 12-task suite | `scripts/episode_task_suite.py`, per-task `metrics.json`, predictions | chronological single-episode split |
 | Single-episode diagnostics | `scripts/single_episode_diagnostics.py`, `results/single_episode_diagnostics/`, `docs/single_episode_explorer.html` | modality ablations, timeline overlay, object-label export, alignment stress tests, and interactive window inspection from one sample episode |
@@ -641,7 +642,7 @@ Current direction-level coverage:
 | --- | --- | --- | --- |
 | A. Human Modeling & Motion Understanding | Partially implemented | Hand Trajectory Forecasting and Contact State Prediction are direct; Action Recognition and Object Relevance Prediction are proxies. Neural MLP improves hand forecasting from `0.8647` to `0.1079` MPJPE. | No full body/shape model, SMPL/MANO target, deformation prior, or multi-episode motion-generation evaluation yet. |
 | B. 3D/4D Reconstruction & Neural Rendering | Proxy tasks only | Cross-Modal Retrieval, Cross-Modal Reconstruction, and Multimodal Synchronization Detection test alignment/reconstruction prerequisites. | No NeRF, Gaussian Splatting, TSDF, mesh, novel-view synthesis, or calibrated 4D reconstruction model yet. |
-| C. Egocentric Vision & Interaction | Strongest implemented track | 6 direct tasks: action, subtask, transition, next-action, object relevance, and caption grounding, plus alignment/order diagnostics. | Single-episode chronological split limits generalization; audio and stronger video-language backbones still need to be added. |
+| C. Egocentric Vision & Interaction | Strongest implemented track | 6 direct tasks: action, subtask, transition, next-action, object relevance, and caption grounding, plus alignment/order diagnostics and audio ablation. | Single-episode chronological split limits generalization; stronger audio and video-language backbones still need multi-episode testing. |
 | D. Scene Reconstruction & World Modeling | Early proxy tasks | Procedure Step Recognition, Next-Action Prediction, Object Relevance Prediction, Cross-Modal Retrieval, Cross-Modal Reconstruction, Temporal Order Verification, and Multimodal Synchronization Detection provide state/world-model probes. | No persistent scene graph, object permanence task, long-term map, or held-out-episode world model yet. |
 
 The important interpretation is that all four directions can be **started** from
@@ -777,6 +778,35 @@ The task-specific heads are:
 | Neural MLP hand forecast | 0.1079 MPJPE | n/a | Same features/split, nonlinear regression head |
 | Neural MLP temporal order | 0.8520 F1 | 0.8578 | Strong improvement on adjacent-window ordering |
 | Neural MLP misalignment | 0.7153 F1 | 0.7009 | Detects shifted motion/visual/audio pairs better than the linear head |
+| Audio ablation | +0.0418 mean delta | n/a | Current AAC audio improves the primary metric on 6 of 12 task contracts |
+| Raw log-mel audio replacement | +0.0936 mean delta | n/a | Raw log-mel replacement beats current handcrafted audio on 6 of 12 task contracts |
+
+## Audio Ablation and Raw-Audio Upgrade
+
+The current AAC audio block is now tested rather than only included. The script
+[`scripts/audio_ablation_and_raw_upgrade.py`](scripts/audio_ablation_and_raw_upgrade.py)
+reuses the real task-suite windows, decodes the local public-sample
+`fisheye_cam0.mp4` audio stream, builds a 588-d raw log-mel window feature, and
+evaluates six variants for every task: current features, no audio,
+handcrafted-audio-only, raw-audio-only, handcrafted audio replaced by raw
+log-mel, and current features plus raw log-mel.
+
+The measured single-episode result is task-specific:
+
+| Readout | Value |
+| --- | ---: |
+| Tasks where current AAC audio improves the primary metric | 6 / 12 |
+| Mean current-audio delta | +0.0418 |
+| Tasks where raw log-mel replacement improves over handcrafted AAC | 6 / 12 |
+| Mean raw-replacement delta vs current audio | +0.0936 |
+
+Full files:
+
+- [`results/audio_ablation/AUDIO_ABLATION_SUMMARY.md`](results/audio_ablation/AUDIO_ABLATION_SUMMARY.md)
+- [`results/audio_ablation/audio_ablation_metrics.csv`](results/audio_ablation/audio_ablation_metrics.csv)
+- [`results/audio_ablation/audio_delta_summary.csv`](results/audio_ablation/audio_delta_summary.csv)
+- [`docs/data/audio_ablation_summary.json`](docs/data/audio_ablation_summary.json)
+- [`docs/assets/charts/audio_ablation_delta.svg`](docs/assets/charts/audio_ablation_delta.svg)
 
 ## Neural MLP Results
 

PROJECT_STATUS.md

@@ -10,6 +10,7 @@ the next development step.
 | Public-sample pipeline | Verified | `results/episode_task_suite/summary_report.json`, `results/episode_task_suite/windows.csv`, `results/episode_task_suite/feature_manifest.json` | One public Xperience-10M sample episode is converted into 5,821 frames, 1,161 aligned 20-frame windows, and an 8,546-dimensional current feature contract. |
 | Task suite | Verified | `scripts/episode_task_suite.py`, `results/episode_task_suite/`, `docs/data/summary_metrics.json` | All 12 task contracts have committed metrics, predictions, and minimal baseline outputs. |
 | Neural heads | Verified | `scripts/neural_task_models.py`, `results/episode_task_suite/neural_mlp/` | Each task also has a compact PyTorch MLP run over the same feature tensor and chronological split. |
+| Audio ablation and raw-audio upgrade | Verified | `scripts/audio_ablation_and_raw_upgrade.py`, `results/audio_ablation/`, `docs/data/audio_ablation_summary.json` | Current AAC audio improves the primary metric on 6 of 12 task contracts; replacing the current handcrafted block with a 588-d raw log-mel feature improves over current audio on 6 of 12 tasks. |
 | Research takeaways | Verified | `RESEARCH_TAKEAWAYS.md`, `docs/data/research_takeaways.json`, `scripts/build_research_takeaways.py` | The main result interpretation is generated from committed metrics: chronological class shift, neural gains on dynamics/order/alignment, open retrieval/reconstruction problems, and the need for held-out episodes. |
 | Research roadmap | Current | `RESEARCH_ROADMAP.md`, `docs/data/research_roadmap.json` | The staged path connects public-sample task development to multi-episode data staging, the 32-episode Qwen3-Omni LoRA pilot, robustness runs, and larger omni-model extensions. |
 | Evaluation protocol | Verified | `EVALUATION_PROTOCOL.md`, `docs/data/evaluation_protocol.json`, `scripts/build_evaluation_protocol.py` | Windowing, chronological split, per-task metrics, leakage controls, and current limitations are generated from committed metric artifacts. |
@@ -32,12 +33,14 @@ the next development step.
    the staged path from public-sample task work to multi-episode modeling.
 5. Inspect `docs/data/summary_metrics.json` and
    `results/episode_task_suite/neural_mlp/` to check the 12-task outputs.
-6. Inspect `EVALUATION_PROTOCOL.md` before judging task metrics or leakage
+6. Inspect `results/audio_ablation/AUDIO_ABLATION_SUMMARY.md` before judging
+   whether audio helps the current task suite.
+7. Inspect `EVALUATION_PROTOCOL.md` before judging task metrics or leakage
    controls.
-7. Inspect `SOURCE_ALIGNMENT_AUDIT.md` and
+8. Inspect `SOURCE_ALIGNMENT_AUDIT.md` and
    `XPERIENCE10M_DATASET_CARD_ALIGNMENT.md` before judging dataset
    wording.
-8. Inspect `results/omni_finetune/DATA_ACCESS_STATUS.md` before judging
+9. Inspect `results/omni_finetune/DATA_ACCESS_STATUS.md` before judging
    Qwen3-Omni scale-up status.
 
 ## Current Reading Notes
@@ -50,3 +53,6 @@ the next development step.
   depth, meshes, NeRF outputs, or Gaussian splats.
 - AAC audio is decoded from `fisheye_cam0.mp4` and included in the current
   8,546-dimensional baseline feature vector.
+- Audio is now evaluated directly: the current AAC block and a raw log-mel
+  replacement are compared across all 12 task contracts in
+  `results/audio_ablation/`.

README.md

@@ -12,6 +12,7 @@ tags:
   - neural-network
   - pytorch
   - retrieval
+  - audio
 task_categories:
   - robotics
   - time-series-forecasting
@@ -26,14 +27,16 @@ size_categories:
 
 This dataset repo stores derived artifacts for the public Ropedia Xperience-10M
 sample project: manifests, metrics, predictions, figures, notes, scripts,
-website data, and compact baseline task-head files. It does not redistribute
-raw Xperience-10M MP4/HDF5/RRD data or full Qwen weights.
+website data, compact baseline task-head files, audio ablation results, and
+derived raw-log-mel audio window features. It does not redistribute raw
+Xperience-10M MP4/HDF5/RRD data or full Qwen weights.
 
 Current public-sample contract:
 
 - 1 public Xperience-10M sample episode
 - 5,821 frames and 1,161 aligned 20-frame windows
 - 8,546 feature dimensions, including `audio_fisheye_cam0_aac`
+- 588-d derived raw log-mel audio window features for the audio-upgrade probe
 - 12 minimal task heads, 12 compact neural MLP heads, and 4 extension probes
 - single-episode chronological split; multi-episode held-out metrics are still data-gated
 
@@ -54,15 +57,19 @@ Useful entry points:
 | `PUBLIC_SURFACE_QA.md` / `docs/data/public_surface_qa.json` | Public project surface across GitHub, website, and Hugging Face |
 | `results/episode_task_suite/summary_report.json` | 12-task minimal and neural metric rollup |
 | `results/episode_task_suite/feature_manifest.json` | 8,546-d feature-block contract |
+| `results/audio_ablation/AUDIO_ABLATION_SUMMARY.md` | Human-readable audio ablation and raw-audio upgrade report |
+| `results/audio_ablation/audio_ablation_summary.json` | Per-task audio deltas for current AAC, no-audio, raw-only, replacement, and all-plus-raw variants |
+| `results/audio_ablation/raw_logmel_fisheye_cam0_sr16000_mels64_fft512_hop160.npz` | Derived raw-log-mel window features; raw audio itself is not redistributed |
 | `docs/single_episode_explorer.html` | Static window-level explorer |
 | `docs/research_roadmap.html` | Interactive roadmap for four research tracks and scale-up |
 
 The mirrored public package also includes `docs/data/figure_index.json`,
 `docs/data/brand_assets.json`, `scripts/build_brand_assets.py`,
-`scripts/build_research_takeaways.py`, the task-first 12-task map, the
-interactive scrub/play walkthrough storyboard, including critical website HTML,
-`docs/data/task_surface_integrity.json`,
-`docs/data/rendered_site_check.json`, and `docs/data/public_surface_qa.json`.
+`scripts/build_research_takeaways.py`, `scripts/audio_ablation_and_raw_upgrade.py`,
+the task-first 12-task map, the interactive scrub/play walkthrough storyboard,
+including critical website HTML, `docs/data/task_surface_integrity.json`,
+`docs/data/rendered_site_check.json`, `docs/data/audio_ablation_summary.json`,
+and `docs/data/public_surface_qa.json`.
 
 The official upstream full dataset is
 [`ropedia-ai/xperience-10m`](https://huggingface.co/datasets/ropedia-ai/xperience-10m),

RESEARCH_TAKEAWAYS.md

@@ -78,6 +78,23 @@ Source: `results/episode_task_suite/cross_modal_retrieval/metrics.json`.
 
 Current scope: The current reconstruction task predicts feature vectors; depth, mesh, NeRF, and Gaussian-splatting outputs are future task variants.
 
+### Audio helps some tasks and hurts others on the public sample
+
+The current AAC audio block improves the primary metric on 6 of 12 tasks, while raw log-mel replacement improves over the current handcrafted block on 6 of 12 tasks. The largest current-audio gain appears in feature reconstruction, not in action classification.
+
+| Metric | Value |
+| --- | ---: |
+| `tasks_where_current_audio_improves` | 6 |
+| `mean_current_audio_delta` | 0.0418 |
+| `tasks_where_raw_replacement_improves` | 6 |
+| `mean_raw_replacement_delta_vs_current` | 0.0936 |
+| `reconstruction_current_audio_delta` | 0.6524 |
+| `object_relevance_current_audio_delta` | 0.0102 |
+
+Source: `results/audio_ablation/audio_ablation_summary.json`.
+
+Current scope: This is a single-episode ablation over fixed ridge heads. It validates that audio is wired into the task suite and shows where it changes metrics; it does not prove cross-episode audio generalization.
+
 ### The next scientific unit is held-out episodes, not more adjacent windows
 
 The prepared Qwen3-Omni path targets 32 episodes from 32 sessions, but it remains data-gated until access and held-out evaluation complete.
@@ -97,5 +114,6 @@ Current scope: The 32-episode Qwen3-Omni fine-tune requires gated data staging a
 - High single-episode scores are useful pipeline checks for the current task contracts.
 - Low chronological action/subtask scores are informative because they expose later-label shift.
 - Neural gains on trajectory/order/alignment make those tasks good candidates for the next fine-tuning stage.
+- Audio ablation is task-specific: current AAC and raw log-mel features help some probes and hurt others.
 - Retrieval and reconstruction remain the main multimodal representation challenges.
 - The next credible model-quality result needs held-out episodes.

docs/data/artifact_index.json

@@ -1,8 +1,8 @@
 {
   "title": "Ropedia Xperience-10M Task Suite Artifact Index",
-  "generated_at_utc": "2026-06-03T13:04:23+00:00",
+  "generated_at_utc": "2026-06-03T14:17:31+00:00",
   "status": "pass",
-  "artifact_count": 65,
+  "artifact_count": 70,
   "missing": [],
   "by_kind": {
     "project_path": 9,
@@ -10,8 +10,10 @@
     "source_alignment": 5,
     "publication_workflow": 1,
     "evaluation_protocol": 3,
-    "result_interpretation": 3,
-    "visual_evidence": 6,
+    "result_interpretation": 5,
+    "metrics_source": 3,
+    "website_data": 3,
+    "visual_evidence": 7,
     "quality_gate": 12,
     "reproducibility": 2,
     "publication_package_check": 1,
@@ -19,8 +21,6 @@
     "mirror_parity": 1,
     "integrity_report": 1,
     "metadata": 1,
-    "metrics_source": 2,
-    "website_data": 2,
     "data_contract": 3,
     "result_directory": 1,
     "taxonomy": 1,
@@ -62,8 +62,8 @@
       "surface": "repo_hf",
       "shows": "Gives a compact current-state table for first-pass readers.",
       "exists": true,
-      "bytes": 5384,
-      "sha256": "3af38f7eefee59b48dfbdfb316f75f428e06fb2a2a5d37efa798ee3e2dae53f0"
+      "bytes": 6031,
+      "sha256": "b2de2d9de04e0a05046cbea8a0b7c2e8766f6171b5c80a1b3f244aa1c51023ea"
     },
     {
       "id": "project_status_json",
@@ -73,8 +73,8 @@
       "surface": "website_hf",
       "shows": "Machine-readable copy of the current project status for website and HF mirrors.",
       "exists": true,
-      "bytes": 7340,
-      "sha256": "3f54a794582f4b61f60fe240a9ea6f1782e29a21498e2d63dad670495d4e0d39"
+      "bytes": 8162,
+      "sha256": "19471dbaf46a16275b8f0deaf67a5326b9508adc911a5c6d3426a2565049dd46"
     },
     {
       "id": "research_roadmap",
@@ -128,8 +128,8 @@
       "surface": "repo_hf",
       "shows": "Gives the human-readable map from project scope to data, tasks, platform mirrors, and scale-up status.",
       "exists": true,
-      "bytes": 14599,
-      "sha256": "b08c9c96b4844e7b2dfb0f35d45c70eda1289c9250e0e348f20a0bad6a353cff"
+      "bytes": 15981,
+      "sha256": "1e79ab74c4dee1e5e5a9b7346a9822977fd557a7350f2426e846211afee2f375"
     },
     {
       "id": "official_dataset_card_alignment",
@@ -173,7 +173,7 @@
       "shows": "Machine-readable source-alignment pass/fail check for repo, website, and HF surfaces.",
       "exists": true,
       "bytes": 4432,
-      "sha256": "056de9ea87b93874260b9a1b6191ae9f2c637b6763bfd3186b08df60eb682e53"
+      "sha256": "9d8eb05aa29f898a28f64b0b23ef0b912ba0f43d6ef558d213e05619da0b3197"
     },
     {
       "id": "source_alignment_validator",
@@ -238,8 +238,8 @@
       "surface": "repo_hf",
       "shows": "Summarizes the main research lessons from committed metrics and identifies which experiments need held-out episodes.",
       "exists": true,
-      "bytes": 3856,
-      "sha256": "3b575d4b47ba4e09f5e16d065dab78194148eb52394b30f588930237e0518f0d"
+      "bytes": 4893,
+      "sha256": "16a002167d0702ea7e85ad1a494800daad4ff080c3751e078275f7c196721046"
     },
     {
       "id": "research_takeaways_json",
@@ -249,8 +249,8 @@
       "surface": "website_hf",
       "shows": "Machine-readable result interpretation for the website, HF cards, and mirror checks.",
       "exists": true,
-      "bytes": 5356,
-      "sha256": "0621f28d602582c70524cd9455b01558e7003267b1f4786d005e31597e46b051"
+      "bytes": 6814,
+      "sha256": "538ec80950dbf5d77cc6ab2bd5a96a3af191b56a257b895e48b6cf28deb0d044"
     },
     {
       "id": "research_takeaways_builder",
@@ -260,8 +260,63 @@
       "surface": "repo_hf",
       "shows": "Regenerates the research takeaways from committed summary metrics and task result artifacts.",
       "exists": true,
-      "bytes": 10709,
-      "sha256": "dfb6cc169b5a5a7e0b2f3d2aeb604fd0eeb586e97cfb114a0fe32844c7ff8617"
+      "bytes": 13304,
+      "sha256": "135f555b17aae5dc5965ec88c3db6e49622a480467909ffa10d53278d4ea054b"
+    },
+    {
+      "id": "audio_ablation_script",
+      "title": "Audio ablation and raw-audio upgrade script",
+      "path": "scripts/audio_ablation_and_raw_upgrade.py",
+      "kind": "result_interpretation",
+      "surface": "repo_hf",
+      "shows": "Measures current AAC audio contribution and a raw log-mel audio feature replacement across all 12 task contracts.",
+      "exists": true,
+      "bytes": 43192,
+      "sha256": "73db45a3ddb50b60f61ca5e55bdf2008e48c268681bfcdc95d35bb56e48f89e3"
+    },
+    {
+      "id": "audio_ablation_summary",
+      "title": "Audio ablation summary",
+      "path": "results/audio_ablation/audio_ablation_summary.json",
+      "kind": "metrics_source",
+      "surface": "repo_hf",
+      "shows": "Stores per-task audio deltas for all current features, no-audio, handcrafted-audio-only, raw-audio-only, raw replacement, and all-plus-raw variants.",
+      "exists": true,
+      "bytes": 9735,
+      "sha256": "514370c53baeb51373160bbd5ab7f5f0de7733301bf1b82981ed799aec41152a"
+    },
+    {
+      "id": "audio_ablation_summary_md",
+      "title": "Audio ablation summary report",
+      "path": "results/audio_ablation/AUDIO_ABLATION_SUMMARY.md",
+      "kind": "result_interpretation",
+      "surface": "repo_hf",
+      "shows": "Human-readable table showing the measured audio contribution and raw-audio replacement delta for every task.",
+      "exists": true,
+      "bytes": 2166,
+      "sha256": "9318bc55b1b051c384ce2fd6e7cd0a56becfa00185a0b597fdc474a40b1f17f9"
+    },
+    {
+      "id": "audio_ablation_website_json",
+      "title": "Audio ablation website JSON",
+      "path": "docs/data/audio_ablation_summary.json",
+      "kind": "website_data",
+      "surface": "website_hf",
+      "shows": "Machine-readable audio ablation summary mirrored into the static website and Hugging Face bundles.",
+      "exists": true,
+      "bytes": 9735,
+      "sha256": "514370c53baeb51373160bbd5ab7f5f0de7733301bf1b82981ed799aec41152a"
+    },
+    {
+      "id": "audio_ablation_delta_chart",
+      "title": "Audio ablation delta chart",
+      "path": "docs/assets/charts/audio_ablation_delta.svg",
+      "kind": "visual_evidence",
+      "surface": "website_hf",
+      "shows": "Bar chart of measured current-audio primary-metric deltas across the 12 tasks.",
+      "exists": true,
+      "bytes": 4146,
+      "sha256": "187dbabe01f9ff18841ff61a1e7fbf85bebdd188cc0f248bb5090d64528e7568"
     },
     {
       "id": "figure_index",
@@ -283,7 +338,7 @@
       "shows": "Machine-readable visual asset index for website and Hugging Face mirrors.",
       "exists": true,
       "bytes": 13434,
-      "sha256": "fc7c9d504509bea2bab08238c96b745f00fdd6a016019d121d4bf0e22b5f3936"
+      "sha256": "1a28548d2838b5a5ea9464761ccb73a3e55d6d48dd4c68d321f056bfdc4ad560"
     },
     {
       "id": "figure_index_builder",
@@ -349,7 +404,7 @@
       "shows": "Machine-readable release-check summary for validators, mirrors, and public project surfaces.",
       "exists": true,
       "bytes": 8147,
-      "sha256": "f5b7fa88aaae118ddf2ceba02a05f961a6a195fa8aa7a1055211c5a7de0143e2"
+      "sha256": "40c9cdc9d5afd15f8c69bda48a0eb36d81df14adff5555de54a481df1baed5ca"
     },
     {
       "id": "public_surface_qa",
@@ -371,7 +426,7 @@
       "volatile": true,
       "shows": "Machine-readable report for SEO/social metadata, accessible tab semantics, public links, project links, and reader-facing copy.",
       "exists": true,
-      "bytes": 5651,
+      "bytes": 5648,
       "hash_policy": "existence_and_size_only"
     },
     {
@@ -452,7 +507,7 @@
       "volatile": true,
       "shows": "Records the last live GitHub/HF URL verification after upload.",
       "exists": true,
-      "bytes": 68740,
+      "bytes": 68744,
       "hash_policy": "existence_and_size_only"
     },
     {
@@ -496,8 +551,8 @@
       "surface": "repo_hf",
       "shows": "Generates the selective artifact catalog from local files.",
       "exists": true,
-      "bytes": 24091,
-      "sha256": "f21aaf9267e93e571b093525525749054bee85872506f01ce534eb4fa131a56c"
+      "bytes": 25927,
+      "sha256": "4a9796faf0cd87b052a69f5fde89a7a3a88008f8e4485c05f9f5cc7630a89164"
     },
     {
       "id": "publication_audit",
@@ -508,7 +563,7 @@
       "volatile": true,
       "shows": "Confirms public bundles exclude raw data, caches, heavy archives, and token strings.",
       "exists": true,
-      "bytes": 7253,
+      "bytes": 7289,
       "hash_policy": "existence_and_size_only"
     },
     {
@@ -532,7 +587,7 @@
       "volatile": true,
       "shows": "Confirms prepared GitHub/HF Space/artifact/model mirrors share the same critical data, figure, website HTML, and validator files.",
       "exists": true,
-      "bytes": 97333,
+      "bytes": 97337,
       "hash_policy": "existence_and_size_only"
     },
     {
@@ -544,7 +599,7 @@
       "volatile": true,
       "shows": "Confirms local website links, anchors, JSON data files, and referenced images resolve.",
       "exists": true,
-      "bytes": 14310,
+      "bytes": 14587,
       "hash_policy": "existence_and_size_only"
     },
     {
@@ -622,7 +677,7 @@
       "shows": "Stores matching PyTorch MLP results for the 12 task contracts.",
       "exists": true,
       "file_count": 60,
-      "bytes": 90612122
+      "bytes": 90608884
     },
     {
       "id": "research_direction_taxonomy",

