README.md

---
license: other
pretty_name: Ropedia Xperience-10M Task Suite Artifacts
tags:
  - robotics
  - embodied-ai
  - multimodal
  - ropedia
  - xperience-10m
  - evaluation
  - baseline
  - neural-network
  - pytorch
  - retrieval
task_categories:
  - robotics
  - time-series-forecasting
  - text-retrieval
language:
  - en
size_categories:
  - n<1K
---

# Ropedia Xperience-10M Task Suite Artifacts

This dataset repo contains the derived research artifact layer for the public
Xperience-10M sample episode released by Ropedia. It is meant for inspecting and
extending the project without downloading raw videos first: task manifests,
metrics, predictions, charts, diagrams, notes, reproduction scripts, modality
metadata, and compact neural task-head artifacts.

![Ropedia Xperience-10M Task Suite logo](assets/brand/xperience10m-logo-social-card.png)

![12-task infographic](assets/task_suite_infographic.png?v=xperience10m-taskfirst-v13-modality-xl)

The logo, figures, cards, and website assets are packaged together so the
artifact repo reads as a cohesive Xperience-10M multimodal task suite. Labels,
dimensions, and metrics are generated from committed result files rather than
hand-edited presentation copy.

The Space is organized as a two-level research dashboard: five top-level tabs
plus subsection tabs for dataset, task-suite, method, result, and resource
views. It still foregrounds the task-first 12-task map and includes a native
responsive modality atlas backed by
`docs/data/modality_atlas.json` and `docs/assets/modalities/`, so each
public-sample stream remains readable on mobile without shipping raw videos or
annotations.
The website task section now reads from `docs/data/task_walkthroughs.json` to
render common research task names, larger task cards, and an interactive scrub/play walkthrough storyboard.
`docs/data/task_surface_integrity.json` verifies that those task cards stay
human-readable, use representative modality thumbnails, and keep the
walkthrough storyboard wired to the generated task metadata.

The artifact bundle now includes
`docs/data/xperience10m_dataset_card_alignment.json` and
`XPERIENCE10M_DATASET_CARD_ALIGNMENT.md`, which align this project's wording
with the official gated `ropedia-ai/xperience-10m` dataset card: manually
reviewed access, full-scale 4D modality coverage, episode layout, intended
uses, limitations, and current project coverage. The same artifact now records the
public sample card (`cc-by-nc-4.0`, HOMIE Toolkit, Rerun 0.29.0 `.rrd`
visualization) and the observed HF API listing snapshot: 803 session folders
and 12,103 episode folders with `annotation.hdf5`, plus the live HF 31.9 TB
file-size display. The 31.9 TB display is tracked separately from the official
card's about-1PB full-scale storage statement. Those counts are upstream
metadata only, not files redistributed in this artifact dataset. The same
source note preserves the official limited in diversity / showcase-quality
disclaimer and excludes identity, surveillance, biometric, sensitive-attribute,
and safety-critical uses.
The generated source-alignment report, `SOURCE_ALIGNMENT_AUDIT.md` plus
`docs/data/source_alignment_audit.json`, checks those full-dataset facts,
public sample-card facts, API-listing caveats, and current-project
markers across the repo, website, and Hugging Face cards.

For first-pass reading, `PROJECT_STATUS.md` and
`docs/data/project_status.json` provide the compact current project state:
implemented public-sample pipeline, 12 task contracts, minimal and neural task
heads, source-aligned dataset wording, data-gated Qwen3-Omni scale-up, and
excluded raw data.
`PROJECT_BRIEF.md` and `docs/data/project_brief.json` provide the shortest
front-door view: what exists now, what the public sample can support, and what
must happen before the 32-episode omni-model stage becomes a held-out
evaluation.
`docs/research_roadmap.html`, `RESEARCH_ROADMAP.md`,
`docs/data/research_roadmap.json`, and
`docs/data/research_roadmap_interactive.json` show how the 12 public-sample
tasks connect to the four research directions, multi-episode data staging, the
32-episode Qwen3-Omni LoRA pilot, robustness runs, and larger omni-model
extensions.
`EVALUATION_PROTOCOL.md` and `docs/data/evaluation_protocol.json` define the
window unit, chronological split, leakage controls, per-task metrics, and
current limitations before a reader compares scores.
`RESEARCH_TAKEAWAYS.md` and `docs/data/research_takeaways.json`, regenerated by
`scripts/build_research_takeaways.py`, summarize what the committed metrics
actually show: chronological class shift, neural gains on
dynamics/order/alignment, harder retrieval/reconstruction probes, and the need
for held-out episodes before final model metrics.
`FIGURE_INDEX.md` and `docs/data/figure_index.json` catalog the public figures,
charts, modality thumbnails, dimensions, stable hashes, and source scripts.
`docs/data/brand_assets.json` catalogs the generated logo variants used for the
favicon, header, README/HF cards, app icon, and social preview.
`PUBLIC_SURFACE_QA.md` and `docs/data/public_surface_qa.json` describe the
public project surface across the repo, website, and Hugging Face cards.

It does **not** contain raw Xperience-10M videos or raw `annotation.hdf5`. Download raw data only from the official Ropedia / Hugging Face sources and follow their terms.