docs/data/audio_ablation_summary.json

@@ -0,0 +1,223 @@
+{
+  "description": "Measured audio ablation and raw log-mel audio upgrade over the single public Xperience-10M sample episode.",
+  "scope": "single public sample episode; chronological split; ridge heads over fixed feature contracts",
+  "raw_audio_metadata": {
+    "source": "local_public_sample/fisheye_cam0.mp4",
+    "exists": true,
+    "has_audio": true,
+    "sample_rate": 16000,
+    "fps": 20.00137419266181,
+    "num_samples": 4656994,
+    "num_windows": 1161,
+    "feature_dim": 588,
+    "mel_bands": 64,
+    "fft_size": 512,
+    "hop_length": 160,
+    "feature_description": "Per-window raw waveform STFT log-mel statistics plus delta and waveform envelope statistics."
+  },
+  "num_tasks": 12,
+  "variants": {
+    "all_handcrafted_audio": "All Current Features",
+    "all_except_audio": "All Except Audio",
+    "handcrafted_audio_only": "Handcrafted AAC Audio Only",
+    "raw_logmel_audio_only": "Raw Log-Mel Audio Only",
+    "replace_handcrafted_with_raw": "Replace AAC Block With Raw Log-Mel",
+    "all_plus_raw_logmel": "All Current Features + Raw Log-Mel"
+  },
+  "task_summaries": [
+    {
+      "task": "timeline_action",
+      "task_display": "Current Action Recognition",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.00905456968081885,
+      "all_except_audio": 0.008771929824561405,
+      "handcrafted_audio_delta": 0.0002826398562574446,
+      "raw_logmel_audio_only": 0.0,
+      "replace_handcrafted_with_raw": 0.0013495276653171392,
+      "raw_replacement_delta_vs_no_audio": -0.007422402159244265,
+      "raw_replacement_delta_vs_handcrafted": -0.00770504201550171,
+      "all_plus_raw_logmel": 0.002734107997265892,
+      "all_plus_raw_delta_vs_handcrafted": -0.006320461683552957
+    },
+    {
+      "task": "timeline_subtask",
+      "task_display": "Current Subtask Recognition",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.011256354393609296,
+      "all_except_audio": 0.0111731843575419,
+      "handcrafted_audio_delta": 8.317003606739606e-05,
+      "raw_logmel_audio_only": 0.0016722408026755855,
+      "replace_handcrafted_with_raw": 0.0008257638315441783,
+      "raw_replacement_delta_vs_no_audio": -0.01034742052599772,
+      "raw_replacement_delta_vs_handcrafted": -0.010430590562065117,
+      "all_plus_raw_logmel": 0.0017889087656529517,
+      "all_plus_raw_delta_vs_handcrafted": -0.009467445627956345
+    },
+    {
+      "task": "transition_detection",
+      "task_display": "Action Transition Detection",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.46213292117465227,
+      "all_except_audio": 0.46870229007633585,
+      "handcrafted_audio_delta": -0.006569368901683581,
+      "raw_logmel_audio_only": 0.4637904468412942,
+      "replace_handcrafted_with_raw": 0.4792100707180375,
+      "raw_replacement_delta_vs_no_audio": 0.010507780641701658,
+      "raw_replacement_delta_vs_handcrafted": 0.01707714954338524,
+      "all_plus_raw_logmel": 0.4816233470132239,
+      "all_plus_raw_delta_vs_handcrafted": 0.019490425838571634
+    },
+    {
+      "task": "next_action",
+      "task_display": "Next-Action Prediction",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.01058201058201058,
+      "all_except_audio": 0.010709504685408301,
+      "handcrafted_audio_delta": -0.0001274941033977215,
+      "raw_logmel_audio_only": 0.0017301038062283738,
+      "replace_handcrafted_with_raw": 0.006006006006006006,
+      "raw_replacement_delta_vs_no_audio": -0.004703498679402295,
+      "raw_replacement_delta_vs_handcrafted": -0.004576004576004574,
+      "all_plus_raw_logmel": 0.0058479532163742695,
+      "all_plus_raw_delta_vs_handcrafted": -0.00473405736563631
+    },
+    {
+      "task": "hand_trajectory_forecast",
+      "task_display": "Future Hand Motion Forecasting",
+      "primary_metric": "mae",
+      "higher_is_better": false,
+      "all_handcrafted_audio": 4.466395378112793,
+      "all_except_audio": 4.303755283355713,
+      "handcrafted_audio_delta": -0.16264009475708008,
+      "raw_logmel_audio_only": 3.1172122955322266,
+      "replace_handcrafted_with_raw": 4.305870532989502,
+      "raw_replacement_delta_vs_no_audio": -0.0021152496337890625,
+      "raw_replacement_delta_vs_handcrafted": 0.16052484512329102,
+      "all_plus_raw_logmel": 4.1367621421813965,
+      "all_plus_raw_delta_vs_handcrafted": 0.3296332359313965
+    },
+    {
+      "task": "contact_prediction",
+      "task_display": "Contact State Prediction",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 1.0,
+      "all_except_audio": 1.0,
+      "handcrafted_audio_delta": 0.0,
+      "raw_logmel_audio_only": 1.0,
+      "replace_handcrafted_with_raw": 1.0,
+      "raw_replacement_delta_vs_no_audio": 0.0,
+      "raw_replacement_delta_vs_handcrafted": 0.0,
+      "all_plus_raw_logmel": 1.0,
+      "all_plus_raw_delta_vs_handcrafted": 0.0
+    },
+    {
+      "task": "object_relevance",
+      "task_display": "Relevant Object Prediction",
+      "primary_metric": "micro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.15813953488372093,
+      "all_except_audio": 0.14793328498912256,
+      "handcrafted_audio_delta": 0.010206249894598368,
+      "raw_logmel_audio_only": 0.15894868585732164,
+      "replace_handcrafted_with_raw": 0.17871759890859482,
+      "raw_replacement_delta_vs_no_audio": 0.030784313919472256,
+      "raw_replacement_delta_vs_handcrafted": 0.020578064024873888,
+      "all_plus_raw_logmel": 0.18262653898768813,
+      "all_plus_raw_delta_vs_handcrafted": 0.024487004103967203
+    },
+    {
+      "task": "caption_grounding",
+      "task_display": "Language-to-Time Grounding",
+      "primary_metric": "mrr",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.03208567947149277,
+      "all_except_audio": 0.027228528633713722,
+      "handcrafted_audio_delta": 0.004857150837779045,
+      "raw_logmel_audio_only": 0.014815197326242924,
+      "replace_handcrafted_with_raw": 0.02484782598912716,
+      "raw_replacement_delta_vs_no_audio": -0.002380702644586563,
+      "raw_replacement_delta_vs_handcrafted": -0.007237853482365608,
+      "all_plus_raw_logmel": 0.02719014883041382,
+      "all_plus_raw_delta_vs_handcrafted": -0.004895530641078949
+    },
+    {
+      "task": "cross_modal_retrieval",
+      "task_display": "Cross-Modal Window Retrieval",
+      "primary_metric": "mrr",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.3751238286495209,
+      "all_except_audio": 0.38921058177948,
+      "handcrafted_audio_delta": -0.014086753129959106,
+      "raw_logmel_audio_only": 0.01806792803108692,
+      "replace_handcrafted_with_raw": 0.32749155163764954,
+      "raw_replacement_delta_vs_no_audio": -0.061719030141830444,
+      "raw_replacement_delta_vs_handcrafted": -0.04763227701187134,
+      "all_plus_raw_logmel": 0.31795138120651245,
+      "all_plus_raw_delta_vs_handcrafted": -0.05717244744300842
+    },
+    {
+      "task": "modality_reconstruction",
+      "task_display": "Sensor-to-Visual Reconstruction",
+      "primary_metric": "mae",
+      "higher_is_better": false,
+      "all_handcrafted_audio": 9.79421329498291,
+      "all_except_audio": 10.446661949157715,
+      "handcrafted_audio_delta": 0.6524486541748047,
+      "raw_logmel_audio_only": 2.6225292682647705,
+      "replace_handcrafted_with_raw": 8.830678939819336,
+      "raw_replacement_delta_vs_no_audio": 1.615983009338379,
+      "raw_replacement_delta_vs_handcrafted": 0.9635343551635742,
+      "all_plus_raw_logmel": 8.392388343811035,
+      "all_plus_raw_delta_vs_handcrafted": 1.401824951171875
+    },
+    {
+      "task": "temporal_order",
+      "task_display": "Temporal Order Verification",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.5172413793103449,
+      "all_except_audio": 0.4942528735632184,
+      "handcrafted_audio_delta": 0.022988505747126464,
+      "raw_logmel_audio_only": 0.5028735632183908,
+      "replace_handcrafted_with_raw": 0.5301714439065678,
+      "raw_replacement_delta_vs_no_audio": 0.03591857034334939,
+      "raw_replacement_delta_vs_handcrafted": 0.012930064596222923,
+      "all_plus_raw_logmel": 0.5330450130569861,
+      "all_plus_raw_delta_vs_handcrafted": 0.015803633746641288
+    },
+    {
+      "task": "misalignment_detection",
+      "task_display": "Cross-Modal Misalignment Detection",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.41734045375379186,
+      "all_except_audio": 0.42258557365378524,
+      "handcrafted_audio_delta": -0.005245119899993378,
+      "raw_logmel_audio_only": 0.47823544277887897,
+      "replace_handcrafted_with_raw": 0.44378951880827355,
+      "raw_replacement_delta_vs_no_audio": 0.021203945154488313,
+      "raw_replacement_delta_vs_handcrafted": 0.02644906505448169,
+      "all_plus_raw_logmel": 0.4373795761078998,
+      "all_plus_raw_delta_vs_handcrafted": 0.02003912235410793
+    }
+  ],
+  "aggregate": {
+    "mean_handcrafted_audio_delta": 0.041849794979543296,
+    "tasks_where_handcrafted_audio_improves": 6,
+    "mean_raw_replacement_delta_vs_handcrafted": 0.09362598132150173,
+    "tasks_where_raw_replacement_improves_over_handcrafted": 6
+  },
+  "provenance": {
+    "suite_dir": "results/episode_task_suite",
+    "shared_windows": "results/episode_task_suite/shared_windows.npz",
+    "feature_manifest": "results/episode_task_suite/feature_manifest.json",
+    "audio_source": "local_public_sample/fisheye_cam0.mp4",
+    "annotation_source": "local_public_sample/annotation.hdf5",
+    "homie_toolkit_available": true
+  }
+}

docs/data/figure_index.json

@@ -1,7 +1,7 @@
 {
   "title": "Ropedia Xperience-10M Figure Index",
   "status": "pass",
-  "generated_at_utc": "2026-06-03T12:56:14+00:00",
+  "generated_at_utc": "2026-06-03T14:16:25+00:00",
   "scope": "Public figures, diagrams, charts, and derived modality thumbnails. Raw Xperience-10M videos, annotations, RRD files, and Qwen weights are excluded.",
   "figure_count": 21,
   "figures": [

docs/data/mirror_parity.json

@@ -1,9 +1,9 @@
 {
   "status": "pass",
-  "generated_at_utc": "2026-06-03T13:54:27+00:00",
+  "generated_at_utc": "2026-06-03T14:25:52+00:00",
   "hf_root": "hf_publish",
   "summary": {
-    "group_count": 91,
+    "group_count": 99,
     "failure_count": 0,
     "failures_by_surface": {}
   },
@@ -34,33 +34,64 @@
     }
   ],
   "groups": [
+    {
+      "name": "data/audio_ablation_summary.json",
+      "status": "pass",
+      "local": {
+        "path": "repo:docs/data/audio_ablation_summary.json",
+        "exists": true,
+        "bytes": 9735,
+        "sha256": "514370c53baeb51373160bbd5ab7f5f0de7733301bf1b82981ed799aec41152a"
+      },
+      "mirrors": {
+        "hf_space": {
+          "path": "hf_space:data/audio_ablation_summary.json",
+          "exists": true,
+          "bytes": 9735,
+          "sha256": "514370c53baeb51373160bbd5ab7f5f0de7733301bf1b82981ed799aec41152a"
+        },
+        "hf_artifacts": {
+          "path": "hf_artifacts:docs/data/audio_ablation_summary.json",
+          "exists": true,
+          "bytes": 9735,
+          "sha256": "514370c53baeb51373160bbd5ab7f5f0de7733301bf1b82981ed799aec41152a"
+        },
+        "hf_model": {
+          "path": "hf_model:metrics/audio_ablation_summary.json",
+          "exists": true,
+          "bytes": 9735,
+          "sha256": "514370c53baeb51373160bbd5ab7f5f0de7733301bf1b82981ed799aec41152a"
+        }
+      },
+      "failures": []
+    },
     {
       "name": "data/artifact_index.json",
       "status": "pass",
       "local": {
         "path": "repo:docs/data/artifact_index.json",
         "exists": true,
-        "bytes": 29016,
-        "sha256": "cbf6469d8ec892ae37bf034e1eb62ce234054955616f4bda5b6609fd47022dea"
+        "bytes": 31423,
+        "sha256": "c696300e33ab3b6fd92accca9160340e68cf99f184ced8a5dcdc3e0baf9f653b"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:data/artifact_index.json",
           "exists": true,
-          "bytes": 29016,
-          "sha256": "cbf6469d8ec892ae37bf034e1eb62ce234054955616f4bda5b6609fd47022dea"
+          "bytes": 31423,
+          "sha256": "c696300e33ab3b6fd92accca9160340e68cf99f184ced8a5dcdc3e0baf9f653b"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:docs/data/artifact_index.json",
           "exists": true,
-          "bytes": 29016,
-          "sha256": "cbf6469d8ec892ae37bf034e1eb62ce234054955616f4bda5b6609fd47022dea"
+          "bytes": 31423,
+          "sha256": "c696300e33ab3b6fd92accca9160340e68cf99f184ced8a5dcdc3e0baf9f653b"
         },
         "hf_model": {
           "path": "hf_model:metrics/artifact_index.json",
           "exists": true,
-          "bytes": 29016,
-          "sha256": "cbf6469d8ec892ae37bf034e1eb62ce234054955616f4bda5b6609fd47022dea"
+          "bytes": 31423,
+          "sha256": "c696300e33ab3b6fd92accca9160340e68cf99f184ced8a5dcdc3e0baf9f653b"
         }
       },
       "failures": []
@@ -165,26 +196,26 @@
         "path": "repo:docs/data/figure_index.json",
         "exists": true,
         "bytes": 13434,
-        "sha256": "fc7c9d504509bea2bab08238c96b745f00fdd6a016019d121d4bf0e22b5f3936"
+        "sha256": "1a28548d2838b5a5ea9464761ccb73a3e55d6d48dd4c68d321f056bfdc4ad560"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:data/figure_index.json",
           "exists": true,
           "bytes": 13434,
-          "sha256": "fc7c9d504509bea2bab08238c96b745f00fdd6a016019d121d4bf0e22b5f3936"
+          "sha256": "1a28548d2838b5a5ea9464761ccb73a3e55d6d48dd4c68d321f056bfdc4ad560"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:docs/data/figure_index.json",
           "exists": true,
           "bytes": 13434,
-          "sha256": "fc7c9d504509bea2bab08238c96b745f00fdd6a016019d121d4bf0e22b5f3936"
+          "sha256": "1a28548d2838b5a5ea9464761ccb73a3e55d6d48dd4c68d321f056bfdc4ad560"
         },
         "hf_model": {
           "path": "hf_model:metrics/figure_index.json",
           "exists": true,
           "bytes": 13434,
-          "sha256": "fc7c9d504509bea2bab08238c96b745f00fdd6a016019d121d4bf0e22b5f3936"
+          "sha256": "1a28548d2838b5a5ea9464761ccb73a3e55d6d48dd4c68d321f056bfdc4ad560"
         }
       },
       "failures": []
@@ -350,27 +381,27 @@
       "local": {
         "path": "repo:docs/data/project_status.json",
         "exists": true,
-        "bytes": 7340,
-        "sha256": "3f54a794582f4b61f60fe240a9ea6f1782e29a21498e2d63dad670495d4e0d39"
+        "bytes": 8162,
+        "sha256": "19471dbaf46a16275b8f0deaf67a5326b9508adc911a5c6d3426a2565049dd46"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:data/project_status.json",
           "exists": true,
-          "bytes": 7340,
-          "sha256": "3f54a794582f4b61f60fe240a9ea6f1782e29a21498e2d63dad670495d4e0d39"
+          "bytes": 8162,
+          "sha256": "19471dbaf46a16275b8f0deaf67a5326b9508adc911a5c6d3426a2565049dd46"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:docs/data/project_status.json",
           "exists": true,
-          "bytes": 7340,
-          "sha256": "3f54a794582f4b61f60fe240a9ea6f1782e29a21498e2d63dad670495d4e0d39"
+          "bytes": 8162,
+          "sha256": "19471dbaf46a16275b8f0deaf67a5326b9508adc911a5c6d3426a2565049dd46"
         },
         "hf_model": {
           "path": "hf_model:metrics/project_status.json",
           "exists": true,
-          "bytes": 7340,
-          "sha256": "3f54a794582f4b61f60fe240a9ea6f1782e29a21498e2d63dad670495d4e0d39"
+          "bytes": 8162,
+          "sha256": "19471dbaf46a16275b8f0deaf67a5326b9508adc911a5c6d3426a2565049dd46"
         }
       },
       "failures": []
@@ -382,26 +413,26 @@
         "path": "repo:docs/data/publication_audit.json",
         "exists": true,
         "bytes": 7289,
-        "sha256": "b2292107e5472a74e5b78dc769d2ff2739c41d7773367d9868b9d57db99a77e1"
+        "sha256": "a236625afaf97b2595e0aeeeab6687f46564d107d6bda3caac7be70d84fc1c38"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:data/publication_audit.json",
           "exists": true,
           "bytes": 7289,
-          "sha256": "b2292107e5472a74e5b78dc769d2ff2739c41d7773367d9868b9d57db99a77e1"
+          "sha256": "a236625afaf97b2595e0aeeeab6687f46564d107d6bda3caac7be70d84fc1c38"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:docs/data/publication_audit.json",
           "exists": true,
           "bytes": 7289,
-          "sha256": "b2292107e5472a74e5b78dc769d2ff2739c41d7773367d9868b9d57db99a77e1"
+          "sha256": "a236625afaf97b2595e0aeeeab6687f46564d107d6bda3caac7be70d84fc1c38"
         },
         "hf_model": {
           "path": "hf_model:metrics/publication_audit.json",
           "exists": true,
           "bytes": 7289,
-          "sha256": "b2292107e5472a74e5b78dc769d2ff2739c41d7773367d9868b9d57db99a77e1"
+          "sha256": "a236625afaf97b2595e0aeeeab6687f46564d107d6bda3caac7be70d84fc1c38"
         }
       },
       "failures": []
@@ -412,27 +443,27 @@
       "local": {
         "path": "repo:docs/data/public_surface_qa.json",
         "exists": true,
-        "bytes": 5647,
-        "sha256": "46576b574f962767b3d52ffd7075659d72ddf3d11154e47997fe7564de1731cb"
+        "bytes": 5648,
+        "sha256": "69b4da08a2603d9d3699125510c9bbd9982a7bf0f2f3236ee5472f60dfe04f5e"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:data/public_surface_qa.json",
           "exists": true,
-          "bytes": 5647,
-          "sha256": "46576b574f962767b3d52ffd7075659d72ddf3d11154e47997fe7564de1731cb"
+          "bytes": 5648,
+          "sha256": "69b4da08a2603d9d3699125510c9bbd9982a7bf0f2f3236ee5472f60dfe04f5e"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:docs/data/public_surface_qa.json",
           "exists": true,
-          "bytes": 5647,
-          "sha256": "46576b574f962767b3d52ffd7075659d72ddf3d11154e47997fe7564de1731cb"
+          "bytes": 5648,
+          "sha256": "69b4da08a2603d9d3699125510c9bbd9982a7bf0f2f3236ee5472f60dfe04f5e"
         },
         "hf_model": {
           "path": "hf_model:metrics/public_surface_qa.json",
           "exists": true,
-          "bytes": 5647,
-          "sha256": "46576b574f962767b3d52ffd7075659d72ddf3d11154e47997fe7564de1731cb"
+          "bytes": 5648,
+          "sha256": "69b4da08a2603d9d3699125510c9bbd9982a7bf0f2f3236ee5472f60dfe04f5e"
         }
       },
       "failures": []
@@ -598,27 +629,27 @@
       "local": {
         "path": "repo:docs/data/research_takeaways.json",
         "exists": true,
-        "bytes": 5356,
-        "sha256": "0621f28d602582c70524cd9455b01558e7003267b1f4786d005e31597e46b051"
+        "bytes": 6814,
+        "sha256": "538ec80950dbf5d77cc6ab2bd5a96a3af191b56a257b895e48b6cf28deb0d044"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:data/research_takeaways.json",
           "exists": true,
-          "bytes": 5356,
-          "sha256": "0621f28d602582c70524cd9455b01558e7003267b1f4786d005e31597e46b051"
+          "bytes": 6814,
+          "sha256": "538ec80950dbf5d77cc6ab2bd5a96a3af191b56a257b895e48b6cf28deb0d044"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:docs/data/research_takeaways.json",
           "exists": true,
-          "bytes": 5356,
-          "sha256": "0621f28d602582c70524cd9455b01558e7003267b1f4786d005e31597e46b051"
+          "bytes": 6814,
+          "sha256": "538ec80950dbf5d77cc6ab2bd5a96a3af191b56a257b895e48b6cf28deb0d044"
         },
         "hf_model": {
           "path": "hf_model:metrics/research_takeaways.json",
           "exists": true,
-          "bytes": 5356,
-          "sha256": "0621f28d602582c70524cd9455b01558e7003267b1f4786d005e31597e46b051"
+          "bytes": 6814,
+          "sha256": "538ec80950dbf5d77cc6ab2bd5a96a3af191b56a257b895e48b6cf28deb0d044"
         }
       },
       "failures": []
@@ -692,26 +723,26 @@
         "path": "repo:docs/data/scope_claims_audit.json",
         "exists": true,
         "bytes": 20066,
-        "sha256": "7ea0cb11edab7c6622888178986c8bc57a45f088bf2eefb4a5ae52fa66f3c5de"
+        "sha256": "263dda0848f698c9a9724f03a3d8263f25134e1d0e4f5273b8ae12dc25e5f271"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:data/scope_claims_audit.json",
           "exists": true,
           "bytes": 20066,
-          "sha256": "7ea0cb11edab7c6622888178986c8bc57a45f088bf2eefb4a5ae52fa66f3c5de"
+          "sha256": "263dda0848f698c9a9724f03a3d8263f25134e1d0e4f5273b8ae12dc25e5f271"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:docs/data/scope_claims_audit.json",
           "exists": true,
           "bytes": 20066,
-          "sha256": "7ea0cb11edab7c6622888178986c8bc57a45f088bf2eefb4a5ae52fa66f3c5de"
+          "sha256": "263dda0848f698c9a9724f03a3d8263f25134e1d0e4f5273b8ae12dc25e5f271"
         },
         "hf_model": {
           "path": "hf_model:metrics/scope_claims_audit.json",
           "exists": true,
           "bytes": 20066,
-          "sha256": "7ea0cb11edab7c6622888178986c8bc57a45f088bf2eefb4a5ae52fa66f3c5de"
+          "sha256": "263dda0848f698c9a9724f03a3d8263f25134e1d0e4f5273b8ae12dc25e5f271"
         }
       },
       "failures": []
@@ -754,26 +785,26 @@
         "path": "repo:docs/data/source_alignment_audit.json",
         "exists": true,
         "bytes": 4432,
-        "sha256": "9d8eb05aa29f898a28f64b0b23ef0b912ba0f43d6ef558d213e05619da0b3197"
+        "sha256": "ab8728f1f7539dd8e9d5a429e1bb01858d701412ab690ccd2f49f39ecbb645d9"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:data/source_alignment_audit.json",
           "exists": true,
           "bytes": 4432,
-          "sha256": "9d8eb05aa29f898a28f64b0b23ef0b912ba0f43d6ef558d213e05619da0b3197"
+          "sha256": "ab8728f1f7539dd8e9d5a429e1bb01858d701412ab690ccd2f49f39ecbb645d9"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:docs/data/source_alignment_audit.json",
           "exists": true,
           "bytes": 4432,
-          "sha256": "9d8eb05aa29f898a28f64b0b23ef0b912ba0f43d6ef558d213e05619da0b3197"
+          "sha256": "ab8728f1f7539dd8e9d5a429e1bb01858d701412ab690ccd2f49f39ecbb645d9"
         },
         "hf_model": {
           "path": "hf_model:metrics/source_alignment_audit.json",
           "exists": true,
           "bytes": 4432,
-          "sha256": "9d8eb05aa29f898a28f64b0b23ef0b912ba0f43d6ef558d213e05619da0b3197"
+          "sha256": "ab8728f1f7539dd8e9d5a429e1bb01858d701412ab690ccd2f49f39ecbb645d9"
         }
       },
       "failures": []
@@ -816,26 +847,26 @@
         "path": "repo:docs/data/task_surface_integrity.json",
         "exists": true,
         "bytes": 45779,
-        "sha256": "8dfb556fc81d931bcaf547c52442e39c5f0579c8cc6508afa117c15aa879e8cb"
+        "sha256": "86031c161ee0163151d1780590f8f1caeead52414fe489f1f433ca26acb4c17b"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:data/task_surface_integrity.json",
           "exists": true,
           "bytes": 45779,
-          "sha256": "8dfb556fc81d931bcaf547c52442e39c5f0579c8cc6508afa117c15aa879e8cb"
+          "sha256": "86031c161ee0163151d1780590f8f1caeead52414fe489f1f433ca26acb4c17b"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:docs/data/task_surface_integrity.json",
           "exists": true,
           "bytes": 45779,
-          "sha256": "8dfb556fc81d931bcaf547c52442e39c5f0579c8cc6508afa117c15aa879e8cb"
+          "sha256": "86031c161ee0163151d1780590f8f1caeead52414fe489f1f433ca26acb4c17b"
         },
         "hf_model": {
           "path": "hf_model:metrics/task_surface_integrity.json",
           "exists": true,
           "bytes": 45779,
-          "sha256": "8dfb556fc81d931bcaf547c52442e39c5f0579c8cc6508afa117c15aa879e8cb"
+          "sha256": "86031c161ee0163151d1780590f8f1caeead52414fe489f1f433ca26acb4c17b"
         }
       },
       "failures": []
@@ -877,27 +908,27 @@
       "local": {
         "path": "repo:docs/data/website_integrity.json",
         "exists": true,
-        "bytes": 14312,
-        "sha256": "5c86a8f27a05c1b0d9f8a407b73be2db1fbd88aafc64687f9f07b9bc4f397305"
+        "bytes": 14588,
+        "sha256": "04b249310051fe621e4ed4a607e999546b64e3c98a0d995e84c3088cd034c04c"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:data/website_integrity.json",
           "exists": true,
-          "bytes": 14312,
-          "sha256": "5c86a8f27a05c1b0d9f8a407b73be2db1fbd88aafc64687f9f07b9bc4f397305"
+          "bytes": 14588,
+          "sha256": "04b249310051fe621e4ed4a607e999546b64e3c98a0d995e84c3088cd034c04c"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:docs/data/website_integrity.json",
           "exists": true,
-          "bytes": 14312,
-          "sha256": "5c86a8f27a05c1b0d9f8a407b73be2db1fbd88aafc64687f9f07b9bc4f397305"
+          "bytes": 14588,
+          "sha256": "04b249310051fe621e4ed4a607e999546b64e3c98a0d995e84c3088cd034c04c"
         },
         "hf_model": {
           "path": "hf_model:metrics/website_integrity.json",
           "exists": true,
-          "bytes": 14312,
-          "sha256": "5c86a8f27a05c1b0d9f8a407b73be2db1fbd88aafc64687f9f07b9bc4f397305"
+          "bytes": 14588,
+          "sha256": "04b249310051fe621e4ed4a607e999546b64e3c98a0d995e84c3088cd034c04c"
         }
       },
       "failures": []
@@ -933,6 +964,43 @@
       },
       "failures": []
     },
+    {
+      "name": "assets/charts/audio_ablation_delta.svg",
+      "status": "pass",
+      "local": {
+        "path": "repo:docs/assets/charts/audio_ablation_delta.svg",
+        "exists": true,
+        "bytes": 4146,
+        "sha256": "187dbabe01f9ff18841ff61a1e7fbf85bebdd188cc0f248bb5090d64528e7568"
+      },
+      "mirrors": {
+        "hf_space": {
+          "path": "hf_space:assets/charts/audio_ablation_delta.svg",
+          "exists": true,
+          "bytes": 4146,
+          "sha256": "187dbabe01f9ff18841ff61a1e7fbf85bebdd188cc0f248bb5090d64528e7568"
+        },
+        "hf_artifacts_docs": {
+          "path": "hf_artifacts:docs/assets/charts/audio_ablation_delta.svg",
+          "exists": true,
+          "bytes": 4146,
+          "sha256": "187dbabe01f9ff18841ff61a1e7fbf85bebdd188cc0f248bb5090d64528e7568"
+        },
+        "hf_artifacts_card": {
+          "path": "hf_artifacts:assets/charts/audio_ablation_delta.svg",
+          "exists": true,
+          "bytes": 4146,
+          "sha256": "187dbabe01f9ff18841ff61a1e7fbf85bebdd188cc0f248bb5090d64528e7568"
+        },
+        "hf_model": {
+          "path": "hf_model:assets/charts/audio_ablation_delta.svg",
+          "exists": true,
+          "bytes": 4146,
+          "sha256": "187dbabe01f9ff18841ff61a1e7fbf85bebdd188cc0f248bb5090d64528e7568"
+        }
+      },
+      "failures": []
+    },
     {
       "name": "assets/brand/xperience10m-logo-apple-touch.png",
       "status": "pass",
@@ -1562,27 +1630,52 @@
       },
       "failures": []
     },
+    {
+      "name": "scripts/audio_ablation_and_raw_upgrade.py",
+      "status": "pass",
+      "local": {
+        "path": "repo:scripts/audio_ablation_and_raw_upgrade.py",
+        "exists": true,
+        "bytes": 43192,
+        "sha256": "73db45a3ddb50b60f61ca5e55bdf2008e48c268681bfcdc95d35bb56e48f89e3"
+      },
+      "mirrors": {
+        "hf_artifacts": {
+          "path": "hf_artifacts:scripts/audio_ablation_and_raw_upgrade.py",
+          "exists": true,
+          "bytes": 43192,
+          "sha256": "73db45a3ddb50b60f61ca5e55bdf2008e48c268681bfcdc95d35bb56e48f89e3"
+        },
+        "hf_model": {
+          "path": "hf_model:scripts/audio_ablation_and_raw_upgrade.py",
+          "exists": true,
+          "bytes": 43192,
+          "sha256": "73db45a3ddb50b60f61ca5e55bdf2008e48c268681bfcdc95d35bb56e48f89e3"
+        }
+      },
+      "failures": []
+    },
     {
       "name": "scripts/build_artifact_index.py",
       "status": "pass",
       "local": {
         "path": "repo:scripts/build_artifact_index.py",
         "exists": true,
-        "bytes": 24091,
-        "sha256": "f21aaf9267e93e571b093525525749054bee85872506f01ce534eb4fa131a56c"
+        "bytes": 25927,
+        "sha256": "4a9796faf0cd87b052a69f5fde89a7a3a88008f8e4485c05f9f5cc7630a89164"
       },
       "mirrors": {
         "hf_artifacts": {
           "path": "hf_artifacts:scripts/build_artifact_index.py",
           "exists": true,
-          "bytes": 24091,
-          "sha256": "f21aaf9267e93e571b093525525749054bee85872506f01ce534eb4fa131a56c"
+          "bytes": 25927,
+          "sha256": "4a9796faf0cd87b052a69f5fde89a7a3a88008f8e4485c05f9f5cc7630a89164"
         },
         "hf_model": {
           "path": "hf_model:scripts/build_artifact_index.py",
           "exists": true,
-          "bytes": 24091,
-          "sha256": "f21aaf9267e93e571b093525525749054bee85872506f01ce534eb4fa131a56c"
+          "bytes": 25927,
+          "sha256": "4a9796faf0cd87b052a69f5fde89a7a3a88008f8e4485c05f9f5cc7630a89164"
         }
       },
       "failures": []
@@ -1793,21 +1886,21 @@
       "local": {
         "path": "repo:scripts/build_research_takeaways.py",
         "exists": true,
-        "bytes": 10709,
-        "sha256": "dfb6cc169b5a5a7e0b2f3d2aeb604fd0eeb586e97cfb114a0fe32844c7ff8617"
+        "bytes": 13304,
+        "sha256": "135f555b17aae5dc5965ec88c3db6e49622a480467909ffa10d53278d4ea054b"
       },
       "mirrors": {
         "hf_artifacts": {
           "path": "hf_artifacts:scripts/build_research_takeaways.py",
           "exists": true,
-          "bytes": 10709,
-          "sha256": "dfb6cc169b5a5a7e0b2f3d2aeb604fd0eeb586e97cfb114a0fe32844c7ff8617"
+          "bytes": 13304,
+          "sha256": "135f555b17aae5dc5965ec88c3db6e49622a480467909ffa10d53278d4ea054b"
         },
         "hf_model": {
           "path": "hf_model:scripts/build_research_takeaways.py",
           "exists": true,
-          "bytes": 10709,
-          "sha256": "dfb6cc169b5a5a7e0b2f3d2aeb604fd0eeb586e97cfb114a0fe32844c7ff8617"
+          "bytes": 13304,
+          "sha256": "135f555b17aae5dc5965ec88c3db6e49622a480467909ffa10d53278d4ea054b"
         }
       },
       "failures": []
@@ -1868,21 +1961,21 @@
       "local": {
         "path": "repo:scripts/validate_mirror_parity.py",
         "exists": true,
-        "bytes": 12057,
-        "sha256": "e2d87770613a36c618bc16c4d72e08064c397eef5860594f04f501446d54ed65"
+        "bytes": 12444,
+        "sha256": "51cb547c4eaaa3999e18aa728fb36ca2a383c363ade0ee2c332b688515838f65"
       },
       "mirrors": {
         "hf_artifacts": {
           "path": "hf_artifacts:scripts/validate_mirror_parity.py",
           "exists": true,
-          "bytes": 12057,
-          "sha256": "e2d87770613a36c618bc16c4d72e08064c397eef5860594f04f501446d54ed65"
+          "bytes": 12444,
+          "sha256": "51cb547c4eaaa3999e18aa728fb36ca2a383c363ade0ee2c332b688515838f65"
         },
         "hf_model": {
           "path": "hf_model:scripts/validate_mirror_parity.py",
           "exists": true,
-          "bytes": 12057,
-          "sha256": "e2d87770613a36c618bc16c4d72e08064c397eef5860594f04f501446d54ed65"
+          "bytes": 12444,
+          "sha256": "51cb547c4eaaa3999e18aa728fb36ca2a383c363ade0ee2c332b688515838f65"
         }
       },
       "failures": []
@@ -2018,21 +2111,21 @@
       "local": {
         "path": "repo:scripts/publish_hf_bundles.py",
         "exists": true,
-        "bytes": 8996,
-        "sha256": "9790951493763e6490d01fec62661df2bbe2a3341466fe41298b95cda0900229"
+        "bytes": 9997,
+        "sha256": "b1482905fac18e598337b2f42028f7ffa46ba71a28726e473dc7cd2af1c70f9c"
       },
       "mirrors": {
         "hf_artifacts": {
           "path": "hf_artifacts:scripts/publish_hf_bundles.py",
           "exists": true,
-          "bytes": 8996,
-          "sha256": "9790951493763e6490d01fec62661df2bbe2a3341466fe41298b95cda0900229"
+          "bytes": 9997,
+          "sha256": "b1482905fac18e598337b2f42028f7ffa46ba71a28726e473dc7cd2af1c70f9c"
         },
         "hf_model": {
           "path": "hf_model:scripts/publish_hf_bundles.py",
           "exists": true,
-          "bytes": 8996,
-          "sha256": "9790951493763e6490d01fec62661df2bbe2a3341466fe41298b95cda0900229"
+          "bytes": 9997,
+          "sha256": "b1482905fac18e598337b2f42028f7ffa46ba71a28726e473dc7cd2af1c70f9c"
         }
       },
       "failures": []
@@ -2093,21 +2186,21 @@
       "local": {
         "path": "repo:docs/index.html",
         "exists": true,
-        "bytes": 166238,
-        "sha256": "4ccdf2d1867793cfa64ff7475ff091300e135935140bd65b3277ac999a317747"
+        "bytes": 167864,
+        "sha256": "871eb24eceb753bda1fc6ba61518ab43c1a076230bf068adf05178081c863ded"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:index.html",
           "exists": true,
-          "bytes": 166238,
-          "sha256": "4ccdf2d1867793cfa64ff7475ff091300e135935140bd65b3277ac999a317747"
+          "bytes": 167864,
+          "sha256": "871eb24eceb753bda1fc6ba61518ab43c1a076230bf068adf05178081c863ded"
         },
         "hf_artifacts_docs": {
           "path": "hf_artifacts:docs/index.html",
           "exists": true,
-          "bytes": 166238,
-          "sha256": "4ccdf2d1867793cfa64ff7475ff091300e135935140bd65b3277ac999a317747"
+          "bytes": 167864,
+          "sha256": "871eb24eceb753bda1fc6ba61518ab43c1a076230bf068adf05178081c863ded"
         }
       },
       "failures": []
@@ -2187,6 +2280,161 @@
       },
       "failures": []
     },
+    {
+      "name": "results/audio_ablation/AUDIO_ABLATION_SUMMARY.md",
+      "status": "pass",
+      "local": {
+        "path": "repo:results/audio_ablation/AUDIO_ABLATION_SUMMARY.md",
+        "exists": true,
+        "bytes": 2166,
+        "sha256": "9318bc55b1b051c384ce2fd6e7cd0a56becfa00185a0b597fdc474a40b1f17f9"
+      },
+      "mirrors": {
+        "hf_space": {
+          "path": "hf_space:results/audio_ablation/AUDIO_ABLATION_SUMMARY.md",
+          "exists": true,
+          "bytes": 2166,
+          "sha256": "9318bc55b1b051c384ce2fd6e7cd0a56becfa00185a0b597fdc474a40b1f17f9"
+        },
+        "hf_artifacts": {
+          "path": "hf_artifacts:results/audio_ablation/AUDIO_ABLATION_SUMMARY.md",
+          "exists": true,
+          "bytes": 2166,
+          "sha256": "9318bc55b1b051c384ce2fd6e7cd0a56becfa00185a0b597fdc474a40b1f17f9"
+        },
+        "hf_model": {
+          "path": "hf_model:results/audio_ablation/AUDIO_ABLATION_SUMMARY.md",
+          "exists": true,
+          "bytes": 2166,
+          "sha256": "9318bc55b1b051c384ce2fd6e7cd0a56becfa00185a0b597fdc474a40b1f17f9"
+        }
+      },
+      "failures": []
+    },
+    {
+      "name": "results/audio_ablation/audio_ablation_metrics.csv",
+      "status": "pass",
+      "local": {
+        "path": "repo:results/audio_ablation/audio_ablation_metrics.csv",
+        "exists": true,
+        "bytes": 24446,
+        "sha256": "674efea54e1be2c835d6c473c9df54ef02fc34e5f2be37728722d914892c8854"
+      },
+      "mirrors": {
+        "hf_space": {
+          "path": "hf_space:results/audio_ablation/audio_ablation_metrics.csv",
+          "exists": true,
+          "bytes": 24446,
+          "sha256": "674efea54e1be2c835d6c473c9df54ef02fc34e5f2be37728722d914892c8854"
+        },
+        "hf_artifacts": {
+          "path": "hf_artifacts:results/audio_ablation/audio_ablation_metrics.csv",
+          "exists": true,
+          "bytes": 24446,
+          "sha256": "674efea54e1be2c835d6c473c9df54ef02fc34e5f2be37728722d914892c8854"
+        },
+        "hf_model": {
+          "path": "hf_model:results/audio_ablation/audio_ablation_metrics.csv",
+          "exists": true,
+          "bytes": 24446,
+          "sha256": "674efea54e1be2c835d6c473c9df54ef02fc34e5f2be37728722d914892c8854"
+        }
+      },
+      "failures": []
+    },
+    {
+      "name": "results/audio_ablation/audio_ablation_summary.json",
+      "status": "pass",
+      "local": {
+        "path": "repo:results/audio_ablation/audio_ablation_summary.json",
+        "exists": true,
+        "bytes": 9735,
+        "sha256": "514370c53baeb51373160bbd5ab7f5f0de7733301bf1b82981ed799aec41152a"
+      },
+      "mirrors": {
+        "hf_space": {
+          "path": "hf_space:results/audio_ablation/audio_ablation_summary.json",
+          "exists": true,
+          "bytes": 9735,
+          "sha256": "514370c53baeb51373160bbd5ab7f5f0de7733301bf1b82981ed799aec41152a"
+        },
+        "hf_artifacts": {
+          "path": "hf_artifacts:results/audio_ablation/audio_ablation_summary.json",
+          "exists": true,
+          "bytes": 9735,
+          "sha256": "514370c53baeb51373160bbd5ab7f5f0de7733301bf1b82981ed799aec41152a"
+        },
+        "hf_model": {
+          "path": "hf_model:results/audio_ablation/audio_ablation_summary.json",
+          "exists": true,
+          "bytes": 9735,
+          "sha256": "514370c53baeb51373160bbd5ab7f5f0de7733301bf1b82981ed799aec41152a"
+        }
+      },
+      "failures": []
+    },
+    {
+      "name": "results/audio_ablation/audio_delta_summary.csv",
+      "status": "pass",
+      "local": {
+        "path": "repo:results/audio_ablation/audio_delta_summary.csv",
+        "exists": true,
+        "bytes": 2630,
+        "sha256": "55638ba2387085e3d21fd58be05de060da572ae009b8019926ab4e1733540f91"
+      },
+      "mirrors": {
+        "hf_space": {
+          "path": "hf_space:results/audio_ablation/audio_delta_summary.csv",
+          "exists": true,
+          "bytes": 2630,
+          "sha256": "55638ba2387085e3d21fd58be05de060da572ae009b8019926ab4e1733540f91"
+        },
+        "hf_artifacts": {
+          "path": "hf_artifacts:results/audio_ablation/audio_delta_summary.csv",
+          "exists": true,
+          "bytes": 2630,
+          "sha256": "55638ba2387085e3d21fd58be05de060da572ae009b8019926ab4e1733540f91"
+        },
+        "hf_model": {
+          "path": "hf_model:results/audio_ablation/audio_delta_summary.csv",
+          "exists": true,
+          "bytes": 2630,
+          "sha256": "55638ba2387085e3d21fd58be05de060da572ae009b8019926ab4e1733540f91"
+        }
+      },
+      "failures": []
+    },
+    {
+      "name": "results/audio_ablation/raw_logmel_fisheye_cam0_sr16000_mels64_fft512_hop160.npz",
+      "status": "pass",
+      "local": {
+        "path": "repo:results/audio_ablation/raw_logmel_fisheye_cam0_sr16000_mels64_fft512_hop160.npz",
+        "exists": true,
+        "bytes": 2363664,
+        "sha256": "6b61a3db5d21c89ba8cfb66b94006cb48640c5ba54d7d6ccd8835d7942534fa1"
+      },
+      "mirrors": {
+        "hf_space": {
+          "path": "hf_space:results/audio_ablation/raw_logmel_fisheye_cam0_sr16000_mels64_fft512_hop160.npz",
+          "exists": true,
+          "bytes": 2363664,
+          "sha256": "6b61a3db5d21c89ba8cfb66b94006cb48640c5ba54d7d6ccd8835d7942534fa1"
+        },
+        "hf_artifacts": {
+          "path": "hf_artifacts:results/audio_ablation/raw_logmel_fisheye_cam0_sr16000_mels64_fft512_hop160.npz",
+          "exists": true,
+          "bytes": 2363664,
+          "sha256": "6b61a3db5d21c89ba8cfb66b94006cb48640c5ba54d7d6ccd8835d7942534fa1"
+        },
+        "hf_model": {
+          "path": "hf_model:results/audio_ablation/raw_logmel_fisheye_cam0_sr16000_mels64_fft512_hop160.npz",
+          "exists": true,
+          "bytes": 2363664,
+          "sha256": "6b61a3db5d21c89ba8cfb66b94006cb48640c5ba54d7d6ccd8835d7942534fa1"
+        }
+      },
+      "failures": []
+    },
     {
       "name": "results/single_episode_diagnostics/provenance.json",
       "status": "pass",
@@ -2658,27 +2906,27 @@
       "local": {
         "path": "repo:PROJECT_STATUS.md",
         "exists": true,
-        "bytes": 5384,
-        "sha256": "3af38f7eefee59b48dfbdfb316f75f428e06fb2a2a5d37efa798ee3e2dae53f0"
+        "bytes": 6031,
+        "sha256": "b2de2d9de04e0a05046cbea8a0b7c2e8766f6171b5c80a1b3f244aa1c51023ea"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:PROJECT_STATUS.md",
           "exists": true,
-          "bytes": 5384,
-          "sha256": "3af38f7eefee59b48dfbdfb316f75f428e06fb2a2a5d37efa798ee3e2dae53f0"
+          "bytes": 6031,
+          "sha256": "b2de2d9de04e0a05046cbea8a0b7c2e8766f6171b5c80a1b3f244aa1c51023ea"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:PROJECT_STATUS.md",
           "exists": true,
-          "bytes": 5384,
-          "sha256": "3af38f7eefee59b48dfbdfb316f75f428e06fb2a2a5d37efa798ee3e2dae53f0"
+          "bytes": 6031,
+          "sha256": "b2de2d9de04e0a05046cbea8a0b7c2e8766f6171b5c80a1b3f244aa1c51023ea"
         },
         "hf_model": {
           "path": "hf_model:PROJECT_STATUS.md",
           "exists": true,
-          "bytes": 5384,
-          "sha256": "3af38f7eefee59b48dfbdfb316f75f428e06fb2a2a5d37efa798ee3e2dae53f0"
+          "bytes": 6031,
+          "sha256": "b2de2d9de04e0a05046cbea8a0b7c2e8766f6171b5c80a1b3f244aa1c51023ea"
         }
       },
       "failures": []
@@ -2720,27 +2968,27 @@
       "local": {
         "path": "repo:RESEARCH_TAKEAWAYS.md",
         "exists": true,
-        "bytes": 3856,
-        "sha256": "3b575d4b47ba4e09f5e16d065dab78194148eb52394b30f588930237e0518f0d"
+        "bytes": 4893,
+        "sha256": "16a002167d0702ea7e85ad1a494800daad4ff080c3751e078275f7c196721046"
       },
       "mirrors": {
         "hf_space": {
           "path": "hf_space:RESEARCH_TAKEAWAYS.md",
           "exists": true,
-          "bytes": 3856,
-          "sha256": "3b575d4b47ba4e09f5e16d065dab78194148eb52394b30f588930237e0518f0d"
+          "bytes": 4893,
+          "sha256": "16a002167d0702ea7e85ad1a494800daad4ff080c3751e078275f7c196721046"
         },
         "hf_artifacts": {
           "path": "hf_artifacts:RESEARCH_TAKEAWAYS.md",
           "exists": true,
-          "bytes": 3856,
-          "sha256": "3b575d4b47ba4e09f5e16d065dab78194148eb52394b30f588930237e0518f0d"
+          "bytes": 4893,
+          "sha256": "16a002167d0702ea7e85ad1a494800daad4ff080c3751e078275f7c196721046"
         },
         "hf_model": {
           "path": "hf_model:RESEARCH_TAKEAWAYS.md",
           "exists": true,
-          "bytes": 3856,
-          "sha256": "3b575d4b47ba4e09f5e16d065dab78194148eb52394b30f588930237e0518f0d"
+          "bytes": 4893,
+          "sha256": "16a002167d0702ea7e85ad1a494800daad4ff080c3751e078275f7c196721046"
         }
       },
       "failures": []

docs/data/project_status.json

@@ -44,6 +44,16 @@
             ],
             "readout": "Each task also has a compact PyTorch MLP run over the same feature tensor and chronological split."
         },
+        {
+            "area": "Audio ablation and raw-audio upgrade",
+            "status": "verified",
+            "evidence": [
+                "scripts/audio_ablation_and_raw_upgrade.py",
+                "results/audio_ablation/",
+                "docs/data/audio_ablation_summary.json"
+            ],
+            "readout": "Current AAC audio improves the primary metric on 6 of 12 task contracts; replacing the current handcrafted block with a 588-d raw log-mel feature improves over current audio on 6 of 12 tasks."
+        },
         {
             "area": "Evaluation protocol",
             "status": "verified",
@@ -147,6 +157,7 @@
         "Inspect RESEARCH_TAKEAWAYS.md and docs/data/research_takeaways.json before interpreting model scores.",
         "Inspect RESEARCH_ROADMAP.md and docs/data/research_roadmap.json for the staged path from public-sample task work to multi-episode modeling.",
         "Inspect docs/data/summary_metrics.json and results/episode_task_suite/neural_mlp/ to check the 12-task outputs.",
+        "Inspect results/audio_ablation/AUDIO_ABLATION_SUMMARY.md before judging whether audio helps the current task suite.",
         "Inspect EVALUATION_PROTOCOL.md before judging task metrics or leakage controls.",
         "Inspect SOURCE_ALIGNMENT_AUDIT.md before judging source-card consistency across public surfaces.",
         "Inspect XPERIENCE10M_DATASET_CARD_ALIGNMENT.md before judging dataset wording.",
@@ -156,6 +167,7 @@
     "Cross-episode generalization is evaluated in the later multi-episode stage.",
     "Historical 32ep path names refer to setup files, not completed 32-episode training results.",
     "The current reconstruction task reconstructs feature vectors, not pixel-depth, mesh, NeRF, or Gaussian reconstruction.",
-    "AAC audio is decoded from fisheye_cam0.mp4 and included in the current 8,546-dimensional baseline feature vector."
+    "AAC audio is decoded from fisheye_cam0.mp4 and included in the current 8,546-dimensional baseline feature vector.",
+    "Audio is now evaluated directly: the current AAC block and a raw log-mel replacement are compared across all 12 task contracts in results/audio_ablation/."
   ]
 }

docs/data/public_surface_qa.json

@@ -1,7 +1,7 @@
 {
   "title": "Ropedia Xperience-10M Public Project Surface",
   "status": "pass",
-  "generated_at_utc": "2026-06-03T13:04:23+00:00",
+  "generated_at_utc": "2026-06-03T14:17:21+00:00",
   "scope": "Repo README, GitHub Pages HTML, Hugging Face Space card, artifact dataset card, and model card.",
   "checks": [
     {
@@ -18,7 +18,7 @@
         "website_integrity": {
           "exists": true,
           "status": "pass",
-          "generated_at_utc": "2026-06-03T07:49:53+00:00"
+          "generated_at_utc": "2026-06-03T13:07:43+00:00"
         },
         "rendered_site_check": {
           "exists": true,
@@ -28,27 +28,27 @@
         "task_surface_integrity": {
           "exists": true,
           "status": "pass",
-          "generated_at_utc": "2026-06-03T12:58:10+00:00"
+          "generated_at_utc": "2026-06-03T13:08:46+00:00"
         },
         "source_alignment": {
           "exists": true,
           "status": "pass",
-          "generated_at_utc": "2026-06-03T13:04:15+00:00"
+          "generated_at_utc": "2026-06-03T13:08:46+00:00"
         },
         "scale_up_status": {
           "exists": true,
           "status": "pass",
-          "generated_at_utc": "2026-06-03T01:38:40+00:00"
+          "generated_at_utc": "2026-06-03T13:07:42+00:00"
         },
         "publication_package": {
           "exists": true,
           "status": "pass",
-          "generated_at_utc": "2026-06-03T07:49:54+00:00"
+          "generated_at_utc": "2026-06-03T13:54:27+00:00"
         },
         "mirror_parity": {
           "exists": true,
           "status": "pass",
-          "generated_at_utc": "2026-06-03T07:52:34+00:00"
+          "generated_at_utc": "2026-06-03T13:54:27+00:00"
         },
         "live_publication": {
           "exists": true,
@@ -102,7 +102,7 @@
       "marker_counts": {
         "Ropedia Xperience-10M Task Suite": 14,
         "Xperience-10M": 93,
-        "12-task": 19,
+        "12-task": 20,
         "Qwen3-Omni": 37,
         "one public Xperience-10M sample episode": 1
       }
@@ -112,7 +112,7 @@
       "status": "pass",
       "reason": "Public cards should link the repo, Space, artifacts, model baselines, upstream dataset, and Ropedia dataset page.",
       "marker_counts": {
-        "https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite": 61,
+        "https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite": 63,
         "https://huggingface.co/spaces/cy0307/ropedia-xperience-10m-task-suite": 5,
         "https://huggingface.co/datasets/cy0307/ropedia-xperience-10m-task-suite-artifacts": 4,
         "https://huggingface.co/cy0307/ropedia-xperience-10m-task-baselines": 3,
@@ -127,11 +127,11 @@
       "marker_counts": {
         "data/project_brief.json": 7,
         "data/website_integrity.json": 7,
-        "data/rendered_site_check.json": 4,
-        "data/task_surface_integrity.json": 9,
+        "data/rendered_site_check.json": 5,
+        "data/task_surface_integrity.json": 10,
         "data/publication_audit.json": 8,
         "data/mirror_parity.json": 6,
-        "data/public_surface_qa.json": 5,
+        "data/public_surface_qa.json": 7,
         "data/research_roadmap.json": 10
       }
     },

docs/data/publication_audit.json

@@ -1,6 +1,6 @@
 {
   "status": "pass",
-  "generated_at_utc": "2026-06-03T13:54:27+00:00",
+  "generated_at_utc": "2026-06-03T14:25:37+00:00",
   "checks": [
     {
       "name": "required_publication_assets_present",
@@ -182,8 +182,8 @@
     "github_repo": {
       "root": "repo",
       "exists": true,
-      "file_count": 355,
-      "text_file_count": 291,
+      "file_count": 363,
+      "text_file_count": 298,
       "largest_file": {
         "path": "results/episode_task_suite/modality_reconstruction/predictions.npz",
         "bytes": 55702978
@@ -193,8 +193,8 @@
     "hf_space_bundle": {
       "root": "hf_publish/space",
       "exists": true,
-      "file_count": 304,
-      "text_file_count": 240,
+      "file_count": 311,
+      "text_file_count": 246,
       "largest_file": {
         "path": "results/episode_task_suite/modality_reconstruction/predictions.npz",
         "bytes": 55702978
@@ -204,8 +204,8 @@
     "hf_artifact_bundle": {
       "root": "hf_publish/artifacts",
       "exists": true,
-      "file_count": 384,
-      "text_file_count": 300,
+      "file_count": 393,
+      "text_file_count": 308,
       "largest_file": {
         "path": "results/episode_task_suite/modality_reconstruction/predictions.npz",
         "bytes": 55702978
@@ -215,8 +215,8 @@
     "hf_model_bundle": {
       "root": "hf_publish/model",
       "exists": true,
-      "file_count": 607,
-      "text_file_count": 486,
+      "file_count": 617,
+      "text_file_count": 495,
       "largest_file": {
         "path": "artifacts/episode_task_suite/modality_reconstruction/predictions.npz",
         "bytes": 55702978

docs/data/research_takeaways.json

@@ -1,11 +1,12 @@
 {
   "title": "Ropedia Xperience-10M Research Takeaways",
   "status": "pass",
-  "generated_at_utc": "2026-06-03T12:56:49+00:00",
+  "generated_at_utc": "2026-06-03T14:16:16+00:00",
   "source_files": [
     "docs/data/summary_metrics.json",
     "results/episode_task_suite/summary_report.json",
     "results/episode_task_suite/neural_mlp/*/metrics.json",
+    "docs/data/audio_ablation_summary.json",
     "results/omni_finetune/MULTI_EPISODE_ACCESS_STATUS.md"
   ],
   "scope": {
@@ -129,6 +130,39 @@
       "source": "results/episode_task_suite/cross_modal_retrieval/metrics.json",
       "current_scope": "The current reconstruction task predicts feature vectors; depth, mesh, NeRF, and Gaussian-splatting outputs are future task variants."
     },
+    {
+      "id": "audio_contribution_is_task_specific",
+      "title": "Audio helps some tasks and hurts others on the public sample",
+      "readout": "The current AAC audio block improves the primary metric on 6 of 12 tasks, while raw log-mel replacement improves over the current handcrafted block on 6 of 12 tasks. The largest current-audio gain appears in feature reconstruction, not in action classification.",
+      "evidence": [
+        {
+          "label": "tasks_where_current_audio_improves",
+          "value": 6
+        },
+        {
+          "label": "mean_current_audio_delta",
+          "value": 0.041849794979543296
+        },
+        {
+          "label": "tasks_where_raw_replacement_improves",
+          "value": 6
+        },
+        {
+          "label": "mean_raw_replacement_delta_vs_current",
+          "value": 0.09362598132150173
+        },
+        {
+          "label": "reconstruction_current_audio_delta",
+          "value": 0.6524486541748047
+        },
+        {
+          "label": "object_relevance_current_audio_delta",
+          "value": 0.010206249894598368
+        }
+      ],
+      "source": "results/audio_ablation/audio_ablation_summary.json",
+      "current_scope": "This is a single-episode ablation over fixed ridge heads. It validates that audio is wired into the task suite and shows where it changes metrics; it does not prove cross-episode audio generalization."
+    },
     {
       "id": "scale_requires_episodes",
       "title": "The next scientific unit is held-out episodes, not more adjacent windows",

docs/data/scope_claims_audit.json

@@ -1,6 +1,6 @@
 {
   "status": "pass",
-  "generated_at_utc": "2026-06-03T13:07:42+00:00",
+  "generated_at_utc": "2026-06-03T14:22:46+00:00",
   "summary": {
     "qwen3_omni_32_episode_claim": false,
     "dataset_manifest_num_episodes": 1,

docs/data/source_alignment_audit.json

@@ -1,7 +1,7 @@
 {
   "title": "Ropedia Xperience-10M Source Alignment Note",
   "status": "pass",
-  "generated_at_utc": "2026-06-03T13:08:46+00:00",
+  "generated_at_utc": "2026-06-03T14:23:21+00:00",
   "alignment_json": "docs/data/xperience10m_dataset_card_alignment.json",
   "alignment_summary": {
     "full_dataset_repo": "ropedia-ai/xperience-10m",

docs/data/task_surface_integrity.json

@@ -1,6 +1,6 @@
 {
   "status": "pass",
-  "generated_at_utc": "2026-06-03T13:08:46+00:00",
+  "generated_at_utc": "2026-06-03T14:23:21+00:00",
   "summary": {
     "task_count": 12,
     "expected_task_count": 12,
@@ -64,27 +64,27 @@
       "observed": "timeline_action"
     },
     {
-      "name": "timeline_action: public_field_research_name_is_human_readable",
+      "name": "timeline_action: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "Egocentric Action Recognition",
+      "value": "Recognize the current manipulation action from synchronized visual, motion, inertial, pose, and annotation context.",
       "raw_hits": []
     },
     {
-      "name": "timeline_action: public_field_output_short_is_human_readable",
+      "name": "timeline_action: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "current action class",
+      "value": "20-frame multimodal window",
       "raw_hits": []
     },
     {
-      "name": "timeline_action: public_field_plain_goal_is_human_readable",
+      "name": "timeline_action: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Look at one short multimodal window and name what action is happening now.",
+      "value": "current action class",
       "raw_hits": []
     },
     {
-      "name": "timeline_action: public_field_input_short_is_human_readable",
+      "name": "timeline_action: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "20-frame multimodal window",
+      "value": "window features -> action label builder -> classifier",
       "raw_hits": []
     },
     {
@@ -94,15 +94,15 @@
       "raw_hits": []
     },
     {
-      "name": "timeline_action: public_field_card_blurb_is_human_readable",
+      "name": "timeline_action: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Recognize the current manipulation action from synchronized visual, motion, inertial, pose, and annotation context.",
+      "value": "Egocentric Action Recognition",
       "raw_hits": []
     },
     {
-      "name": "timeline_action: public_field_process_short_is_human_readable",
+      "name": "timeline_action: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "window features -> action label builder -> classifier",
+      "value": "Look at one short multimodal window and name what action is happening now.",
       "raw_hits": []
     },
     {
@@ -184,27 +184,27 @@
       "observed": "timeline_subtask"
     },
     {
-      "name": "timeline_subtask: public_field_research_name_is_human_readable",
+      "name": "timeline_subtask: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "Temporal Subtask Recognition",
+      "value": "Recognize the broader activity stage so fine actions become a readable procedure timeline.",
       "raw_hits": []
     },
     {
-      "name": "timeline_subtask: public_field_output_short_is_human_readable",
+      "name": "timeline_subtask: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "current procedure step",
+      "value": "20-frame multimodal window",
       "raw_hits": []
     },
     {
-      "name": "timeline_subtask: public_field_plain_goal_is_human_readable",
+      "name": "timeline_subtask: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Predict the higher-level task stage for the current window.",
+      "value": "current procedure step",
       "raw_hits": []
     },
     {
-      "name": "timeline_subtask: public_field_input_short_is_human_readable",
+      "name": "timeline_subtask: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "20-frame multimodal window",
+      "value": "window features -> subtask label builder -> classifier",
       "raw_hits": []
     },
     {
@@ -214,15 +214,15 @@
       "raw_hits": []
     },
     {
-      "name": "timeline_subtask: public_field_card_blurb_is_human_readable",
+      "name": "timeline_subtask: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Recognize the broader activity stage so fine actions become a readable procedure timeline.",
+      "value": "Temporal Subtask Recognition",
       "raw_hits": []
     },
     {
-      "name": "timeline_subtask: public_field_process_short_is_human_readable",
+      "name": "timeline_subtask: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "window features -> subtask label builder -> classifier",
+      "value": "Predict the higher-level task stage for the current window.",
       "raw_hits": []
     },
     {
@@ -304,27 +304,27 @@
       "observed": "transition_detection"
     },
     {
-      "name": "transition_detection: public_field_research_name_is_human_readable",
+      "name": "transition_detection: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "Temporal Action Segmentation",
+      "value": "Detect the local moment where the episode changes from one action segment to the next.",
       "raw_hits": []
     },
     {
-      "name": "transition_detection: public_field_output_short_is_human_readable",
+      "name": "transition_detection: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "boundary or steady",
+      "value": "current window with boundary target",
       "raw_hits": []
     },
     {
-      "name": "transition_detection: public_field_plain_goal_is_human_readable",
+      "name": "transition_detection: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Detect whether the current window is near a boundary between actions.",
+      "value": "boundary or steady",
       "raw_hits": []
     },
     {
-      "name": "transition_detection: public_field_input_short_is_human_readable",
+      "name": "transition_detection: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "current window with boundary target",
+      "value": "action changes -> boundary labels -> binary classifier",
       "raw_hits": []
     },
     {
@@ -334,15 +334,15 @@
       "raw_hits": []
     },
     {
-      "name": "transition_detection: public_field_card_blurb_is_human_readable",
+      "name": "transition_detection: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Detect the local moment where the episode changes from one action segment to the next.",
+      "value": "Temporal Action Segmentation",
       "raw_hits": []
     },
     {
-      "name": "transition_detection: public_field_process_short_is_human_readable",
+      "name": "transition_detection: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "action changes -> boundary labels -> binary classifier",
+      "value": "Detect whether the current window is near a boundary between actions.",
       "raw_hits": []
     },
     {
@@ -422,27 +422,27 @@
       "observed": "next_action"
     },
     {
-      "name": "next_action: public_field_research_name_is_human_readable",
+      "name": "next_action: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "Short-Horizon Intention Prediction",
+      "value": "Forecast the near-future action from the current observations only.",
       "raw_hits": []
     },
     {
-      "name": "next_action: public_field_output_short_is_human_readable",
+      "name": "next_action: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "action at t+20 frames",
+      "value": "current window at time t",
       "raw_hits": []
     },
     {
-      "name": "next_action: public_field_plain_goal_is_human_readable",
+      "name": "next_action: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Use the current window to guess the action that will happen shortly after it.",
+      "value": "action at t+20 frames",
       "raw_hits": []
     },
     {
-      "name": "next_action: public_field_input_short_is_human_readable",
+      "name": "next_action: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "current window at time t",
+      "value": "current features -> future label shift -> classifier",
       "raw_hits": []
     },
     {
@@ -452,15 +452,15 @@
       "raw_hits": []
     },
     {
-      "name": "next_action: public_field_card_blurb_is_human_readable",
+      "name": "next_action: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Forecast the near-future action from the current observations only.",
+      "value": "Short-Horizon Intention Prediction",
       "raw_hits": []
     },
     {
-      "name": "next_action: public_field_process_short_is_human_readable",
+      "name": "next_action: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "current features -> future label shift -> classifier",
+      "value": "Use the current window to guess the action that will happen shortly after it.",
       "raw_hits": []
     },
     {
@@ -540,27 +540,27 @@
       "observed": "hand_trajectory_forecast"
     },
     {
-      "name": "hand_trajectory_forecast: public_field_research_name_is_human_readable",
+      "name": "hand_trajectory_forecast: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "3D Hand Motion Forecasting",
+      "value": "Predict the future 3D left/right hand path from the current multimodal state.",
       "raw_hits": []
     },
     {
-      "name": "hand_trajectory_forecast: public_field_output_short_is_human_readable",
+      "name": "hand_trajectory_forecast: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "future hand-joint trajectory",
+      "value": "current multimodal window",
       "raw_hits": []
     },
     {
-      "name": "hand_trajectory_forecast: public_field_plain_goal_is_human_readable",
+      "name": "hand_trajectory_forecast: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Predict where the hands will move over the next few frames.",
+      "value": "future hand-joint trajectory",
       "raw_hits": []
     },
     {
-      "name": "hand_trajectory_forecast: public_field_input_short_is_human_readable",
+      "name": "hand_trajectory_forecast: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "current multimodal window",
+      "value": "current features -> future mocap target -> regression head",
       "raw_hits": []
     },
     {
@@ -570,15 +570,15 @@
       "raw_hits": []
     },
     {
-      "name": "hand_trajectory_forecast: public_field_card_blurb_is_human_readable",
+      "name": "hand_trajectory_forecast: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Predict the future 3D left/right hand path from the current multimodal state.",
+      "value": "3D Hand Motion Forecasting",
       "raw_hits": []
     },
     {
-      "name": "hand_trajectory_forecast: public_field_process_short_is_human_readable",
+      "name": "hand_trajectory_forecast: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "current features -> future mocap target -> regression head",
+      "value": "Predict where the hands will move over the next few frames.",
       "raw_hits": []
     },
     {
@@ -658,27 +658,27 @@
       "observed": "contact_prediction"
     },
     {
-      "name": "contact_prediction: public_field_research_name_is_human_readable",
+      "name": "contact_prediction: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "Human-Object Contact Prediction",
+      "value": "Predict whether body or hand contact with the scene is occurring without leaking contact labels.",
       "raw_hits": []
     },
     {
-      "name": "contact_prediction: public_field_output_short_is_human_readable",
+      "name": "contact_prediction: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "contact or no contact",
+      "value": "non-contact, non-caption features",
       "raw_hits": []
     },
     {
-      "name": "contact_prediction: public_field_plain_goal_is_human_readable",
+      "name": "contact_prediction: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Predict whether the body or hand is in contact with something.",
+      "value": "contact or no contact",
       "raw_hits": []
     },
     {
-      "name": "contact_prediction: public_field_input_short_is_human_readable",
+      "name": "contact_prediction: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "non-contact, non-caption features",
+      "value": "feature filter -> contact target -> binary classifier",
       "raw_hits": []
     },
     {
@@ -688,15 +688,15 @@
       "raw_hits": []
     },
     {
-      "name": "contact_prediction: public_field_card_blurb_is_human_readable",
+      "name": "contact_prediction: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Predict whether body or hand contact with the scene is occurring without leaking contact labels.",
+      "value": "Human-Object Contact Prediction",
       "raw_hits": []
     },
     {
-      "name": "contact_prediction: public_field_process_short_is_human_readable",
+      "name": "contact_prediction: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "feature filter -> contact target -> binary classifier",
+      "value": "Predict whether the body or hand is in contact with something.",
       "raw_hits": []
     },
     {
@@ -774,27 +774,27 @@
       "observed": "object_relevance"
     },
     {
-      "name": "object_relevance: public_field_research_name_is_human_readable",
+      "name": "object_relevance: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "Object-Centric Interaction Recognition",
+      "value": "Infer which objects are relevant to the current manipulation window from non-caption features.",
       "raw_hits": []
     },
     {
-      "name": "object_relevance: public_field_output_short_is_human_readable",
+      "name": "object_relevance: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "relevant object set",
+      "value": "non-caption multimodal features",
       "raw_hits": []
     },
     {
-      "name": "object_relevance: public_field_plain_goal_is_human_readable",
+      "name": "object_relevance: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Predict which objects matter in the current window.",
+      "value": "relevant object set",
       "raw_hits": []
     },
     {
-      "name": "object_relevance: public_field_input_short_is_human_readable",
+      "name": "object_relevance: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "non-caption multimodal features",
+      "value": "object vocabulary -> multi-hot labels -> sigmoid heads",
       "raw_hits": []
     },
     {
@@ -804,15 +804,15 @@
       "raw_hits": []
     },
     {
-      "name": "object_relevance: public_field_card_blurb_is_human_readable",
+      "name": "object_relevance: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Infer which objects are relevant to the current manipulation window from non-caption features.",
+      "value": "Object-Centric Interaction Recognition",
       "raw_hits": []
     },
     {
-      "name": "object_relevance: public_field_process_short_is_human_readable",
+      "name": "object_relevance: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "object vocabulary -> multi-hot labels -> sigmoid heads",
+      "value": "Predict which objects matter in the current window.",
       "raw_hits": []
     },
     {
@@ -892,27 +892,27 @@
       "observed": "caption_grounding"
     },
     {
-      "name": "caption_grounding: public_field_research_name_is_human_readable",
+      "name": "caption_grounding: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "Language-to-Moment Grounding",
+      "value": "Retrieve the matching time window for an annotation-derived text query.",
       "raw_hits": []
     },
     {
-      "name": "caption_grounding: public_field_output_short_is_human_readable",
+      "name": "caption_grounding: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "ranked matching moments",
+      "value": "text-like query and candidate windows",
       "raw_hits": []
     },
     {
-      "name": "caption_grounding: public_field_plain_goal_is_human_readable",
+      "name": "caption_grounding: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Given a text-like query from annotation, find the matching time window.",
+      "value": "ranked matching moments",
       "raw_hits": []
     },
     {
-      "name": "caption_grounding: public_field_input_short_is_human_readable",
+      "name": "caption_grounding: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "text-like query and candidate windows",
+      "value": "query features -> candidate index -> cosine ranker",
       "raw_hits": []
     },
     {
@@ -922,15 +922,15 @@
       "raw_hits": []
     },
     {
-      "name": "caption_grounding: public_field_card_blurb_is_human_readable",
+      "name": "caption_grounding: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Retrieve the matching time window for an annotation-derived text query.",
+      "value": "Language-to-Moment Grounding",
       "raw_hits": []
     },
     {
-      "name": "caption_grounding: public_field_process_short_is_human_readable",
+      "name": "caption_grounding: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "query features -> candidate index -> cosine ranker",
+      "value": "Given a text-like query from annotation, find the matching time window.",
       "raw_hits": []
     },
     {
@@ -1008,27 +1008,27 @@
       "observed": "cross_modal_retrieval"
     },
     {
-      "name": "cross_modal_retrieval: public_field_research_name_is_human_readable",
+      "name": "cross_modal_retrieval: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "Multimodal Representation Retrieval",
+      "value": "Use motion, IMU, and camera-pose signals to retrieve the matching depth/video window.",
       "raw_hits": []
     },
     {
-      "name": "cross_modal_retrieval: public_field_output_short_is_human_readable",
+      "name": "cross_modal_retrieval: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "ranked visual windows",
+      "value": "motion/IMU/pose query; depth/video candidates",
       "raw_hits": []
     },
     {
-      "name": "cross_modal_retrieval: public_field_plain_goal_is_human_readable",
+      "name": "cross_modal_retrieval: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Use one group of modalities to retrieve the matching window from another group.",
+      "value": "ranked visual windows",
       "raw_hits": []
     },
     {
-      "name": "cross_modal_retrieval: public_field_input_short_is_human_readable",
+      "name": "cross_modal_retrieval: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "motion/IMU/pose query; depth/video candidates",
+      "value": "modality split -> projection -> nearest-neighbor ranker",
       "raw_hits": []
     },
     {
@@ -1038,15 +1038,15 @@
       "raw_hits": []
     },
     {
-      "name": "cross_modal_retrieval: public_field_card_blurb_is_human_readable",
+      "name": "cross_modal_retrieval: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Use motion, IMU, and camera-pose signals to retrieve the matching depth/video window.",
+      "value": "Multimodal Representation Retrieval",
       "raw_hits": []
     },
     {
-      "name": "cross_modal_retrieval: public_field_process_short_is_human_readable",
+      "name": "cross_modal_retrieval: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "modality split -> projection -> nearest-neighbor ranker",
+      "value": "Use one group of modalities to retrieve the matching window from another group.",
       "raw_hits": []
     },
     {
@@ -1126,27 +1126,27 @@
       "observed": "modality_reconstruction"
     },
     {
-      "name": "modality_reconstruction: public_field_research_name_is_human_readable",
+      "name": "modality_reconstruction: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "Modality Feature Reconstruction",
+      "value": "Predict compressed depth/video feature vectors from motion, IMU, and camera-pose features.",
       "raw_hits": []
     },
     {
-      "name": "modality_reconstruction: public_field_output_short_is_human_readable",
+      "name": "modality_reconstruction: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "reconstructed depth/video vector",
+      "value": "motion, IMU, and camera/pose features",
       "raw_hits": []
     },
     {
-      "name": "modality_reconstruction: public_field_plain_goal_is_human_readable",
+      "name": "modality_reconstruction: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Predict one modality feature block from other modality blocks.",
+      "value": "reconstructed depth/video vector",
       "raw_hits": []
     },
     {
-      "name": "modality_reconstruction: public_field_input_short_is_human_readable",
+      "name": "modality_reconstruction: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "motion, IMU, and camera/pose features",
+      "value": "source-target split -> scaler -> regression head",
       "raw_hits": []
     },
     {
@@ -1156,15 +1156,15 @@
       "raw_hits": []
     },
     {
-      "name": "modality_reconstruction: public_field_card_blurb_is_human_readable",
+      "name": "modality_reconstruction: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Predict compressed depth/video feature vectors from motion, IMU, and camera-pose features.",
+      "value": "Modality Feature Reconstruction",
       "raw_hits": []
     },
     {
-      "name": "modality_reconstruction: public_field_process_short_is_human_readable",
+      "name": "modality_reconstruction: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "source-target split -> scaler -> regression head",
+      "value": "Predict one modality feature block from other modality blocks.",
       "raw_hits": []
     },
     {
@@ -1244,27 +1244,27 @@
       "observed": "temporal_order"
     },
     {
-      "name": "temporal_order: public_field_research_name_is_human_readable",
+      "name": "temporal_order: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "Temporal Order Verification",
+      "value": "Tell whether two neighboring windows are in chronological order or reversed.",
       "raw_hits": []
     },
     {
-      "name": "temporal_order: public_field_output_short_is_human_readable",
+      "name": "temporal_order: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "correct or reversed",
+      "value": "two adjacent windows plus difference vector",
       "raw_hits": []
     },
     {
-      "name": "temporal_order: public_field_plain_goal_is_human_readable",
+      "name": "temporal_order: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Tell whether two nearby windows are in the correct time order.",
+      "value": "correct or reversed",
       "raw_hits": []
     },
     {
-      "name": "temporal_order: public_field_input_short_is_human_readable",
+      "name": "temporal_order: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "two adjacent windows plus difference vector",
+      "value": "pair builder -> feature combiner -> binary classifier",
       "raw_hits": []
     },
     {
@@ -1274,15 +1274,15 @@
       "raw_hits": []
     },
     {
-      "name": "temporal_order: public_field_card_blurb_is_human_readable",
+      "name": "temporal_order: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Tell whether two neighboring windows are in chronological order or reversed.",
+      "value": "Temporal Order Verification",
       "raw_hits": []
     },
     {
-      "name": "temporal_order: public_field_process_short_is_human_readable",
+      "name": "temporal_order: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "pair builder -> feature combiner -> binary classifier",
+      "value": "Tell whether two nearby windows are in the correct time order.",
       "raw_hits": []
     },
     {
@@ -1360,27 +1360,27 @@
       "observed": "misalignment_detection"
     },
     {
-      "name": "misalignment_detection: public_field_research_name_is_human_readable",
+      "name": "misalignment_detection: public_field_card_blurb_is_human_readable",
       "status": "pass",
-      "value": "Cross-Modal Misalignment Detection",
+      "value": "Detect whether motion and visual/depth streams have been artificially shifted out of sync.",
       "raw_hits": []
     },
     {
-      "name": "misalignment_detection: public_field_output_short_is_human_readable",
+      "name": "misalignment_detection: public_field_input_short_is_human_readable",
       "status": "pass",
-      "value": "aligned or shifted",
+      "value": "motion-side and visual/depth-side feature groups",
       "raw_hits": []
     },
     {
-      "name": "misalignment_detection: public_field_plain_goal_is_human_readable",
+      "name": "misalignment_detection: public_field_output_short_is_human_readable",
       "status": "pass",
-      "value": "Detect when modalities that should match are shifted out of sync.",
+      "value": "aligned or shifted",
       "raw_hits": []
     },
     {
-      "name": "misalignment_detection: public_field_input_short_is_human_readable",
+      "name": "misalignment_detection: public_field_process_short_is_human_readable",
       "status": "pass",
-      "value": "motion-side and visual/depth-side feature groups",
+      "value": "aligned/shifted pairs -> feature combiner -> binary classifier",
       "raw_hits": []
     },
     {
@@ -1390,15 +1390,15 @@
       "raw_hits": []
     },
     {
-      "name": "misalignment_detection: public_field_card_blurb_is_human_readable",
+      "name": "misalignment_detection: public_field_research_name_is_human_readable",
       "status": "pass",
-      "value": "Detect whether motion and visual/depth streams have been artificially shifted out of sync.",
+      "value": "Cross-Modal Misalignment Detection",
       "raw_hits": []
     },
     {
-      "name": "misalignment_detection: public_field_process_short_is_human_readable",
+      "name": "misalignment_detection: public_field_plain_goal_is_human_readable",
       "status": "pass",
-      "value": "aligned/shifted pairs -> feature combiner -> binary classifier",
+      "value": "Detect when modalities that should match are shifted out of sync.",
       "raw_hits": []
     },
     {

docs/data/website_integrity.json

@@ -1,14 +1,14 @@
 {
   "status": "pass",
-  "generated_at_utc": "2026-06-03T13:07:43+00:00",
+  "generated_at_utc": "2026-06-03T14:22:46+00:00",
   "docs_root": "docs",
   "site_base": "/ropedia-xperience-10m-task-suite/",
   "summary": {
     "html_pages": 4,
-    "local_references": 131,
-    "external_reference_count": 88,
-    "json_files": 30,
-    "image_assets_referenced": 20,
+    "local_references": 136,
+    "external_reference_count": 90,
+    "json_files": 31,
+    "image_assets_referenced": 21,
     "failure_count": 0
   },
   "failures": {
@@ -75,7 +75,7 @@
       "status": "pass",
       "reason": "The project overview should appear before the deeper progress ledger.",
       "overview_index": 66066,
-      "evidence_index": 80505
+      "evidence_index": 80778
     },
     {
       "name": "project_status_links_json",
@@ -150,7 +150,7 @@
       "reason": "The evaluation protocol should appear before the deeper evidence ledger.",
       "overview_index": 66066,
       "protocol_index": 77908,
-      "evidence_index": 80505
+      "evidence_index": 80778
     },
     {
       "name": "evaluation_protocol_links_json",
@@ -224,8 +224,8 @@
     {
       "path": "index.html",
       "id_count": 75,
-      "reference_count": 109,
-      "image_count": 22
+      "reference_count": 114,
+      "image_count": 23
     },
     {
       "path": "research_roadmap.html",
@@ -243,7 +243,12 @@
   "json_files": [
     {
       "path": "data/artifact_index.json",
-      "bytes": 29016,
+      "bytes": 31423,
+      "top_level_type": "dict"
+    },
+    {
+      "path": "data/audio_ablation_summary.json",
+      "bytes": 9735,
       "top_level_type": "dict"
     },
     {
@@ -268,12 +273,12 @@
     },
     {
       "path": "data/live_publication_status.json",
-      "bytes": 68740,
+      "bytes": 68744,
       "top_level_type": "dict"
     },
     {
       "path": "data/mirror_parity.json",
-      "bytes": 103779,
+      "bytes": 108702,
       "top_level_type": "dict"
     },
     {
@@ -298,17 +303,17 @@
     },
     {
       "path": "data/project_status.json",
-      "bytes": 7340,
+      "bytes": 8162,
       "top_level_type": "dict"
     },
     {
       "path": "data/public_surface_qa.json",
-      "bytes": 5647,
+      "bytes": 5648,
       "top_level_type": "dict"
     },
     {
       "path": "data/publication_audit.json",
-      "bytes": 8317,
+      "bytes": 7289,
       "top_level_type": "dict"
     },
     {
@@ -348,7 +353,7 @@
     },
     {
       "path": "data/research_takeaways.json",
-      "bytes": 5356,
+      "bytes": 6814,
       "top_level_type": "dict"
     },
     {
@@ -363,7 +368,7 @@
     },
     {
       "path": "data/source_alignment_audit.json",
-      "bytes": 4432,
+      "bytes": 4874,
       "top_level_type": "dict"
     },
     {
@@ -383,7 +388,7 @@
     },
     {
       "path": "data/website_integrity.json",
-      "bytes": 14311,
+      "bytes": 14587,
       "top_level_type": "dict"
     },
     {
@@ -409,6 +414,13 @@
       "height": 192,
       "format": "PNG"
     },
+    {
+      "path": "assets/charts/audio_ablation_delta.svg",
+      "exists": true,
+      "bytes": 4146,
+      "format": "SVG",
+      "has_viewbox": true
+    },
     {
       "path": "assets/charts/cross_modal_retrieval.svg",
       "exists": true,

docs/index.html

@@ -2372,6 +2372,7 @@
           <article class="artifact"><h3>Split policy</h3><p>Single-episode chronological 70/30 train/test split. This avoids random future-window mixing; cross-episode generalization is measured in the later multi-episode pilot.</p><a href="https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite/blob/main/EVALUATION_PROTOCOL.md">protocol doc</a></article>
           <article class="artifact"><h3>Metric contract</h3><p>All 12 tasks list input, target, primary metric, minimal baseline score, and neural MLP score from committed result files.</p><a href="data/summary_metrics.json">summary metrics</a></article>
           <article class="artifact"><h3>Leakage controls</h3><p>Scalers fit on train windows only; future labels, target feature blocks, caption/object labels, and contact labels stay on the target side unless explicitly queried.</p><a href="https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite/blob/main/scripts/build_evaluation_protocol.py">builder script</a></article>
+          <article class="artifact"><h3>Audio ablation</h3><p>The current AAC block and a 588-d raw log-mel replacement are evaluated across all 12 task contracts under the same chronological split.</p><a href="data/audio_ablation_summary.json">audio summary</a></article>
           <article class="artifact"><h3>Next evaluation stage</h3><p>This public-sample run covers single-episode task development. Cross-episode generalization, audio-visual learning, pixel-depth reconstruction, neural rendering, and full 32-episode Qwen3-Omni training move to the multi-episode stage.</p><a href="data/scope_claims_audit.json">pilot status</a></article>
           <article class="artifact"><h3>Scale-up requirement</h3><p>The Omni pilot requires at least 32 valid episodes, held-out episode splits, no train/test episode leakage, training metadata, predictions, metrics, and a run report.</p><a href="https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite/blob/main/results/omni_finetune/DATA_ACCESS_STATUS.md">data status</a></article>
         </div>
@@ -2403,6 +2404,16 @@
               <a href="https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite/tree/main/results/episode_task_suite/neural_mlp">neural_mlp/</a>
             </div>
           </article>
+          <article class="evidence-card">
+            <span class="status-pill">verified</span>
+            <h3>Audio contribution is measured task by task</h3>
+            <p>Current AAC audio improves the primary metric on 6 of 12 task contracts; raw log-mel replacement improves over current audio on 6 of 12 tasks.</p>
+            <div class="evidence-links">
+              <a href="data/audio_ablation_summary.json">audio summary</a>
+              <a href="assets/charts/audio_ablation_delta.svg">delta chart</a>
+              <a href="https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite/blob/main/results/audio_ablation/AUDIO_ABLATION_SUMMARY.md">report</a>
+            </div>
+          </article>
           <article class="evidence-card">
             <span class="status-pill">verified</span>
             <h3>Four research directions are mapped by evidence type</h3>
@@ -2844,7 +2855,7 @@
       <div class="wrap">
         <div class="section-head">
           <h2>The 12 tasks share four head families.</h2>
-          <p>The diagram separates the shared episode-window feature pipeline from the task-specific heads, and notes that audio remains dataset context rather than a current baseline feature block.</p>
+          <p>The diagram separates the shared episode-window feature pipeline from the task-specific heads. AAC audio is part of the current baseline feature block, and raw log-mel audio is now measured in the ablation upgrade.</p>
         </div>
         <img class="architecture-image" src="assets/task_architectures.png?v=xperience10m-nn" alt="Verified minimal and neural architecture diagram for all 12 Ropedia Xperience-10M tasks">
       </div>
@@ -2942,8 +2953,9 @@
           <img class="chart" src="assets/charts/cross_modal_retrieval.svg" alt="Cross modal retrieval chart">
           <img class="chart" src="assets/charts/episode_task_scores_neural_mlp.svg" alt="Neural MLP task score chart">
           <img class="chart" src="assets/charts/episode_task_scores_minimal_vs_neural.svg" alt="Minimal versus neural score chart">
+          <img class="chart" src="assets/charts/audio_ablation_delta.svg" alt="Measured audio delta chart across 12 task contracts">
         </div>
-        <p class="section-note"><a href="single_episode_explorer.html">Open the single-episode explorer</a> to inspect window-level labels, predictions, feature-block statistics, object labels, and diagnostic scores.</p>
+        <p class="section-note"><a href="single_episode_explorer.html">Open the single-episode explorer</a> to inspect window-level labels, predictions, feature-block statistics, object labels, and diagnostic scores. The audio ablation report is available at <a href="data/audio_ablation_summary.json">audio_ablation_summary.json</a>.</p>
       </div>
     </section>
 
@@ -2985,6 +2997,7 @@
               <article class="artifact"><h3>Four-direction taxonomy</h3><p>Generated JSON, CSV, Markdown, and website data mapping all 12 tasks to the four research tracks.</p><a href="https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite/tree/main/results/episode_task_suite/research_directions">research_directions/</a></article>
               <article class="artifact"><h3>Direction extension probes</h3><p>Four coded probes, one per research direction, with minimal and neural metrics plus prediction/rank CSVs.</p><a href="https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite/tree/main/results/episode_task_suite/research_direction_extensions">research_direction_extensions/</a></article>
               <article class="artifact"><h3>Task walkthroughs</h3><p>Case studies for all 12 tasks, including input, middle process modules, output, metric, limitation, and task-player data.</p><a href="https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite/tree/main/results/episode_task_suite/task_walkthroughs">task_walkthroughs/</a></article>
+              <article class="artifact"><h3>Audio ablation and raw upgrade</h3><p>All 72 task/variant rows comparing current audio, no audio, handcrafted-audio only, raw-audio only, raw replacement, and all-plus-raw.</p><a href="https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite/tree/main/results/audio_ablation">audio_ablation/</a></article>
               <article class="artifact"><h3>Single-episode explorer</h3><p>Interactive window-level view of labels, predictions, feature-block statistics, object labels, and diagnostics.</p><a href="single_episode_explorer.html">single_episode_explorer.html</a></article>
               <article class="artifact"><h3>Cross-modal retrieval</h3><p>The strongest self-supervised signal from the single episode.</p><a href="https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite/blob/main/results/episode_task_suite/cross_modal_retrieval/metrics.json">metrics.json</a></article>
             </div>

results/audio_ablation/AUDIO_ABLATION_SUMMARY.md

@@ -0,0 +1,38 @@
+# Audio Ablation and Raw-Audio Upgrade
+
+This report is generated from committed task-suite artifacts plus the local public-sample MP4 audio stream.
+It measures whether audio changes each single-episode task under the same chronological split.
+
+## Raw Audio Feature
+
+- Source: `local_public_sample/fisheye_cam0.mp4`
+- Has audio: `True`
+- Sample rate: `16000`
+- Window feature dim: `588`
+- Feature: Per-window raw waveform STFT log-mel statistics plus delta and waveform envelope statistics.
+
+## Task Deltas
+
+| Task | Metric | Current audio | No audio | Current audio delta | Raw replaces audio | Raw replacement delta |
+| --- | --- | ---: | ---: | ---: | ---: | ---: |
+| Current Action Recognition | macro_f1 | 0.0091 | 0.0088 | 0.0003 | 0.0013 | -0.0077 |
+| Current Subtask Recognition | macro_f1 | 0.0113 | 0.0112 | 0.0001 | 0.0008 | -0.0104 |
+| Action Transition Detection | macro_f1 | 0.4621 | 0.4687 | -0.0066 | 0.4792 | 0.0171 |
+| Next-Action Prediction | macro_f1 | 0.0106 | 0.0107 | -0.0001 | 0.0060 | -0.0046 |
+| Future Hand Motion Forecasting | mae | 4.4664 | 4.3038 | -0.1626 | 4.3059 | 0.1605 |
+| Contact State Prediction | macro_f1 | 1.0000 | 1.0000 | 0.0000 | 1.0000 | 0.0000 |
+| Relevant Object Prediction | micro_f1 | 0.1581 | 0.1479 | 0.0102 | 0.1787 | 0.0206 |
+| Language-to-Time Grounding | mrr | 0.0321 | 0.0272 | 0.0049 | 0.0248 | -0.0072 |
+| Cross-Modal Window Retrieval | mrr | 0.3751 | 0.3892 | -0.0141 | 0.3275 | -0.0476 |
+| Sensor-to-Visual Reconstruction | mae | 9.7942 | 10.4467 | 0.6524 | 8.8307 | 0.9635 |
+| Temporal Order Verification | macro_f1 | 0.5172 | 0.4943 | 0.0230 | 0.5302 | 0.0129 |
+| Cross-Modal Misalignment Detection | macro_f1 | 0.4173 | 0.4226 | -0.0052 | 0.4438 | 0.0264 |
+
+## Aggregate
+
+- Mean current-audio delta: `0.041849794979543296`
+- Tasks where current handcrafted audio improves the primary metric: `6`
+- Mean raw-replacement delta vs current handcrafted audio: `0.09362598132150173`
+- Tasks where raw log-mel replacement improves over current handcrafted audio: `6`
+
+Positive deltas always mean better according to each task's primary metric. For MAE tasks, lower MAE is converted into a positive improvement.

results/audio_ablation/audio_ablation_metrics.csv

@@ -0,0 +1,73 @@
+task,task_display,variant,variant_display,status,primary_metric,primary_value,higher_is_better,feature_dim,num_train,num_test,input_contract,target_variant,reason,accuracy,macro_f1,balanced_accuracy,num_classes,unseen_test_classes,unseen_test_class_count,mse,mae,r2,target_dim,micro_f1,exact_match,precision,recall,num_objects,mrr,top1_accuracy,top5_accuracy,top10_accuracy,median_rank,mean_rank,num_queries
+timeline_action,Current Action Recognition,all_handcrafted_audio,All Current Features,computed,macro_f1,0.00905456968081885,true,8546,813,348,task contract feature blocks with handcrafted AAC audio where applicable (8546 dims),,,0.017241379310344827,0.00905456968081885,0.01720647773279352,19,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+timeline_action,Current Action Recognition,all_except_audio,All Except Audio,computed,macro_f1,0.008771929824561405,true,8378,813,348,same task contract with handcrafted AAC audio columns removed (8378 dims),,,0.020114942528735632,0.008771929824561405,0.005668016194331984,19,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+timeline_action,Current Action Recognition,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,macro_f1,0.006925207756232688,true,168,813,348,handcrafted AAC audio block only (168 dims),,,0.014367816091954023,0.006925207756232688,0.004048582995951417,19,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+timeline_action,Current Action Recognition,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,macro_f1,0.0,true,588,813,348,raw waveform log-mel embedding only (588 dims),,,0.0,0.0,0.0,19,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+timeline_action,Current Action Recognition,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,macro_f1,0.0013495276653171392,true,8966,813,348,task contract with handcrafted AAC removed and raw log-mel added (8966 dims),,,0.0028735632183908046,0.0013495276653171392,0.0008097165991902835,19,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+timeline_action,Current Action Recognition,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,macro_f1,0.002734107997265892,true,9134,813,348,task contract with existing handcrafted AAC plus raw log-mel (9134 dims),,,0.005747126436781609,0.002734107997265892,0.001619433198380567,19,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+timeline_subtask,Current Subtask Recognition,all_handcrafted_audio,All Current Features,computed,macro_f1,0.011256354393609296,true,8546,813,348,task contract feature blocks with handcrafted AAC audio where applicable (8546 dims),,,0.02586206896551724,0.011256354393609296,0.02788220551378446,15,"['Move bottle to coffee equipment', 'Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+timeline_subtask,Current Subtask Recognition,all_except_audio,All Except Audio,computed,macro_f1,0.0111731843575419,true,8378,813,348,same task contract with handcrafted AAC audio columns removed (8378 dims),,,0.040229885057471264,0.0111731843575419,0.017543859649122806,15,"['Move bottle to coffee equipment', 'Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+timeline_subtask,Current Subtask Recognition,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,macro_f1,0.016194331983805668,true,168,813,348,handcrafted AAC audio block only (168 dims),,,0.022988505747126436,0.016194331983805668,0.010796221322537112,15,"['Move bottle to coffee equipment', 'Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+timeline_subtask,Current Subtask Recognition,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,macro_f1,0.0016722408026755855,true,588,813,348,raw waveform log-mel embedding only (588 dims),,,0.0028735632183908046,0.0016722408026755855,0.001349527665317139,15,"['Move bottle to coffee equipment', 'Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+timeline_subtask,Current Subtask Recognition,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,macro_f1,0.0008257638315441783,true,8966,813,348,task contract with handcrafted AAC removed and raw log-mel added (8966 dims),,,0.0028735632183908046,0.0008257638315441783,0.0012531328320802004,15,"['Move bottle to coffee equipment', 'Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+timeline_subtask,Current Subtask Recognition,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,macro_f1,0.0017889087656529517,true,9134,813,348,task contract with existing handcrafted AAC plus raw log-mel (9134 dims),,,0.005747126436781609,0.0017889087656529517,0.002699055330634278,15,"['Move bottle to coffee equipment', 'Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+transition_detection,Action Transition Detection,all_handcrafted_audio,All Current Features,computed,macro_f1,0.46213292117465227,true,8546,813,348,task contract feature blocks with handcrafted AAC audio where applicable (8546 dims),,,0.8591954022988506,0.46213292117465227,0.4503012048192771,2,[],0,,,,,,,,,,,,,,,,
+transition_detection,Action Transition Detection,all_except_audio,All Except Audio,computed,macro_f1,0.46870229007633585,true,8378,813,348,same task contract with handcrafted AAC audio columns removed (8378 dims),,,0.882183908045977,0.46870229007633585,0.4623493975903614,2,[],0,,,,,,,,,,,,,,,,
+transition_detection,Action Transition Detection,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,macro_f1,0.48444444444444446,true,168,813,348,handcrafted AAC audio block only (168 dims),,,0.9396551724137931,0.48444444444444446,0.4924698795180723,2,[],0,,,,,,,,,,,,,,,,
+transition_detection,Action Transition Detection,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,macro_f1,0.4637904468412942,true,588,813,348,raw waveform log-mel embedding only (588 dims),,,0.8649425287356322,0.4637904468412942,0.45331325301204817,2,[],0,,,,,,,,,,,,,,,,
+transition_detection,Action Transition Detection,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,macro_f1,0.4792100707180375,true,8966,813,348,task contract with handcrafted AAC removed and raw log-mel added (8966 dims),,,0.853448275862069,0.4792100707180375,0.4770331325301205,2,[],0,,,,,,,,,,,,,,,,
+transition_detection,Action Transition Detection,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,macro_f1,0.4816233470132239,true,9134,813,348,task contract with existing handcrafted AAC plus raw log-mel (9134 dims),,,0.8591954022988506,0.4816233470132239,0.48004518072289154,2,[],0,,,,,,,,,,,,,,,,
+next_action,Next-Action Prediction,all_handcrafted_audio,All Current Features,computed,macro_f1,0.01058201058201058,true,8546,813,348,task contract feature blocks with handcrafted AAC audio where applicable (8546 dims),future action from annotation frame labels,,0.022988505747126436,0.01058201058201058,0.007407407407407408,18,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+next_action,Next-Action Prediction,all_except_audio,All Except Audio,computed,macro_f1,0.010709504685408301,true,8378,813,348,same task contract with handcrafted AAC audio columns removed (8378 dims),future action from annotation frame labels,,0.022988505747126436,0.010709504685408301,0.007407407407407408,18,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+next_action,Next-Action Prediction,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,macro_f1,0.008561643835616438,true,168,813,348,handcrafted AAC audio block only (168 dims),future action from annotation frame labels,,0.014367816091954023,0.008561643835616438,0.005208333333333333,18,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+next_action,Next-Action Prediction,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,macro_f1,0.0017301038062283738,true,588,813,348,raw waveform log-mel embedding only (588 dims),future action from annotation frame labels,,0.0028735632183908046,0.0017301038062283738,0.000980392156862745,18,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+next_action,Next-Action Prediction,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,macro_f1,0.006006006006006006,true,8966,813,348,task contract with handcrafted AAC removed and raw log-mel added (8966 dims),future action from annotation frame labels,,0.011494252873563218,0.006006006006006006,0.003703703703703704,18,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+next_action,Next-Action Prediction,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,macro_f1,0.0058479532163742695,true,9134,813,348,task contract with existing handcrafted AAC plus raw log-mel (9134 dims),future action from annotation frame labels,,0.011494252873563218,0.0058479532163742695,0.003703703703703704,18,"['Place item on table', 'Wait/Prepare for pouring', 'Pour coffee', 'Pour milk into coffee']",4,,,,,,,,,,,,,,,,
+hand_trajectory_forecast,Future Hand Motion Forecasting,all_handcrafted_audio,All Current Features,computed,mae,4.466395378112793,false,8546,811,348,task contract feature blocks with handcrafted AAC audio where applicable (8546 dims),future hand joints from annotation.hdf5,,,,,,,,1557.2276611328125,4.466395378112793,-1588.9305967247317,1260,,,,,,,,,,,,
+hand_trajectory_forecast,Future Hand Motion Forecasting,all_except_audio,All Except Audio,computed,mae,4.303755283355713,false,8378,811,348,same task contract with handcrafted AAC audio columns removed (8378 dims),future hand joints from annotation.hdf5,,,,,,,,1229.3953857421875,4.303755283355713,-1254.2135816756352,1260,,,,,,,,,,,,
+hand_trajectory_forecast,Future Hand Motion Forecasting,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,mae,1.1956232786178589,false,168,811,348,handcrafted AAC audio block only (168 dims),future hand joints from annotation.hdf5,,,,,,,,3.1508853435516357,1.1956232786178589,-2.2170563799739673,1260,,,,,,,,,,,,
+hand_trajectory_forecast,Future Hand Motion Forecasting,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,mae,3.1172122955322266,false,588,811,348,raw waveform log-mel embedding only (588 dims),future hand joints from annotation.hdf5,,,,,,,,63.60802459716797,3.1172122955322266,-63.94383890504609,1260,,,,,,,,,,,,
+hand_trajectory_forecast,Future Hand Motion Forecasting,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,mae,4.305870532989502,false,8966,811,348,task contract with handcrafted AAC removed and raw log-mel added (8966 dims),future hand joints from annotation.hdf5,,,,,,,,1241.380126953125,4.305870532989502,-1266.4500920921807,1260,,,,,,,,,,,,
+hand_trajectory_forecast,Future Hand Motion Forecasting,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,mae,4.1367621421813965,false,9134,811,348,task contract with existing handcrafted AAC plus raw log-mel (9134 dims),future hand joints from annotation.hdf5,,,,,,,,1217.459716796875,4.1367621421813965,-1242.0272923501784,1260,,,,,,,,,,,,
+contact_prediction,Contact State Prediction,all_handcrafted_audio,All Current Features,computed,macro_f1,1.0,true,7503,813,348,task contract feature blocks with handcrafted AAC audio where applicable (7503 dims),,,1.0,1.0,1.0,1,[],0,,,,,,,,,,,,,,,,
+contact_prediction,Contact State Prediction,all_except_audio,All Except Audio,computed,macro_f1,1.0,true,7335,813,348,same task contract with handcrafted AAC audio columns removed (7335 dims),,,1.0,1.0,1.0,1,[],0,,,,,,,,,,,,,,,,
+contact_prediction,Contact State Prediction,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,macro_f1,1.0,true,168,813,348,handcrafted AAC audio block only (168 dims),,,1.0,1.0,1.0,1,[],0,,,,,,,,,,,,,,,,
+contact_prediction,Contact State Prediction,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,macro_f1,1.0,true,588,813,348,raw waveform log-mel embedding only (588 dims),,,1.0,1.0,1.0,1,[],0,,,,,,,,,,,,,,,,
+contact_prediction,Contact State Prediction,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,macro_f1,1.0,true,7923,813,348,task contract with handcrafted AAC removed and raw log-mel added (7923 dims),,,1.0,1.0,1.0,1,[],0,,,,,,,,,,,,,,,,
+contact_prediction,Contact State Prediction,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,macro_f1,1.0,true,8091,813,348,task contract with existing handcrafted AAC plus raw log-mel (8091 dims),,,1.0,1.0,1.0,1,[],0,,,,,,,,,,,,,,,,
+object_relevance,Relevant Object Prediction,all_handcrafted_audio,All Current Features,computed,micro_f1,0.15813953488372093,true,7650,813,348,task contract feature blocks with handcrafted AAC audio where applicable (7650 dims),,,,0.05335536055344564,,,,,,,,,0.15813953488372093,0.011494252873563218,0.15368567454798332,0.16285924834193072,34,,,,,,,
+object_relevance,Relevant Object Prediction,all_except_audio,All Except Audio,computed,micro_f1,0.14793328498912256,true,7482,813,348,same task contract with handcrafted AAC audio columns removed (7482 dims),,,,0.05137956064750565,,,,,,,,,0.14793328498912256,0.008620689655172414,0.145610278372591,0.1503316138540899,34,,,,,,,
+object_relevance,Relevant Object Prediction,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,micro_f1,0.15894039735099336,true,168,813,348,handcrafted AAC audio block only (168 dims),,,,0.0640376063191102,,,,,,,,,0.15894039735099336,0.005747126436781609,0.16859504132231404,0.1503316138540899,34,,,,,,,
+object_relevance,Relevant Object Prediction,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,micro_f1,0.15894868585732164,true,588,813,348,raw waveform log-mel embedding only (588 dims),,,,0.06183171604594236,,,,,,,,,0.15894868585732164,0.0,0.13811854268624252,0.18717759764185704,34,,,,,,,
+object_relevance,Relevant Object Prediction,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,micro_f1,0.17871759890859482,true,8070,813,348,task contract with handcrafted AAC removed and raw log-mel added (8070 dims),,,,0.05575181300519915,,,,,,,,,0.17871759890859482,0.0028735632183908046,0.16634920634920636,0.19307295504789979,34,,,,,,,
+object_relevance,Relevant Object Prediction,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,micro_f1,0.18262653898768813,true,8238,813,348,task contract with existing handcrafted AAC plus raw log-mel (8238 dims),,,,0.05735108836010565,,,,,,,,,0.18262653898768813,0.0028735632183908046,0.17038927887683472,0.1967575534266765,34,,,,,,,
+caption_grounding,Language-to-Time Grounding,all_handcrafted_audio,All Current Features,computed,mrr,0.03208567947149277,true,7650,813,348,task contract feature blocks with handcrafted AAC audio where applicable (7650 dims),,,,,,,,,,,,896,,,,,,0.03208567947149277,0.0028735632183908046,0.040229885057471264,0.06896551724137931,132.0,137.4022979736328,348
+caption_grounding,Language-to-Time Grounding,all_except_audio,All Except Audio,computed,mrr,0.027228528633713722,true,7482,813,348,same task contract with handcrafted AAC audio columns removed (7482 dims),,,,,,,,,,,,896,,,,,,0.027228528633713722,0.005747126436781609,0.028735632183908046,0.04597701149425287,134.0,142.62930297851562,348
+caption_grounding,Language-to-Time Grounding,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,mrr,0.03902389109134674,true,168,813,348,handcrafted AAC audio block only (168 dims),,,,,,,,,,,,896,,,,,,0.03902389109134674,0.011494252873563218,0.04885057471264368,0.07758620689655173,141.0,152.14942932128906,348
+caption_grounding,Language-to-Time Grounding,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,mrr,0.014815197326242924,true,588,813,348,raw waveform log-mel embedding only (588 dims),,,,,,,,,,,,896,,,,,,0.014815197326242924,0.0,0.005747126436781609,0.022988505747126436,164.5,168.51437377929688,348
+caption_grounding,Language-to-Time Grounding,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,mrr,0.02484782598912716,true,8070,813,348,task contract with handcrafted AAC removed and raw log-mel added (8070 dims),,,,,,,,,,,,896,,,,,,0.02484782598912716,0.0028735632183908046,0.022988505747126436,0.05172413793103448,120.5,137.83908081054688,348
+caption_grounding,Language-to-Time Grounding,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,mrr,0.02719014883041382,true,8238,813,348,task contract with existing handcrafted AAC plus raw log-mel (8238 dims),,,,,,,,,,,,896,,,,,,0.02719014883041382,0.005747126436781609,0.02586206896551724,0.05747126436781609,116.0,136.37930297851562,348
+cross_modal_retrieval,Cross-Modal Window Retrieval,all_handcrafted_audio,All Current Features,computed,mrr,0.3751238286495209,true,2415,813,348,task contract feature blocks with handcrafted AAC audio where applicable (2415 dims),,,,,,,,,,,,5096,,,,,,0.3751238286495209,0.26436781609195403,0.47988505747126436,0.5545977011494253,6.5,25.83333396911621,348
+cross_modal_retrieval,Cross-Modal Window Retrieval,all_except_audio,All Except Audio,computed,mrr,0.38921058177948,true,2247,813,348,same task contract with handcrafted AAC audio columns removed (2247 dims),,,,,,,,,,,,5096,,,,,,0.38921058177948,0.28448275862068967,0.4827586206896552,0.5718390804597702,6.0,25.27298927307129,348
+cross_modal_retrieval,Cross-Modal Window Retrieval,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,mrr,0.02334633097052574,true,168,813,348,handcrafted AAC audio block only (168 dims),,,,,,,,,,,,5096,,,,,,0.02334633097052574,0.005747126436781609,0.014367816091954023,0.031609195402298854,152.5,161.44540405273438,348
+cross_modal_retrieval,Cross-Modal Window Retrieval,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,mrr,0.01806792803108692,true,588,813,348,raw waveform log-mel embedding only (588 dims),,,,,,,,,,,,5096,,,,,,0.01806792803108692,0.0028735632183908046,0.008620689655172414,0.022988505747126436,162.5,165.3275909423828,348
+cross_modal_retrieval,Cross-Modal Window Retrieval,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,mrr,0.32749155163764954,true,2835,813,348,task contract with handcrafted AAC removed and raw log-mel added (2835 dims),,,,,,,,,,,,5096,,,,,,0.32749155163764954,0.22126436781609196,0.4367816091954023,0.514367816091954,9.0,32.57183837890625,348
+cross_modal_retrieval,Cross-Modal Window Retrieval,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,mrr,0.31795138120651245,true,3003,813,348,task contract with existing handcrafted AAC plus raw log-mel (3003 dims),,,,,,,,,,,,5096,,,,,,0.31795138120651245,0.20689655172413793,0.4396551724137931,0.5287356321839081,8.5,33.75,348
+modality_reconstruction,Sensor-to-Visual Reconstruction,all_handcrafted_audio,All Current Features,computed,mae,9.79421329498291,false,2415,813,348,task contract feature blocks with handcrafted AAC audio where applicable (2415 dims),,,,,,,,,13864.333984375,9.79421329498291,-0.6094599339266962,5096,,,,,,,,,,,,
+modality_reconstruction,Sensor-to-Visual Reconstruction,all_except_audio,All Except Audio,computed,mae,10.446661949157715,false,2247,813,348,same task contract with handcrafted AAC audio columns removed (2247 dims),,,,,,,,,14634.3974609375,10.446661949157715,-0.6988538186806226,5096,,,,,,,,,,,,
+modality_reconstruction,Sensor-to-Visual Reconstruction,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,mae,1.359641671180725,false,168,813,348,handcrafted AAC audio block only (168 dims),,,,,,,,,8681.8916015625,1.359641671180725,-0.007849137404613904,5096,,,,,,,,,,,,
+modality_reconstruction,Sensor-to-Visual Reconstruction,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,mae,2.6225292682647705,false,588,813,348,raw waveform log-mel embedding only (588 dims),,,,,,,,,8708.6181640625,2.6225292682647705,-0.010951711094258076,5096,,,,,,,,,,,,
+modality_reconstruction,Sensor-to-Visual Reconstruction,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,mae,8.830678939819336,false,2835,813,348,task contract with handcrafted AAC removed and raw log-mel added (2835 dims),,,,,,,,,12454.744140625,8.830678939819336,-0.44582574939092745,5096,,,,,,,,,,,,
+modality_reconstruction,Sensor-to-Visual Reconstruction,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,mae,8.392388343811035,false,3003,813,348,task contract with existing handcrafted AAC plus raw log-mel (3003 dims),,,,,,,,,12078.8935546875,8.392388343811035,-0.4021946589830052,5096,,,,,,,,,,,,
+temporal_order,Temporal Order Verification,all_handcrafted_audio,All Current Features,computed,macro_f1,0.5172413793103449,true,25638,1624,696,task contract feature blocks with handcrafted AAC audio where applicable (8546 dims),,,0.5172413793103449,0.5172413793103449,0.5172413793103449,2,[],0,,,,,,,,,,,,,,,,
+temporal_order,Temporal Order Verification,all_except_audio,All Except Audio,computed,macro_f1,0.4942528735632184,true,25134,1624,696,same task contract with handcrafted AAC audio columns removed (8378 dims),,,0.4942528735632184,0.4942528735632184,0.4942528735632184,2,[],0,,,,,,,,,,,,,,,,
+temporal_order,Temporal Order Verification,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,macro_f1,0.4425287356321839,true,504,1624,696,handcrafted AAC audio block only (168 dims),,,0.4425287356321839,0.4425287356321839,0.4425287356321839,2,[],0,,,,,,,,,,,,,,,,
+temporal_order,Temporal Order Verification,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,macro_f1,0.5028735632183908,true,1764,1624,696,raw waveform log-mel embedding only (588 dims),,,0.5028735632183908,0.5028735632183908,0.5028735632183908,2,[],0,,,,,,,,,,,,,,,,
+temporal_order,Temporal Order Verification,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,macro_f1,0.5301714439065678,true,26898,1624,696,task contract with handcrafted AAC removed and raw log-mel added (8966 dims),,,0.5301724137931034,0.5301714439065678,0.5301724137931034,2,[],0,,,,,,,,,,,,,,,,
+temporal_order,Temporal Order Verification,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,macro_f1,0.5330450130569861,true,27402,1624,696,task contract with existing handcrafted AAC plus raw log-mel (9134 dims),,,0.5330459770114943,0.5330450130569861,0.5330459770114943,2,[],0,,,,,,,,,,,,,,,,
+misalignment_detection,Cross-Modal Misalignment Detection,all_handcrafted_audio,All Current Features,computed,macro_f1,0.41734045375379186,true,7511,1614,692,motion/current visual+handcrafted audio pair,,,0.4725433526011561,0.41734045375379186,0.4725433526011561,2,[],0,,,,,,,,,,,,,,,,
+misalignment_detection,Cross-Modal Misalignment Detection,all_except_audio,All Except Audio,computed,macro_f1,0.42258557365378524,true,7343,1614,692,motion/current visual pair with audio removed,,,0.47832369942196534,0.42258557365378524,0.4783236994219653,2,[],0,,,,,,,,,,,,,,,,
+misalignment_detection,Cross-Modal Misalignment Detection,handcrafted_audio_only,Handcrafted AAC Audio Only,computed,macro_f1,0.5102351916376306,true,504,1614,692,handcrafted AAC audio self-alignment pair,,,0.5115606936416185,0.5102351916376306,0.5115606936416185,2,[],0,,,,,,,,,,,,,,,,
+misalignment_detection,Cross-Modal Misalignment Detection,raw_logmel_audio_only,Raw Log-Mel Audio Only,computed,macro_f1,0.47823544277887897,true,1764,1614,692,raw log-mel audio self-alignment pair,,,0.47832369942196534,0.47823544277887897,0.4783236994219653,2,[],0,,,,,,,,,,,,,,,,
+misalignment_detection,Cross-Modal Misalignment Detection,replace_handcrafted_with_raw,Replace AAC Block With Raw Log-Mel,computed,macro_f1,0.44378951880827355,true,7931,1614,692,motion/current visual pair with raw log-mel replacing handcrafted audio,,,0.4797687861271676,0.44378951880827355,0.47976878612716767,2,[],0,,,,,,,,,,,,,,,,
+misalignment_detection,Cross-Modal Misalignment Detection,all_plus_raw_logmel,All Current Features + Raw Log-Mel,computed,macro_f1,0.4373795761078998,true,8099,1614,692,motion/current visual+handcrafted audio pair plus raw log-mel,,,0.4725433526011561,0.4373795761078998,0.4725433526011561,2,[],0,,,,,,,,,,,,,,,,

results/audio_ablation/audio_ablation_summary.json

@@ -0,0 +1,223 @@
+{
+  "description": "Measured audio ablation and raw log-mel audio upgrade over the single public Xperience-10M sample episode.",
+  "scope": "single public sample episode; chronological split; ridge heads over fixed feature contracts",
+  "raw_audio_metadata": {
+    "source": "local_public_sample/fisheye_cam0.mp4",
+    "exists": true,
+    "has_audio": true,
+    "sample_rate": 16000,
+    "fps": 20.00137419266181,
+    "num_samples": 4656994,
+    "num_windows": 1161,
+    "feature_dim": 588,
+    "mel_bands": 64,
+    "fft_size": 512,
+    "hop_length": 160,
+    "feature_description": "Per-window raw waveform STFT log-mel statistics plus delta and waveform envelope statistics."
+  },
+  "num_tasks": 12,
+  "variants": {
+    "all_handcrafted_audio": "All Current Features",
+    "all_except_audio": "All Except Audio",
+    "handcrafted_audio_only": "Handcrafted AAC Audio Only",
+    "raw_logmel_audio_only": "Raw Log-Mel Audio Only",
+    "replace_handcrafted_with_raw": "Replace AAC Block With Raw Log-Mel",
+    "all_plus_raw_logmel": "All Current Features + Raw Log-Mel"
+  },
+  "task_summaries": [
+    {
+      "task": "timeline_action",
+      "task_display": "Current Action Recognition",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.00905456968081885,
+      "all_except_audio": 0.008771929824561405,
+      "handcrafted_audio_delta": 0.0002826398562574446,
+      "raw_logmel_audio_only": 0.0,
+      "replace_handcrafted_with_raw": 0.0013495276653171392,
+      "raw_replacement_delta_vs_no_audio": -0.007422402159244265,
+      "raw_replacement_delta_vs_handcrafted": -0.00770504201550171,
+      "all_plus_raw_logmel": 0.002734107997265892,
+      "all_plus_raw_delta_vs_handcrafted": -0.006320461683552957
+    },
+    {
+      "task": "timeline_subtask",
+      "task_display": "Current Subtask Recognition",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.011256354393609296,
+      "all_except_audio": 0.0111731843575419,
+      "handcrafted_audio_delta": 8.317003606739606e-05,
+      "raw_logmel_audio_only": 0.0016722408026755855,
+      "replace_handcrafted_with_raw": 0.0008257638315441783,
+      "raw_replacement_delta_vs_no_audio": -0.01034742052599772,
+      "raw_replacement_delta_vs_handcrafted": -0.010430590562065117,
+      "all_plus_raw_logmel": 0.0017889087656529517,
+      "all_plus_raw_delta_vs_handcrafted": -0.009467445627956345
+    },
+    {
+      "task": "transition_detection",
+      "task_display": "Action Transition Detection",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.46213292117465227,
+      "all_except_audio": 0.46870229007633585,
+      "handcrafted_audio_delta": -0.006569368901683581,
+      "raw_logmel_audio_only": 0.4637904468412942,
+      "replace_handcrafted_with_raw": 0.4792100707180375,
+      "raw_replacement_delta_vs_no_audio": 0.010507780641701658,
+      "raw_replacement_delta_vs_handcrafted": 0.01707714954338524,
+      "all_plus_raw_logmel": 0.4816233470132239,
+      "all_plus_raw_delta_vs_handcrafted": 0.019490425838571634
+    },
+    {
+      "task": "next_action",
+      "task_display": "Next-Action Prediction",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.01058201058201058,
+      "all_except_audio": 0.010709504685408301,
+      "handcrafted_audio_delta": -0.0001274941033977215,
+      "raw_logmel_audio_only": 0.0017301038062283738,
+      "replace_handcrafted_with_raw": 0.006006006006006006,
+      "raw_replacement_delta_vs_no_audio": -0.004703498679402295,
+      "raw_replacement_delta_vs_handcrafted": -0.004576004576004574,
+      "all_plus_raw_logmel": 0.0058479532163742695,
+      "all_plus_raw_delta_vs_handcrafted": -0.00473405736563631
+    },
+    {
+      "task": "hand_trajectory_forecast",
+      "task_display": "Future Hand Motion Forecasting",
+      "primary_metric": "mae",
+      "higher_is_better": false,
+      "all_handcrafted_audio": 4.466395378112793,
+      "all_except_audio": 4.303755283355713,
+      "handcrafted_audio_delta": -0.16264009475708008,
+      "raw_logmel_audio_only": 3.1172122955322266,
+      "replace_handcrafted_with_raw": 4.305870532989502,
+      "raw_replacement_delta_vs_no_audio": -0.0021152496337890625,
+      "raw_replacement_delta_vs_handcrafted": 0.16052484512329102,
+      "all_plus_raw_logmel": 4.1367621421813965,
+      "all_plus_raw_delta_vs_handcrafted": 0.3296332359313965
+    },
+    {
+      "task": "contact_prediction",
+      "task_display": "Contact State Prediction",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 1.0,
+      "all_except_audio": 1.0,
+      "handcrafted_audio_delta": 0.0,
+      "raw_logmel_audio_only": 1.0,
+      "replace_handcrafted_with_raw": 1.0,
+      "raw_replacement_delta_vs_no_audio": 0.0,
+      "raw_replacement_delta_vs_handcrafted": 0.0,
+      "all_plus_raw_logmel": 1.0,
+      "all_plus_raw_delta_vs_handcrafted": 0.0
+    },
+    {
+      "task": "object_relevance",
+      "task_display": "Relevant Object Prediction",
+      "primary_metric": "micro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.15813953488372093,
+      "all_except_audio": 0.14793328498912256,
+      "handcrafted_audio_delta": 0.010206249894598368,
+      "raw_logmel_audio_only": 0.15894868585732164,
+      "replace_handcrafted_with_raw": 0.17871759890859482,
+      "raw_replacement_delta_vs_no_audio": 0.030784313919472256,
+      "raw_replacement_delta_vs_handcrafted": 0.020578064024873888,
+      "all_plus_raw_logmel": 0.18262653898768813,
+      "all_plus_raw_delta_vs_handcrafted": 0.024487004103967203
+    },
+    {
+      "task": "caption_grounding",
+      "task_display": "Language-to-Time Grounding",
+      "primary_metric": "mrr",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.03208567947149277,
+      "all_except_audio": 0.027228528633713722,
+      "handcrafted_audio_delta": 0.004857150837779045,
+      "raw_logmel_audio_only": 0.014815197326242924,
+      "replace_handcrafted_with_raw": 0.02484782598912716,
+      "raw_replacement_delta_vs_no_audio": -0.002380702644586563,
+      "raw_replacement_delta_vs_handcrafted": -0.007237853482365608,
+      "all_plus_raw_logmel": 0.02719014883041382,
+      "all_plus_raw_delta_vs_handcrafted": -0.004895530641078949
+    },
+    {
+      "task": "cross_modal_retrieval",
+      "task_display": "Cross-Modal Window Retrieval",
+      "primary_metric": "mrr",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.3751238286495209,
+      "all_except_audio": 0.38921058177948,
+      "handcrafted_audio_delta": -0.014086753129959106,
+      "raw_logmel_audio_only": 0.01806792803108692,
+      "replace_handcrafted_with_raw": 0.32749155163764954,
+      "raw_replacement_delta_vs_no_audio": -0.061719030141830444,
+      "raw_replacement_delta_vs_handcrafted": -0.04763227701187134,
+      "all_plus_raw_logmel": 0.31795138120651245,
+      "all_plus_raw_delta_vs_handcrafted": -0.05717244744300842
+    },
+    {
+      "task": "modality_reconstruction",
+      "task_display": "Sensor-to-Visual Reconstruction",
+      "primary_metric": "mae",
+      "higher_is_better": false,
+      "all_handcrafted_audio": 9.79421329498291,
+      "all_except_audio": 10.446661949157715,
+      "handcrafted_audio_delta": 0.6524486541748047,
+      "raw_logmel_audio_only": 2.6225292682647705,
+      "replace_handcrafted_with_raw": 8.830678939819336,
+      "raw_replacement_delta_vs_no_audio": 1.615983009338379,
+      "raw_replacement_delta_vs_handcrafted": 0.9635343551635742,
+      "all_plus_raw_logmel": 8.392388343811035,
+      "all_plus_raw_delta_vs_handcrafted": 1.401824951171875
+    },
+    {
+      "task": "temporal_order",
+      "task_display": "Temporal Order Verification",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.5172413793103449,
+      "all_except_audio": 0.4942528735632184,
+      "handcrafted_audio_delta": 0.022988505747126464,
+      "raw_logmel_audio_only": 0.5028735632183908,
+      "replace_handcrafted_with_raw": 0.5301714439065678,
+      "raw_replacement_delta_vs_no_audio": 0.03591857034334939,
+      "raw_replacement_delta_vs_handcrafted": 0.012930064596222923,
+      "all_plus_raw_logmel": 0.5330450130569861,
+      "all_plus_raw_delta_vs_handcrafted": 0.015803633746641288
+    },
+    {
+      "task": "misalignment_detection",
+      "task_display": "Cross-Modal Misalignment Detection",
+      "primary_metric": "macro_f1",
+      "higher_is_better": true,
+      "all_handcrafted_audio": 0.41734045375379186,
+      "all_except_audio": 0.42258557365378524,
+      "handcrafted_audio_delta": -0.005245119899993378,
+      "raw_logmel_audio_only": 0.47823544277887897,
+      "replace_handcrafted_with_raw": 0.44378951880827355,
+      "raw_replacement_delta_vs_no_audio": 0.021203945154488313,
+      "raw_replacement_delta_vs_handcrafted": 0.02644906505448169,
+      "all_plus_raw_logmel": 0.4373795761078998,
+      "all_plus_raw_delta_vs_handcrafted": 0.02003912235410793
+    }
+  ],
+  "aggregate": {
+    "mean_handcrafted_audio_delta": 0.041849794979543296,
+    "tasks_where_handcrafted_audio_improves": 6,
+    "mean_raw_replacement_delta_vs_handcrafted": 0.09362598132150173,
+    "tasks_where_raw_replacement_improves_over_handcrafted": 6
+  },
+  "provenance": {
+    "suite_dir": "results/episode_task_suite",
+    "shared_windows": "results/episode_task_suite/shared_windows.npz",
+    "feature_manifest": "results/episode_task_suite/feature_manifest.json",
+    "audio_source": "local_public_sample/fisheye_cam0.mp4",
+    "annotation_source": "local_public_sample/annotation.hdf5",
+    "homie_toolkit_available": true
+  }
+}

results/audio_ablation/audio_delta_summary.csv

@@ -0,0 +1,13 @@
+task,task_display,metric,current_audio,no_audio,current_audio_delta,raw_audio_only,replace_with_raw,raw_replacement_delta_vs_current,all_plus_raw,all_plus_raw_delta_vs_current
+timeline_action,Current Action Recognition,macro_f1,0.00905456968081885,0.008771929824561405,0.0002826398562574446,0.0,0.0013495276653171392,-0.00770504201550171,0.002734107997265892,-0.006320461683552957
+timeline_subtask,Current Subtask Recognition,macro_f1,0.011256354393609296,0.0111731843575419,8.317003606739606e-05,0.0016722408026755855,0.0008257638315441783,-0.010430590562065117,0.0017889087656529517,-0.009467445627956345
+transition_detection,Action Transition Detection,macro_f1,0.46213292117465227,0.46870229007633585,-0.006569368901683581,0.4637904468412942,0.4792100707180375,0.01707714954338524,0.4816233470132239,0.019490425838571634
+next_action,Next-Action Prediction,macro_f1,0.01058201058201058,0.010709504685408301,-0.0001274941033977215,0.0017301038062283738,0.006006006006006006,-0.004576004576004574,0.0058479532163742695,-0.00473405736563631
+hand_trajectory_forecast,Future Hand Motion Forecasting,mae,4.466395378112793,4.303755283355713,-0.16264009475708008,3.1172122955322266,4.305870532989502,0.16052484512329102,4.1367621421813965,0.3296332359313965
+contact_prediction,Contact State Prediction,macro_f1,1.0,1.0,0.0,1.0,1.0,0.0,1.0,0.0
+object_relevance,Relevant Object Prediction,micro_f1,0.15813953488372093,0.14793328498912256,0.010206249894598368,0.15894868585732164,0.17871759890859482,0.020578064024873888,0.18262653898768813,0.024487004103967203
+caption_grounding,Language-to-Time Grounding,mrr,0.03208567947149277,0.027228528633713722,0.004857150837779045,0.014815197326242924,0.02484782598912716,-0.007237853482365608,0.02719014883041382,-0.004895530641078949
+cross_modal_retrieval,Cross-Modal Window Retrieval,mrr,0.3751238286495209,0.38921058177948,-0.014086753129959106,0.01806792803108692,0.32749155163764954,-0.04763227701187134,0.31795138120651245,-0.05717244744300842
+modality_reconstruction,Sensor-to-Visual Reconstruction,mae,9.79421329498291,10.446661949157715,0.6524486541748047,2.6225292682647705,8.830678939819336,0.9635343551635742,8.392388343811035,1.401824951171875
+temporal_order,Temporal Order Verification,macro_f1,0.5172413793103449,0.4942528735632184,0.022988505747126464,0.5028735632183908,0.5301714439065678,0.012930064596222923,0.5330450130569861,0.015803633746641288
+misalignment_detection,Cross-Modal Misalignment Detection,macro_f1,0.41734045375379186,0.42258557365378524,-0.005245119899993378,0.47823544277887897,0.44378951880827355,0.02644906505448169,0.4373795761078998,0.02003912235410793

scripts/audio_ablation_and_raw_upgrade.py

@@ -0,0 +1,954 @@
+#!/usr/bin/env python3
+"""Audio ablation and raw-audio feature upgrade for the Xperience-10M task suite.
+
+This script is artifact-driven where possible. It consumes the committed
+single-episode task-suite windows and feature manifest, decodes the real AAC
+stream from the local public sample MP4, and writes measured task deltas.
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import math
+import shutil
+import subprocess
+import sys
+from collections import OrderedDict
+from pathlib import Path
+from typing import Iterable
+
+import numpy as np
+
+from single_episode_diagnostics import (
+    TASKS,
+    TASK_DISPLAY,
+    block_indices,
+    chronological_split,
+    classification_metrics,
+    encode_labels,
+    frame_centers,
+    labels_from_windows,
+    load_inputs,
+    multilabel_metrics,
+    onehot,
+    read_csv,
+    regression_metrics,
+    retrieval_metrics,
+    ridge_predict,
+    standardize,
+    transition_labels_from_boundaries,
+    write_csv,
+    write_json,
+)
+
+
+VARIANTS = [
+    "all_handcrafted_audio",
+    "all_except_audio",
+    "handcrafted_audio_only",
+    "raw_logmel_audio_only",
+    "replace_handcrafted_with_raw",
+    "all_plus_raw_logmel",
+]
+
+VARIANT_DISPLAY = {
+    "all_handcrafted_audio": "All Current Features",
+    "all_except_audio": "All Except Audio",
+    "handcrafted_audio_only": "Handcrafted AAC Audio Only",
+    "raw_logmel_audio_only": "Raw Log-Mel Audio Only",
+    "replace_handcrafted_with_raw": "Replace AAC Block With Raw Log-Mel",
+    "all_plus_raw_logmel": "All Current Features + Raw Log-Mel",
+}
+
+PRIMARY_METRIC_HIGHER_IS_BETTER = {
+    "timeline_action": True,
+    "timeline_subtask": True,
+    "transition_detection": True,
+    "next_action": True,
+    "hand_trajectory_forecast": False,
+    "contact_prediction": True,
+    "object_relevance": True,
+    "caption_grounding": True,
+    "cross_modal_retrieval": True,
+    "modality_reconstruction": False,
+    "temporal_order": True,
+    "misalignment_detection": True,
+}
+
+
+def parse_args() -> argparse.Namespace:
+    root = Path(__file__).resolve().parents[1]
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--workspace", type=Path, default=root)
+    parser.add_argument("--suite-dir", type=Path, default=root / "results/episode_task_suite")
+    parser.add_argument("--output-dir", type=Path, default=root / "results/audio_ablation")
+    parser.add_argument("--raw-sample-dir", type=Path, default=None)
+    parser.add_argument("--annotation", type=Path, default=None)
+    parser.add_argument("--homie-toolkit", type=Path, default=None)
+    parser.add_argument("--audio-source", default="fisheye_cam0.mp4")
+    parser.add_argument("--sample-rate", type=int, default=16000)
+    parser.add_argument("--mel-bands", type=int, default=64)
+    parser.add_argument("--fft-size", type=int, default=512)
+    parser.add_argument("--hop-length", type=int, default=160)
+    parser.add_argument("--ridge-l2", type=float, default=10.0)
+    parser.add_argument("--test-fraction", type=float, default=0.30)
+    parser.add_argument("--future-offset-windows", type=int, default=4)
+    parser.add_argument("--forecast-frames", type=int, default=10)
+    parser.add_argument("--misalignment-shift-windows", type=int, default=8)
+    parser.add_argument("--force", action="store_true")
+    return parser.parse_args()
+
+
+def infer_raw_sample_dir(workspace: Path, explicit: Path | None) -> Path | None:
+    if explicit is not None:
+        return explicit.expanduser().resolve()
+    candidates = [
+        workspace / "data/sample/xperience-10m-sample",
+        workspace.parent / "data/sample/xperience-10m-sample",
+        Path.home() / "Library/CloudStorage/Dropbox/Ropedia/data/sample/xperience-10m-sample",
+    ]
+    for candidate in candidates:
+        if (candidate / "fisheye_cam0.mp4").exists():
+            return candidate.resolve()
+    return None
+
+
+def infer_homie_toolkit(raw_sample_dir: Path | None, explicit: Path | None) -> Path | None:
+    if explicit is not None:
+        return explicit.expanduser().resolve()
+    candidates = []
+    if raw_sample_dir is not None:
+        for parent in raw_sample_dir.parents:
+            candidates.append(parent / "HOMIE-toolkit")
+    candidates.append(Path.home() / "Library/CloudStorage/Dropbox/Ropedia/HOMIE-toolkit")
+    for candidate in candidates:
+        if candidate.exists():
+            return candidate.resolve()
+    return None
+
+
+def public_raw_sample_ref(path: Path | None) -> str:
+    if path is None:
+        return "not_available"
+    if path.name == "fisheye_cam0.mp4":
+        return "local_public_sample/fisheye_cam0.mp4"
+    if path.name == "annotation.hdf5":
+        return "local_public_sample/annotation.hdf5"
+    return f"local_public_sample/{path.name}"
+
+
+def decode_audio_mono(path: Path, sample_rate: int) -> np.ndarray:
+    if not path.exists() or shutil.which("ffmpeg") is None:
+        return np.zeros(0, dtype=np.float32)
+    cmd = [
+        "ffmpeg",
+        "-v",
+        "error",
+        "-i",
+        str(path),
+        "-vn",
+        "-ac",
+        "1",
+        "-ar",
+        str(sample_rate),
+        "-f",
+        "f32le",
+        "-",
+    ]
+    try:
+        proc = subprocess.run(cmd, check=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+    except (subprocess.CalledProcessError, FileNotFoundError):
+        return np.zeros(0, dtype=np.float32)
+    audio = np.frombuffer(proc.stdout, dtype=np.float32)
+    return np.nan_to_num(audio, nan=0.0, posinf=0.0, neginf=0.0).astype(np.float32)
+
+
+def video_fps(path: Path) -> float | None:
+    if not path.exists() or shutil.which("ffprobe") is None:
+        return None
+    cmd = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-select_streams",
+        "v:0",
+        "-show_entries",
+        "stream=avg_frame_rate,r_frame_rate",
+        "-of",
+        "json",
+        str(path),
+    ]
+    try:
+        proc = subprocess.run(cmd, check=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+        payload = json.loads(proc.stdout)
+    except (subprocess.CalledProcessError, FileNotFoundError, json.JSONDecodeError):
+        return None
+    streams = payload.get("streams") or []
+    for stream in streams:
+        for key in ("avg_frame_rate", "r_frame_rate"):
+            value = str(stream.get(key) or "")
+            if "/" in value:
+                num, den = value.split("/", 1)
+                try:
+                    fps = float(num) / max(float(den), 1e-12)
+                except ValueError:
+                    continue
+            else:
+                try:
+                    fps = float(value)
+                except ValueError:
+                    continue
+            if np.isfinite(fps) and fps > 0:
+                return fps
+    return None
+
+
+def hz_to_mel(hz: np.ndarray) -> np.ndarray:
+    return 2595.0 * np.log10(1.0 + hz / 700.0)
+
+
+def mel_to_hz(mel: np.ndarray) -> np.ndarray:
+    return 700.0 * (10.0 ** (mel / 2595.0) - 1.0)
+
+
+def mel_filterbank(sample_rate: int, fft_size: int, n_mels: int, f_min: float = 40.0) -> np.ndarray:
+    n_freqs = fft_size // 2 + 1
+    f_max = sample_rate / 2.0
+    mel_points = np.linspace(hz_to_mel(np.asarray([f_min]))[0], hz_to_mel(np.asarray([f_max]))[0], n_mels + 2)
+    hz_points = mel_to_hz(mel_points)
+    bins = np.floor((fft_size + 1) * hz_points / sample_rate).astype(int)
+    bins = np.clip(bins, 0, n_freqs - 1)
+    fb = np.zeros((n_mels, n_freqs), dtype=np.float32)
+    for i in range(n_mels):
+        left, center, right = int(bins[i]), int(bins[i + 1]), int(bins[i + 2])
+        if center <= left:
+            center = min(left + 1, n_freqs - 1)
+        if right <= center:
+            right = min(center + 1, n_freqs)
+        if center > left:
+            fb[i, left:center] = (np.arange(left, center) - left) / max(center - left, 1)
+        if right > center:
+            fb[i, center:right] = (right - np.arange(center, right)) / max(right - center, 1)
+    denom = fb.sum(axis=1, keepdims=True)
+    denom[denom < 1e-8] = 1.0
+    return fb / denom
+
+
+def stft_power(segment: np.ndarray, fft_size: int, hop_length: int) -> np.ndarray:
+    segment = np.asarray(segment, dtype=np.float32).reshape(-1)
+    if segment.size == 0:
+        return np.zeros((1, fft_size // 2 + 1), dtype=np.float32)
+    if segment.size < fft_size:
+        segment = np.pad(segment, (0, fft_size - segment.size))
+    n_frames = 1 + max(0, (segment.size - fft_size) // hop_length)
+    if n_frames <= 0:
+        n_frames = 1
+    window = np.hanning(fft_size).astype(np.float32)
+    frames = np.zeros((n_frames, fft_size), dtype=np.float32)
+    for i in range(n_frames):
+        start = i * hop_length
+        chunk = segment[start : start + fft_size]
+        if chunk.size < fft_size:
+            chunk = np.pad(chunk, (0, fft_size - chunk.size))
+        frames[i] = chunk * window
+    spec = np.fft.rfft(frames, n=fft_size, axis=1)
+    return (np.abs(spec) ** 2).astype(np.float32)
+
+
+def raw_audio_segment_embedding(segment: np.ndarray, sample_rate: int, mel_fb: np.ndarray, fft_size: int, hop_length: int) -> np.ndarray:
+    segment = np.asarray(segment, dtype=np.float32).reshape(-1)
+    if segment.size == 0:
+        return np.zeros(mel_fb.shape[0] * 9 + 12, dtype=np.float32)
+    segment = np.nan_to_num(segment, nan=0.0, posinf=0.0, neginf=0.0)
+    power = stft_power(segment, fft_size, hop_length)
+    mel = np.log1p(power @ mel_fb.T)
+    delta = np.diff(mel, axis=0) if mel.shape[0] > 1 else np.zeros_like(mel)
+    stats = [
+        mel.mean(axis=0),
+        mel.std(axis=0),
+        mel.min(axis=0),
+        mel.max(axis=0),
+        np.percentile(mel, 10, axis=0),
+        np.percentile(mel, 50, axis=0),
+        np.percentile(mel, 90, axis=0),
+        delta.mean(axis=0),
+        delta.std(axis=0),
+    ]
+
+    abs_seg = np.abs(segment)
+    rms = float(np.sqrt(np.mean(segment * segment)))
+    zcr = float(np.mean(segment[1:] * segment[:-1] < 0.0)) if segment.size > 1 else 0.0
+    energy = abs_seg.reshape(-1)
+    thirds = np.array_split(energy, 3)
+    third_means = [float(x.mean()) if len(x) else 0.0 for x in thirds]
+    waveform = np.asarray(
+        [
+            rms,
+            float(abs_seg.mean()),
+            float(abs_seg.std()),
+            float(abs_seg.max(initial=0.0)),
+            zcr,
+            float(np.log1p(np.mean(segment * segment))),
+            *third_means,
+            float(third_means[-1] - third_means[0]),
+            float(segment.size / max(sample_rate, 1)),
+            float(mel.shape[0]),
+        ],
+        dtype=np.float32,
+    )
+    return np.concatenate([*stats, waveform]).astype(np.float32)
+
+
+def extract_raw_audio_window_features(
+    audio_path: Path,
+    windows: list[dict],
+    n_frames: int,
+    output_dir: Path,
+    sample_rate: int,
+    mel_bands: int,
+    fft_size: int,
+    hop_length: int,
+    force: bool,
+) -> tuple[np.ndarray, dict]:
+    output_dir.mkdir(parents=True, exist_ok=True)
+    cache_path = output_dir / f"raw_logmel_{audio_path.stem}_sr{sample_rate}_mels{mel_bands}_fft{fft_size}_hop{hop_length}.npz"
+    if cache_path.exists() and not force:
+        data = np.load(cache_path, allow_pickle=True)
+        return data["features"].astype(np.float32), json.loads(str(data["metadata"].item()))
+
+    audio = decode_audio_mono(audio_path, sample_rate)
+    fps = video_fps(audio_path)
+    has_audio = bool(audio.size > 0)
+    if has_audio and fps is None:
+        fps = n_frames / max(audio.size / float(sample_rate), 1e-6)
+    mel_fb = mel_filterbank(sample_rate, fft_size, mel_bands)
+    feature_dim = mel_bands * 9 + 12
+    features = np.zeros((len(windows), feature_dim), dtype=np.float32)
+    if has_audio and fps is not None:
+        for i, row in enumerate(windows):
+            start_frame = int(row["start_frame"])
+            end_frame = int(row["end_frame"]) + 1
+            start_sample = int(round((start_frame / fps) * sample_rate))
+            end_sample = int(round((end_frame / fps) * sample_rate))
+            start_sample = max(0, min(start_sample, audio.size))
+            end_sample = max(start_sample + 1, min(end_sample, audio.size))
+            features[i] = raw_audio_segment_embedding(audio[start_sample:end_sample], sample_rate, mel_fb, fft_size, hop_length)
+            if i and i % 250 == 0:
+                print(f"    raw log-mel audio windows: {i}/{len(windows)}")
+
+    metadata = {
+        "source": public_raw_sample_ref(audio_path),
+        "exists": bool(audio_path.exists()),
+        "has_audio": has_audio,
+        "sample_rate": int(sample_rate),
+        "fps": float(fps) if fps is not None else None,
+        "num_samples": int(audio.size),
+        "num_windows": int(len(windows)),
+        "feature_dim": int(features.shape[1]),
+        "mel_bands": int(mel_bands),
+        "fft_size": int(fft_size),
+        "hop_length": int(hop_length),
+        "feature_description": "Per-window raw waveform STFT log-mel statistics plus delta and waveform envelope statistics.",
+    }
+    np.savez_compressed(cache_path, features=features, metadata=json.dumps(metadata, sort_keys=True))
+    return features, metadata
+
+
+def load_annotation(annotation: Path | None, toolkit: Path | None) -> dict | None:
+    if annotation is None or not annotation.exists() or toolkit is None or not toolkit.exists():
+        return None
+    sys.path.insert(0, str(toolkit))
+    from data_loader import load_from_annotation_hdf5
+
+    return load_from_annotation_hdf5(annotation, 0, None, load_slam_point_cloud=False)
+
+
+def object_targets_from_annotation(ann: dict | None, windows: list[dict]) -> dict | None:
+    if ann is None:
+        return None
+    frame_info = ann.get("caption_frame_info_map")
+    if frame_info is None:
+        return None
+    vocab: OrderedDict[str, int] = OrderedDict()
+    labels: list[list[str]] = []
+    for row in windows:
+        objects: OrderedDict[str, None] = OrderedDict()
+        for frame in range(int(row["start_frame"]), int(row["end_frame"]) + 1):
+            info = frame_info.get(frame, {})
+            raw_objects = info.get("objects")
+            if isinstance(raw_objects, list):
+                for obj in raw_objects:
+                    text = str(obj).strip()
+                    if text:
+                        objects.setdefault(text, None)
+            elif raw_objects:
+                text = str(raw_objects).strip()
+                if text:
+                    objects.setdefault(text, None)
+        obj_list = list(objects.keys())
+        for obj in obj_list:
+            if obj not in vocab:
+                vocab[obj] = len(vocab)
+        labels.append(obj_list)
+    if not vocab:
+        return None
+    Y = np.zeros((len(windows), len(vocab)), dtype=np.float32)
+    for i, obj_list in enumerate(labels):
+        for obj in obj_list:
+            Y[i, vocab[obj]] = 1.0
+    return {"Y": Y, "vocab": list(vocab.keys())}
+
+
+def exact_hand_targets_from_annotation(ann: dict | None, windows: list[dict], forecast_frames: int) -> tuple[np.ndarray, np.ndarray] | None:
+    if ann is None:
+        return None
+    left = ann.get("hand_left_joints")
+    right = ann.get("hand_right_joints")
+    body = ann.get("smplh_body_joints")
+    if left is None or right is None:
+        return None
+    valid, targets = [], []
+    n_frames = len(left)
+    for i, row in enumerate(windows):
+        future_start = int(row["end_frame"]) + 1
+        future_end = future_start + forecast_frames
+        if future_end > n_frames:
+            continue
+        hand = np.concatenate([left[future_start:future_end], right[future_start:future_end]], axis=1)
+        if body is not None and future_end <= len(body):
+            root = body[future_start:future_end, :1, :]
+            hand = hand - root
+        valid.append(i)
+        targets.append(hand.reshape(-1))
+    if not targets:
+        return None
+    return np.asarray(valid, dtype=np.int64), np.stack(targets).astype(np.float32)
+
+
+def exact_contact_labels_from_annotation(ann: dict | None, windows: list[dict]) -> np.ndarray | None:
+    if ann is None or ann.get("contacts") is None:
+        return None
+    contacts = ann["contacts"]
+    labels = []
+    for row in windows:
+        c = contacts[int(row["start_frame"]) : int(row["end_frame"]) + 1]
+        labels.append("contact" if np.any(c > 0) else "no_contact")
+    return np.asarray(labels, dtype=object)
+
+
+def exact_next_action_labels_from_annotation(ann: dict | None, windows: list[dict], future_frames: int = 20) -> np.ndarray | None:
+    if ann is None or ann.get("caption_frame_info_map") is None:
+        return None
+    frame_info = ann["caption_frame_info_map"]
+    n_frames = len(ann["img_names"])
+    labels = []
+    for row in windows:
+        future_frame = min(n_frames - 1, int(row["end_frame"]) + future_frames)
+        info = frame_info.get(future_frame, {})
+        label = info.get("action_label") or info.get("action") or ""
+        labels.append(str(label))
+    return np.asarray(labels, dtype=object)
+
+
+def setdiff_idx(a: np.ndarray, b: np.ndarray) -> np.ndarray:
+    return np.setdiff1d(np.asarray(a, dtype=np.int64), np.asarray(b, dtype=np.int64), assume_unique=False)
+
+
+def task_base_indices(task: str, manifest: list[dict]) -> np.ndarray:
+    audio = block_indices(manifest, ["audio_"])
+    caption = block_indices(manifest, ["caption_objects_interaction_text"])
+    contact = block_indices(manifest, ["body_contacts"])
+    all_idx = block_indices(manifest)
+    sensor = setdiff_idx(all_idx, caption)
+    if task in {"caption_grounding"}:
+        return sensor
+    if task in {"cross_modal_retrieval", "modality_reconstruction"}:
+        return block_indices(manifest, ["hand_", "body_joints", "body_contacts", "camera_", "imu_", "audio_"])
+    if task == "contact_prediction":
+        return setdiff_idx(sensor, contact)
+    if task == "object_relevance":
+        return sensor
+    return all_idx
+
+
+def feature_matrix_for_variant(task: str, variant: str, X: np.ndarray, raw_audio: np.ndarray, manifest: list[dict]) -> tuple[np.ndarray, str]:
+    base = task_base_indices(task, manifest)
+    audio = block_indices(manifest, ["audio_"])
+    base_no_audio = setdiff_idx(base, audio)
+    if variant == "all_handcrafted_audio":
+        return X[:, base], f"task contract feature blocks with handcrafted AAC audio where applicable ({len(base)} dims)"
+    if variant == "all_except_audio":
+        return X[:, base_no_audio], f"same task contract with handcrafted AAC audio columns removed ({len(base_no_audio)} dims)"
+    if variant == "handcrafted_audio_only":
+        return X[:, audio], f"handcrafted AAC audio block only ({len(audio)} dims)"
+    if variant == "raw_logmel_audio_only":
+        return raw_audio, f"raw waveform log-mel embedding only ({raw_audio.shape[1]} dims)"
+    if variant == "replace_handcrafted_with_raw":
+        return np.concatenate([X[:, base_no_audio], raw_audio], axis=1), (
+            f"task contract with handcrafted AAC removed and raw log-mel added ({len(base_no_audio) + raw_audio.shape[1]} dims)"
+        )
+    if variant == "all_plus_raw_logmel":
+        return np.concatenate([X[:, base], raw_audio], axis=1), (
+            f"task contract with existing handcrafted AAC plus raw log-mel ({len(base) + raw_audio.shape[1]} dims)"
+        )
+    raise KeyError(variant)
+
+
+def one_dim_target_standardize(train: np.ndarray, test: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
+    return standardize(train, test)
+
+
+def fit_classification_matrix(Xv: np.ndarray, labels: np.ndarray, train_idx: np.ndarray, test_idx: np.ndarray, l2: float) -> dict:
+    y, class_names = encode_labels(labels)
+    train_classes = set(int(x) for x in y[train_idx])
+    test_classes = set(int(x) for x in y[test_idx])
+    unseen = [class_names[i] for i in sorted(test_classes - train_classes)]
+    X_train, X_test = standardize(Xv[train_idx], Xv[test_idx])
+    scores = ridge_predict(X_train, onehot(y[train_idx], len(class_names)), X_test, l2)
+    pred = scores.argmax(axis=1)
+    metrics = classification_metrics(y[test_idx], pred)
+    metrics.update({
+        "num_classes": int(len(class_names)),
+        "num_train": int(len(train_idx)),
+        "num_test": int(len(test_idx)),
+        "unseen_test_classes": unseen,
+        "unseen_test_class_count": int(len(unseen)),
+    })
+    return metrics
+
+
+def fit_multilabel_matrix(Xv: np.ndarray, Y: np.ndarray, train_idx: np.ndarray, test_idx: np.ndarray, l2: float) -> dict:
+    X_train, X_test = standardize(Xv[train_idx], Xv[test_idx])
+    scores = ridge_predict(X_train, Y[train_idx], X_test, l2)
+    pred = (scores >= 0.5).astype(np.float32)
+    empty = np.where(pred.sum(axis=1) == 0)[0]
+    if len(empty):
+        pred[empty, np.argmax(scores[empty], axis=1)] = 1.0
+    metrics = multilabel_metrics(Y[test_idx], pred)
+    metrics.update({"num_objects": int(Y.shape[1]), "num_train": int(len(train_idx)), "num_test": int(len(test_idx))})
+    return metrics
+
+
+def fit_regression_matrix(Xv: np.ndarray, Y: np.ndarray, train_idx: np.ndarray, test_idx: np.ndarray, l2: float) -> dict:
+    X_train, X_test = standardize(Xv[train_idx], Xv[test_idx])
+    Y_train, Y_test = standardize(Y[train_idx], Y[test_idx])
+    pred = ridge_predict(X_train, Y_train, X_test, l2)
+    metrics = regression_metrics(Y_test, pred)
+    metrics.update({"num_train": int(len(train_idx)), "num_test": int(len(test_idx)), "target_dim": int(Y.shape[1])})
+    return metrics
+
+
+def fit_retrieval_matrix(Xv: np.ndarray, Y: np.ndarray, train_idx: np.ndarray, test_idx: np.ndarray, l2: float) -> dict:
+    X_train, X_test = standardize(Xv[train_idx], Xv[test_idx])
+    Y_train, Y_test = standardize(Y[train_idx], Y[test_idx])
+    pred = ridge_predict(X_train, Y_train, X_test, l2)
+    metrics = retrieval_metrics(pred, Y_test)
+    metrics.update({"num_train": int(len(train_idx)), "num_test": int(len(test_idx)), "target_dim": int(Y.shape[1])})
+    return metrics
+
+
+def pair_features_generic(F: np.ndarray, pairs: np.ndarray) -> np.ndarray:
+    left = F[pairs[:, 0]]
+    right = F[pairs[:, 1]]
+    return np.concatenate([left, right, right - left], axis=1).astype(np.float32)
+
+
+def misalignment_features(
+    variant: str,
+    X: np.ndarray,
+    raw_audio: np.ndarray,
+    manifest: list[dict],
+    pairs: np.ndarray,
+) -> tuple[np.ndarray, str]:
+    motion = block_indices(manifest, ["hand_", "body_joints", "body_contacts", "camera_", "imu_"])
+    visual_audio = block_indices(manifest, ["depth_confidence", "video_", "audio_"])
+    audio = block_indices(manifest, ["audio_"])
+    visual_no_audio = setdiff_idx(visual_audio, audio)
+    if variant == "all_handcrafted_audio":
+        left = X[pairs[:, 0]][:, motion]
+        right = X[pairs[:, 1]][:, visual_audio]
+        return np.concatenate([left, right], axis=1).astype(np.float32), "motion/current visual+handcrafted audio pair"
+    if variant == "all_except_audio":
+        left = X[pairs[:, 0]][:, motion]
+        right = X[pairs[:, 1]][:, visual_no_audio]
+        return np.concatenate([left, right], axis=1).astype(np.float32), "motion/current visual pair with audio removed"
+    if variant == "handcrafted_audio_only":
+        return pair_features_generic(X[:, audio], pairs), "handcrafted AAC audio self-alignment pair"
+    if variant == "raw_logmel_audio_only":
+        return pair_features_generic(raw_audio, pairs), "raw log-mel audio self-alignment pair"
+    if variant == "replace_handcrafted_with_raw":
+        left = X[pairs[:, 0]][:, motion]
+        right = np.concatenate([X[pairs[:, 1]][:, visual_no_audio], raw_audio[pairs[:, 1]]], axis=1)
+        return np.concatenate([left, right], axis=1).astype(np.float32), "motion/current visual pair with raw log-mel replacing handcrafted audio"
+    if variant == "all_plus_raw_logmel":
+        left = X[pairs[:, 0]][:, motion]
+        right = np.concatenate([X[pairs[:, 1]][:, visual_audio], raw_audio[pairs[:, 1]]], axis=1)
+        return np.concatenate([left, right], axis=1).astype(np.float32), "motion/current visual+handcrafted audio pair plus raw log-mel"
+    raise KeyError(variant)
+
+
+def task_target(
+    task: str,
+    X: np.ndarray,
+    windows: list[dict],
+    manifest: list[dict],
+    suite_dir: Path,
+    args: argparse.Namespace,
+    raw_targets: dict,
+) -> dict:
+    n = len(windows)
+    all_rows = np.arange(n, dtype=np.int64)
+    if task == "timeline_action":
+        return {"kind": "classification", "labels": labels_from_windows(windows, "action_label"), "rows": all_rows, "metric": "macro_f1"}
+    if task == "timeline_subtask":
+        return {"kind": "classification", "labels": labels_from_windows(windows, "subtask_label"), "rows": all_rows, "metric": "macro_f1"}
+    if task == "transition_detection":
+        labels = transition_labels_from_boundaries(suite_dir, frame_centers(windows))
+        return {"kind": "classification", "labels": labels, "rows": all_rows, "metric": "macro_f1"}
+    if task == "next_action":
+        labels = raw_targets.get("next_action_labels")
+        if labels is not None:
+            return {"kind": "classification", "labels": labels, "rows": all_rows, "metric": "macro_f1", "target_variant": "future action from annotation frame labels"}
+        rows = np.arange(0, n - args.future_offset_windows, dtype=np.int64)
+        labels = labels_from_windows(windows, "action_label")[rows + args.future_offset_windows]
+        return {"kind": "classification", "labels": labels, "rows": rows, "metric": "macro_f1", "target_variant": "future action from windows.csv"}
+    if task == "contact_prediction":
+        labels = raw_targets.get("contact_labels")
+        if labels is None:
+            contacts = block_indices(manifest, ["body_contacts"])
+            labels = np.where(np.abs(X[:, contacts]).sum(axis=1) > 1e-8, "contact", "no_contact")
+        return {"kind": "classification", "labels": labels, "rows": all_rows, "metric": "macro_f1"}
+    if task == "object_relevance":
+        obj = raw_targets.get("object_targets")
+        if obj is None:
+            return {"kind": "not_available", "reason": "object labels require local annotation.hdf5"}
+        return {"kind": "multilabel", "target": obj["Y"], "rows": all_rows, "metric": "micro_f1", "num_objects": len(obj["vocab"])}
+    if task == "hand_trajectory_forecast":
+        exact = raw_targets.get("hand_targets")
+        if exact is not None:
+            rows, target = exact
+            return {"kind": "regression", "target": target, "rows": rows, "metric": "mae", "target_variant": "future hand joints from annotation.hdf5"}
+        rows = np.arange(0, n - args.future_offset_windows, dtype=np.int64)
+        hand = block_indices(manifest, ["hand_left_joints", "hand_right_joints"])
+        return {"kind": "regression", "target": X[rows + args.future_offset_windows][:, hand], "rows": rows, "metric": "mae", "target_variant": "future hand feature block"}
+    if task == "caption_grounding":
+        text = block_indices(manifest, ["caption_objects_interaction_text"])
+        return {"kind": "retrieval", "target": X[:, text], "rows": all_rows, "metric": "mrr"}
+    if task in {"cross_modal_retrieval", "modality_reconstruction"}:
+        visual = block_indices(manifest, ["depth_confidence", "video_"])
+        return {"kind": "retrieval" if task == "cross_modal_retrieval" else "regression", "target": X[:, visual], "rows": all_rows, "metric": "mrr" if task == "cross_modal_retrieval" else "mae"}
+    if task == "temporal_order":
+        pairs, labels = [], []
+        for i in range(n - 1):
+            pairs.append((i, i + 1))
+            labels.append("forward")
+            pairs.append((i + 1, i))
+            labels.append("reversed")
+        return {"kind": "pair_classification", "pairs": np.asarray(pairs, dtype=np.int64), "labels": np.asarray(labels, dtype=object), "metric": "macro_f1"}
+    if task == "misalignment_detection":
+        pairs, labels = [], []
+        shift = args.misalignment_shift_windows
+        for i in range(n - shift):
+            pairs.append((i, i))
+            labels.append("aligned")
+            pairs.append((i, i + shift))
+            labels.append("shifted")
+        return {"kind": "misalignment", "pairs": np.asarray(pairs, dtype=np.int64), "labels": np.asarray(labels, dtype=object), "metric": "macro_f1"}
+    raise KeyError(task)
+
+
+def evaluate_task_variant(
+    task: str,
+    variant: str,
+    X: np.ndarray,
+    raw_audio: np.ndarray,
+    windows: list[dict],
+    manifest: list[dict],
+    suite_dir: Path,
+    args: argparse.Namespace,
+    raw_targets: dict,
+) -> dict:
+    info = task_target(task, X, windows, manifest, suite_dir, args, raw_targets)
+    row = {
+        "task": task,
+        "task_display": TASK_DISPLAY.get(task, task),
+        "variant": variant,
+        "variant_display": VARIANT_DISPLAY[variant],
+        "status": "computed",
+        "primary_metric": info.get("metric", ""),
+        "primary_value": "",
+        "higher_is_better": str(PRIMARY_METRIC_HIGHER_IS_BETTER[task]).lower(),
+        "feature_dim": "",
+        "num_train": "",
+        "num_test": "",
+        "input_contract": "",
+        "target_variant": info.get("target_variant", ""),
+        "reason": "",
+    }
+    if info["kind"] == "not_available":
+        row.update({"status": "not_computed", "reason": info["reason"]})
+        return row
+    try:
+        if info["kind"] == "misalignment":
+            feats, desc = misalignment_features(variant, X, raw_audio, manifest, np.asarray(info["pairs"], dtype=np.int64))
+            labels = np.asarray(info["labels"], dtype=object)
+            train_idx, test_idx = chronological_split(len(labels), args.test_fraction)
+            metrics = fit_classification_matrix(feats, labels, train_idx, test_idx, args.ridge_l2)
+            row["input_contract"] = desc
+        elif info["kind"] == "pair_classification":
+            F, desc = feature_matrix_for_variant(task, variant, X, raw_audio, manifest)
+            feats = pair_features_generic(F, np.asarray(info["pairs"], dtype=np.int64))
+            labels = np.asarray(info["labels"], dtype=object)
+            train_idx, test_idx = chronological_split(len(labels), args.test_fraction)
+            metrics = fit_classification_matrix(feats, labels, train_idx, test_idx, args.ridge_l2)
+            row["input_contract"] = desc
+        else:
+            F, desc = feature_matrix_for_variant(task, variant, X, raw_audio, manifest)
+            data_rows = np.asarray(info["rows"], dtype=np.int64)
+            train_idx, test_idx = chronological_split(len(data_rows), args.test_fraction)
+            if info["kind"] == "classification":
+                metrics = fit_classification_matrix(F[data_rows], np.asarray(info["labels"], dtype=object), train_idx, test_idx, args.ridge_l2)
+            elif info["kind"] == "multilabel":
+                metrics = fit_multilabel_matrix(F[data_rows], np.asarray(info["target"], dtype=np.float32), train_idx, test_idx, args.ridge_l2)
+            elif info["kind"] == "regression":
+                metrics = fit_regression_matrix(F[data_rows], np.asarray(info["target"], dtype=np.float32), train_idx, test_idx, args.ridge_l2)
+            elif info["kind"] == "retrieval":
+                metrics = fit_retrieval_matrix(F[data_rows], np.asarray(info["target"], dtype=np.float32), train_idx, test_idx, args.ridge_l2)
+            else:
+                raise KeyError(info["kind"])
+            row["input_contract"] = desc
+            row["feature_dim"] = int(F.shape[1])
+        row.update(metrics)
+        row["primary_value"] = float(metrics[info["metric"]])
+        row["num_train"] = int(metrics.get("num_train", row.get("num_train") or 0))
+        row["num_test"] = int(metrics.get("num_test", row.get("num_test") or 0))
+        if row["feature_dim"] == "":
+            row["feature_dim"] = int(feats.shape[1])
+    except Exception as exc:
+        row.update({"status": "not_computed", "reason": f"{type(exc).__name__}: {exc}"})
+    return row
+
+
+def delta(base: float, compare: float, higher_is_better: bool) -> float:
+    return compare - base if higher_is_better else base - compare
+
+
+def build_summary(rows: list[dict], raw_meta: dict) -> dict:
+    by_task: dict[str, dict[str, dict]] = {}
+    for row in rows:
+        if row.get("status") != "computed":
+            continue
+        by_task.setdefault(row["task"], {})[row["variant"]] = row
+    task_summaries = []
+    for task in TASKS:
+        variants = by_task.get(task, {})
+        base = variants.get("all_handcrafted_audio")
+        no_audio = variants.get("all_except_audio")
+        raw_only = variants.get("raw_logmel_audio_only")
+        replace = variants.get("replace_handcrafted_with_raw")
+        plus = variants.get("all_plus_raw_logmel")
+        if not base:
+            continue
+        higher = PRIMARY_METRIC_HIGHER_IS_BETTER[task]
+        item = {
+            "task": task,
+            "task_display": TASK_DISPLAY.get(task, task),
+            "primary_metric": base["primary_metric"],
+            "higher_is_better": higher,
+            "all_handcrafted_audio": float(base["primary_value"]),
+        }
+        if no_audio:
+            item["all_except_audio"] = float(no_audio["primary_value"])
+            item["handcrafted_audio_delta"] = delta(float(no_audio["primary_value"]), float(base["primary_value"]), higher)
+        if raw_only:
+            item["raw_logmel_audio_only"] = float(raw_only["primary_value"])
+        if replace and no_audio:
+            item["replace_handcrafted_with_raw"] = float(replace["primary_value"])
+            item["raw_replacement_delta_vs_no_audio"] = delta(float(no_audio["primary_value"]), float(replace["primary_value"]), higher)
+            item["raw_replacement_delta_vs_handcrafted"] = delta(float(base["primary_value"]), float(replace["primary_value"]), higher)
+        if plus:
+            item["all_plus_raw_logmel"] = float(plus["primary_value"])
+            item["all_plus_raw_delta_vs_handcrafted"] = delta(float(base["primary_value"]), float(plus["primary_value"]), higher)
+        task_summaries.append(item)
+
+    handcrafted_deltas = [x["handcrafted_audio_delta"] for x in task_summaries if "handcrafted_audio_delta" in x]
+    raw_replace_deltas = [x["raw_replacement_delta_vs_handcrafted"] for x in task_summaries if "raw_replacement_delta_vs_handcrafted" in x]
+    return {
+        "description": "Measured audio ablation and raw log-mel audio upgrade over the single public Xperience-10M sample episode.",
+        "scope": "single public sample episode; chronological split; ridge heads over fixed feature contracts",
+        "raw_audio_metadata": raw_meta,
+        "num_tasks": len(task_summaries),
+        "variants": VARIANT_DISPLAY,
+        "task_summaries": task_summaries,
+        "aggregate": {
+            "mean_handcrafted_audio_delta": float(np.mean(handcrafted_deltas)) if handcrafted_deltas else None,
+            "tasks_where_handcrafted_audio_improves": int(sum(1 for x in handcrafted_deltas if x > 0)),
+            "mean_raw_replacement_delta_vs_handcrafted": float(np.mean(raw_replace_deltas)) if raw_replace_deltas else None,
+            "tasks_where_raw_replacement_improves_over_handcrafted": int(sum(1 for x in raw_replace_deltas if x > 0)),
+        },
+    }
+
+
+def write_summary_markdown(path: Path, summary: dict) -> None:
+    lines = [
+        "# Audio Ablation and Raw-Audio Upgrade",
+        "",
+        "This report is generated from committed task-suite artifacts plus the local public-sample MP4 audio stream.",
+        "It measures whether audio changes each single-episode task under the same chronological split.",
+        "",
+        "## Raw Audio Feature",
+        "",
+    ]
+    meta = summary["raw_audio_metadata"]
+    lines.extend([
+        f"- Source: `{meta.get('source')}`",
+        f"- Has audio: `{meta.get('has_audio')}`",
+        f"- Sample rate: `{meta.get('sample_rate')}`",
+        f"- Window feature dim: `{meta.get('feature_dim')}`",
+        f"- Feature: {meta.get('feature_description')}",
+        "",
+        "## Task Deltas",
+        "",
+        "| Task | Metric | Current audio | No audio | Current audio delta | Raw replaces audio | Raw replacement delta |",
+        "| --- | --- | ---: | ---: | ---: | ---: | ---: |",
+    ])
+    for item in summary["task_summaries"]:
+        lines.append(
+            "| {task} | {metric} | {cur:.4f} | {no:.4f} | {d1:.4f} | {raw:.4f} | {d2:.4f} |".format(
+                task=item["task_display"],
+                metric=item["primary_metric"],
+                cur=item.get("all_handcrafted_audio", float("nan")),
+                no=item.get("all_except_audio", float("nan")),
+                d1=item.get("handcrafted_audio_delta", float("nan")),
+                raw=item.get("replace_handcrafted_with_raw", float("nan")),
+                d2=item.get("raw_replacement_delta_vs_handcrafted", float("nan")),
+            )
+        )
+    agg = summary["aggregate"]
+    lines.extend([
+        "",
+        "## Aggregate",
+        "",
+        f"- Mean current-audio delta: `{agg['mean_handcrafted_audio_delta']}`",
+        f"- Tasks where current handcrafted audio improves the primary metric: `{agg['tasks_where_handcrafted_audio_improves']}`",
+        f"- Mean raw-replacement delta vs current handcrafted audio: `{agg['mean_raw_replacement_delta_vs_handcrafted']}`",
+        f"- Tasks where raw log-mel replacement improves over current handcrafted audio: `{agg['tasks_where_raw_replacement_improves_over_handcrafted']}`",
+        "",
+        "Positive deltas always mean better according to each task's primary metric. For MAE tasks, lower MAE is converted into a positive improvement.",
+        "",
+    ])
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text("\n".join(lines), encoding="utf-8")
+
+
+def write_delta_chart(path: Path, summary: dict) -> None:
+    items = summary["task_summaries"]
+    width = 1320
+    row_h = 42
+    height = 120 + row_h * len(items)
+    max_abs = max([abs(x.get("handcrafted_audio_delta", 0.0)) for x in items] + [1e-6])
+    left = 410
+    mid = 680
+    scale = 240 / max_abs
+    lines = [
+        f'<svg xmlns="http://www.w3.org/2000/svg" width="{width}" height="{height}" viewBox="0 0 {width} {height}">',
+        '<rect width="100%" height="100%" fill="#07110d"/>',
+        '<text x="36" y="42" fill="#e6f7ea" font-family="Arial, sans-serif" font-size="28" font-weight="700">Measured Audio Delta Across 12 Xperience-10M Tasks</text>',
+        '<text x="36" y="70" fill="#a7b8ab" font-family="Arial, sans-serif" font-size="15">Positive means audio improved the task primary metric on the single public sample split.</text>',
+        f'<line x1="{mid}" y1="92" x2="{mid}" y2="{height - 24}" stroke="#5b6f61" stroke-width="1"/>',
+    ]
+    for i, item in enumerate(items):
+        y = 112 + i * row_h
+        task = item["task_display"].replace("&", "&amp;")
+        value = float(item.get("handcrafted_audio_delta", 0.0))
+        bar_w = abs(value) * scale
+        x = mid if value >= 0 else mid - bar_w
+        color = "#7ae5c3" if value >= 0 else "#ff8a6a"
+        lines.extend([
+            f'<text x="36" y="{y + 18}" fill="#d8eadc" font-family="Arial, sans-serif" font-size="15">{task}</text>',
+            f'<rect x="{x:.2f}" y="{y}" width="{bar_w:.2f}" height="22" rx="3" fill="{color}"/>',
+            f'<text x="{mid + 270}" y="{y + 17}" fill="#d8eadc" font-family="Arial, sans-serif" font-size="14">{value:+.4f} {item["primary_metric"]}</text>',
+        ])
+    lines.append("</svg>")
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text("\n".join(lines), encoding="utf-8")
+
+
+def main() -> int:
+    args = parse_args()
+    args.output_dir.mkdir(parents=True, exist_ok=True)
+    raw_sample_dir = infer_raw_sample_dir(args.workspace, args.raw_sample_dir)
+    audio_path = raw_sample_dir / args.audio_source if raw_sample_dir is not None else Path(args.audio_source)
+    annotation = args.annotation or (raw_sample_dir / "annotation.hdf5" if raw_sample_dir is not None else None)
+    toolkit = infer_homie_toolkit(raw_sample_dir, args.homie_toolkit)
+
+    if raw_sample_dir is None or not audio_path.exists():
+        raise FileNotFoundError("Local public sample MP4 is required for raw-audio upgrade. Pass --raw-sample-dir.")
+    if shutil.which("ffmpeg") is None:
+        raise RuntimeError("ffmpeg is required to decode the MP4 audio stream.")
+
+    X, _starts, _ends, windows, manifest, _summary = load_inputs(args.suite_dir)
+    raw_audio, raw_meta = extract_raw_audio_window_features(
+        audio_path,
+        windows,
+        X.shape[0],
+        args.output_dir,
+        args.sample_rate,
+        args.mel_bands,
+        args.fft_size,
+        args.hop_length,
+        args.force,
+    )
+    ann = load_annotation(annotation, toolkit)
+    raw_targets = {
+        "object_targets": object_targets_from_annotation(ann, windows),
+        "hand_targets": exact_hand_targets_from_annotation(ann, windows, args.forecast_frames),
+        "contact_labels": exact_contact_labels_from_annotation(ann, windows),
+        "next_action_labels": exact_next_action_labels_from_annotation(ann, windows),
+    }
+
+    rows: list[dict] = []
+    for task in TASKS:
+        print(f"Audio ablation task: {task}")
+        for variant in VARIANTS:
+            rows.append(evaluate_task_variant(task, variant, X, raw_audio, windows, manifest, args.suite_dir, args, raw_targets))
+
+    write_csv(args.output_dir / "audio_ablation_metrics.csv", rows)
+    summary = build_summary(rows, raw_meta)
+    summary["provenance"] = {
+        "suite_dir": "results/episode_task_suite",
+        "shared_windows": "results/episode_task_suite/shared_windows.npz",
+        "feature_manifest": "results/episode_task_suite/feature_manifest.json",
+        "audio_source": public_raw_sample_ref(audio_path),
+        "annotation_source": public_raw_sample_ref(annotation) if annotation is not None and annotation.exists() else "not_available",
+        "homie_toolkit_available": bool(toolkit is not None and toolkit.exists()),
+    }
+    write_json(args.output_dir / "audio_ablation_summary.json", summary)
+    write_summary_markdown(args.output_dir / "AUDIO_ABLATION_SUMMARY.md", summary)
+    write_delta_chart(args.workspace / "docs/assets/charts/audio_ablation_delta.svg", summary)
+    write_json(args.workspace / "docs/data/audio_ablation_summary.json", summary)
+
+    compact_rows = []
+    for item in summary["task_summaries"]:
+        compact_rows.append({
+            "task": item["task"],
+            "task_display": item["task_display"],
+            "metric": item["primary_metric"],
+            "current_audio": item.get("all_handcrafted_audio", ""),
+            "no_audio": item.get("all_except_audio", ""),
+            "current_audio_delta": item.get("handcrafted_audio_delta", ""),
+            "raw_audio_only": item.get("raw_logmel_audio_only", ""),
+            "replace_with_raw": item.get("replace_handcrafted_with_raw", ""),
+            "raw_replacement_delta_vs_current": item.get("raw_replacement_delta_vs_handcrafted", ""),
+            "all_plus_raw": item.get("all_plus_raw_logmel", ""),
+            "all_plus_raw_delta_vs_current": item.get("all_plus_raw_delta_vs_handcrafted", ""),
+        })
+    write_csv(args.output_dir / "audio_delta_summary.csv", compact_rows)
+    print(f"Wrote {args.output_dir}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/build_artifact_index.py

@@ -185,6 +185,46 @@ ARTIFACTS = [
         "surface": "repo_hf",
         "shows": "Regenerates the research takeaways from committed summary metrics and task result artifacts.",
     },
+    {
+        "id": "audio_ablation_script",
+        "title": "Audio ablation and raw-audio upgrade script",
+        "path": "scripts/audio_ablation_and_raw_upgrade.py",
+        "kind": "result_interpretation",
+        "surface": "repo_hf",
+        "shows": "Measures current AAC audio contribution and a raw log-mel audio feature replacement across all 12 task contracts.",
+    },
+    {
+        "id": "audio_ablation_summary",
+        "title": "Audio ablation summary",
+        "path": "results/audio_ablation/audio_ablation_summary.json",
+        "kind": "metrics_source",
+        "surface": "repo_hf",
+        "shows": "Stores per-task audio deltas for all current features, no-audio, handcrafted-audio-only, raw-audio-only, raw replacement, and all-plus-raw variants.",
+    },
+    {
+        "id": "audio_ablation_summary_md",
+        "title": "Audio ablation summary report",
+        "path": "results/audio_ablation/AUDIO_ABLATION_SUMMARY.md",
+        "kind": "result_interpretation",
+        "surface": "repo_hf",
+        "shows": "Human-readable table showing the measured audio contribution and raw-audio replacement delta for every task.",
+    },
+    {
+        "id": "audio_ablation_website_json",
+        "title": "Audio ablation website JSON",
+        "path": "docs/data/audio_ablation_summary.json",
+        "kind": "website_data",
+        "surface": "website_hf",
+        "shows": "Machine-readable audio ablation summary mirrored into the static website and Hugging Face bundles.",
+    },
+    {
+        "id": "audio_ablation_delta_chart",
+        "title": "Audio ablation delta chart",
+        "path": "docs/assets/charts/audio_ablation_delta.svg",
+        "kind": "visual_evidence",
+        "surface": "website_hf",
+        "shows": "Bar chart of measured current-audio primary-metric deltas across the 12 tasks.",
+    },
     {
         "id": "figure_index",
         "title": "Figure index",

scripts/build_research_takeaways.py

@@ -10,6 +10,7 @@ from pathlib import Path
 
 ROOT = Path(__file__).resolve().parents[1]
 SUMMARY_PATH = ROOT / "docs/data/summary_metrics.json"
+AUDIO_PATH = ROOT / "docs/data/audio_ablation_summary.json"
 OUTPUT_JSON = ROOT / "docs/data/research_takeaways.json"
 OUTPUT_MD = ROOT / "RESEARCH_TAKEAWAYS.md"
 
@@ -36,6 +37,7 @@ def task_metric(tasks: dict, task: str, key: str) -> float:
 
 def build_payload() -> dict:
     summary = json.loads(SUMMARY_PATH.read_text(encoding="utf-8"))
+    audio_summary = json.loads(AUDIO_PATH.read_text(encoding="utf-8")) if AUDIO_PATH.exists() else None
     suite = summary["suite"]
     tasks = suite["tasks"]
     neural = suite.get("neural_tasks", {})
@@ -124,6 +126,44 @@ def build_payload() -> dict:
             "source": "results/episode_task_suite/cross_modal_retrieval/metrics.json",
             "current_scope": "The current reconstruction task predicts feature vectors; depth, mesh, NeRF, and Gaussian-splatting outputs are future task variants.",
         },
+    ]
+
+    if audio_summary is not None:
+        audio_aggregate = audio_summary["aggregate"]
+        modality_recon = next(
+            (item for item in audio_summary["task_summaries"] if item["task"] == "modality_reconstruction"),
+            {},
+        )
+        object_relevance = next(
+            (item for item in audio_summary["task_summaries"] if item["task"] == "object_relevance"),
+            {},
+        )
+        takeaways.append(
+            {
+                "id": "audio_contribution_is_task_specific",
+                "title": "Audio helps some tasks and hurts others on the public sample",
+                "readout": (
+                    "The current AAC audio block improves the primary metric on 6 of 12 tasks, "
+                    "while raw log-mel replacement improves over the current handcrafted block on 6 of 12 tasks. "
+                    "The largest current-audio gain appears in feature reconstruction, not in action classification."
+                ),
+                "evidence": [
+                    {"label": "tasks_where_current_audio_improves", "value": audio_aggregate["tasks_where_handcrafted_audio_improves"]},
+                    {"label": "mean_current_audio_delta", "value": audio_aggregate["mean_handcrafted_audio_delta"]},
+                    {"label": "tasks_where_raw_replacement_improves", "value": audio_aggregate["tasks_where_raw_replacement_improves_over_handcrafted"]},
+                    {"label": "mean_raw_replacement_delta_vs_current", "value": audio_aggregate["mean_raw_replacement_delta_vs_handcrafted"]},
+                    {"label": "reconstruction_current_audio_delta", "value": modality_recon.get("handcrafted_audio_delta")},
+                    {"label": "object_relevance_current_audio_delta", "value": object_relevance.get("handcrafted_audio_delta")},
+                ],
+                "source": "results/audio_ablation/audio_ablation_summary.json",
+                "current_scope": (
+                    "This is a single-episode ablation over fixed ridge heads. It validates that audio is wired into the task suite "
+                    "and shows where it changes metrics; it does not prove cross-episode audio generalization."
+                ),
+            }
+        )
+
+    takeaways.append(
         {
             "id": "scale_requires_episodes",
             "title": "The next scientific unit is held-out episodes, not more adjacent windows",
@@ -141,8 +181,8 @@ def build_payload() -> dict:
                 "current_scope",
                 "The 32-episode fine-tune requires gated data staging and held-out evaluation.",
             ),
-        },
-    ]
+        }
+    )
 
     return {
         "title": "Ropedia Xperience-10M Research Takeaways",
@@ -152,6 +192,7 @@ def build_payload() -> dict:
             "docs/data/summary_metrics.json",
             "results/episode_task_suite/summary_report.json",
             "results/episode_task_suite/neural_mlp/*/metrics.json",
+            "docs/data/audio_ablation_summary.json",
             "results/omni_finetune/MULTI_EPISODE_ACCESS_STATUS.md",
         ],
         "scope": {
@@ -215,8 +256,9 @@ def render_md(payload: dict) -> str:
             "",
             "- High single-episode scores are useful pipeline checks for the current task contracts.",
             "- Low chronological action/subtask scores are informative because they expose later-label shift.",
-            "- Neural gains on trajectory/order/alignment make those tasks good candidates for the next fine-tuning stage.",
-            "- Retrieval and reconstruction remain the main multimodal representation challenges.",
+        "- Neural gains on trajectory/order/alignment make those tasks good candidates for the next fine-tuning stage.",
+        "- Audio ablation is task-specific: current AAC and raw log-mel features help some probes and hurt others.",
+        "- Retrieval and reconstruction remain the main multimodal representation challenges.",
             "- The next credible model-quality result needs held-out episodes.",
             "",
         ]

scripts/publish_hf_bundles.py

@@ -69,6 +69,10 @@ STALE_MODEL_REMOTE_FILES = [
     "metrics/" + LEGACY_SCORECARD_JSON,
 ]
 
+ARTIFACT_BINARY_ALLOWLIST = [
+    "results/audio_ablation/raw_logmel_fisheye_cam0_sr16000_mels64_fft512_hop160.npz",
+]
+
 
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(description=__doc__)
@@ -171,6 +175,29 @@ def delete_remote_folder_if_present(
         print(f"Remote stale-folder cleanup skipped for {repo_id}/{path_in_repo}: {exc}")
 
 
+def upload_allowlisted_artifact_binaries(
+    api: HfApi,
+    token: str,
+    repo_id: str,
+    artifact_root: Path,
+) -> None:
+    """Upload approved derived binary artifacts without exposing model weights."""
+    for relative_path in ARTIFACT_BINARY_ALLOWLIST:
+        path = artifact_root / relative_path
+        if not path.exists():
+            print(f"Allowlisted artifact binary absent: {relative_path}")
+            continue
+        api.upload_file(
+            path_or_fileobj=str(path),
+            path_in_repo=relative_path,
+            repo_id=repo_id,
+            repo_type="dataset",
+            token=token,
+            commit_message=f"Publish derived artifact {relative_path}",
+        )
+        print(f"Uploaded allowlisted artifact binary: {repo_id}/{relative_path}")
+
+
 def main() -> int:
     args = parse_args()
     hf_root = args.hf_root.resolve()
@@ -214,6 +241,7 @@ def main() -> int:
         "Publish Ropedia Xperience-10M derived artifacts",
         ignore_patterns=["**/*.pt", "**/*.npz"],
     )
+    upload_allowlisted_artifact_binaries(api, token, artifact_repo, hf_root / "artifacts")
     for path_in_repo in STALE_ARTIFACT_REMOTE_FILES:
         delete_remote_file_if_present(api, token, artifact_repo, "dataset", path_in_repo)
     for path_in_repo in STALE_ARTIFACT_REMOTE_FOLDERS:

scripts/validate_mirror_parity.py

@@ -20,6 +20,7 @@ DEFAULT_HF_ROOT = ROOT.parent / "hf_publish"
 DEFAULT_OUTPUT = ROOT / "docs/data/mirror_parity.json"
 
 DATA_FILES = [
+    "audio_ablation_summary.json",
     "artifact_index.json",
     "brand_assets.json",
     "evidence_contract.json",
@@ -52,6 +53,7 @@ DATA_FILES = [
 ]
 
 ASSET_FILES = [
+    "charts/audio_ablation_delta.svg",
     "brand/xperience10m-logo-apple-touch.png",
     "brand/xperience10m-logo-favicon-32.png",
     "brand/xperience10m-logo-favicon-64.png",
@@ -72,6 +74,7 @@ ASSET_FILES = [
 ]
 
 SCRIPT_FILES = [
+    "audio_ablation_and_raw_upgrade.py",
     "build_artifact_index.py",
     "build_brand_assets.py",
     "build_evaluation_protocol.py",
@@ -103,6 +106,11 @@ WEBSITE_FILES = [
 ]
 
 RESULT_FILES = [
+    "audio_ablation/AUDIO_ABLATION_SUMMARY.md",
+    "audio_ablation/audio_ablation_metrics.csv",
+    "audio_ablation/audio_ablation_summary.json",
+    "audio_ablation/audio_delta_summary.csv",
+    "audio_ablation/raw_logmel_fisheye_cam0_sr16000_mels64_fft512_hop160.npz",
     "single_episode_diagnostics/provenance.json",
     "single_episode_diagnostics/README.md",
     "single_episode_diagnostics/modality_ablation/ablation_metrics.csv",