Current scale-up status: the full `ropedia-ai/xperience-10m` Hugging Face
dataset is still gated for this account. The multi-episode workflow is prepared
to select, download, validate, and stage a 32-episode held-out pilot after
access approval. Until that completes, the committed Qwen3-Omni artifacts
remain setup-stage evidence; 32-episode held-out metrics require the full pilot.

## Why This Repo Exists

This is the explorable artifact half of the project. You can inspect the task outputs, compare the committed metrics, and understand the single-episode limitations without downloading the raw videos first.

The latest bundle also includes `docs/single_episode_explorer.html` and
`docs/data/single_episode_explorer.json`, a static window-level explorer for
labels, object sets, predictions, feature-block statistics, modality ablations,
and alignment stress diagnostics.

## 90-Second Research Project Path

| Step | Question | Primary artifacts |
| --- | --- | --- |
| 1 | What is the project in one page? | `PROJECT_BRIEF.md`, `docs/data/project_brief.json` |
| 2 | What has been implemented? | `PROJECT_STATUS.md`, `docs/data/project_status.json`, `EVIDENCE_CONTRACT.md`, `ARTIFACT_GUIDE.md`, `QUALITY_GATES.md`, `PUBLIC_SURFACE_QA.md`, `FIGURE_INDEX.md`, `docs/data/evidence_contract.json`, `docs/data/artifact_index.json`, `docs/data/figure_index.json`, `docs/data/live_publication_status.json`, `docs/data/quality_gates.json`, `docs/data/mirror_parity.json`, `docs/data/public_surface_qa.json`, `docs/data/scope_claims_audit.json`, `docs/data/publication_audit.json`, `docs/data/task_surface_integrity.json`, `docs/data/rendered_site_check.json`, `docs/data/website_integrity.json` |
| 3 | What is the staged roadmap? | `docs/research_roadmap.html`, `RESEARCH_ROADMAP.md`, `docs/data/research_roadmap.json`, `docs/data/research_roadmap_interactive.json` |
| 4 | Are source facts consistently presented? | `SOURCE_ALIGNMENT_AUDIT.md`, `docs/data/source_alignment_audit.json`, `scripts/validate_source_alignment.py` |
| 5 | What do the current results mean? | `RESEARCH_TAKEAWAYS.md`, `docs/data/research_takeaways.json`, `docs/data/summary_metrics.json` |
| 6 | How do I reproduce it? | `REPRODUCIBILITY.md`, `docs/data/reproducibility_matrix.json`, `notes/reproducibility_audit.md` |
| 7 | What is one model input? | `results/episode_task_suite/windows.csv`, `results/episode_task_suite/feature_manifest.json`, `results/episode_task_suite/available_modalities.json` |
| 8 | Are the task results backed by files? | `results/episode_task_suite/summary_report.json`, `results/episode_task_suite/neural_mlp/`, `docs/data/summary_metrics.json` |
| 9 | Can I inspect one window interactively? | `docs/single_episode_explorer.html`, `docs/data/single_episode_explorer.json`, `results/single_episode_diagnostics/` |
| 10 | What is still pending? | `results/omni_finetune/DATA_ACCESS_STATUS.md`, `results/omni_finetune/MULTI_EPISODE_ACCESS_STATUS.md`, `scripts/omni/discover_xperience10m_sources.py` |

Human-readable artifact guide: `ARTIFACT_GUIDE.md`.
Project brief: `PROJECT_BRIEF.md` and `docs/data/project_brief.json`.
Project status: `PROJECT_STATUS.md` and `docs/data/project_status.json`.
Research roadmap: `docs/research_roadmap.html`, `RESEARCH_ROADMAP.md`,
`docs/data/research_roadmap.json`, and
`docs/data/research_roadmap_interactive.json`.
Research Takeaways: `RESEARCH_TAKEAWAYS.md` and `docs/data/research_takeaways.json`.
Multi-episode data status: `results/omni_finetune/DATA_ACCESS_STATUS.md`.
Official dataset-card alignment: `XPERIENCE10M_DATASET_CARD_ALIGNMENT.md` and `docs/data/xperience10m_dataset_card_alignment.json`.
Source alignment: `SOURCE_ALIGNMENT_AUDIT.md` and `docs/data/source_alignment_audit.json`.
Release checks: `QUALITY_GATES.md` and `docs/data/quality_gates.json`.
Public project surface: `PUBLIC_SURFACE_QA.md` and `docs/data/public_surface_qa.json`.
Live publication status: `docs/data/live_publication_status.json`.
Machine-readable project packet: `docs/data/project_packet.json`.
Source-of-truth artifact index: `docs/data/artifact_index.json`.
Source-of-truth figure index: `FIGURE_INDEX.md` and `docs/data/figure_index.json`.
Source-of-truth brand asset index: `docs/data/brand_assets.json`.

## Current Research Scope And Supporting Checks

| Project layer | Evidence | Current scope |
| --- | --- | --- |
| Project status | `PROJECT_STATUS.md`, `docs/data/project_status.json` | compact current-state table |
| Research roadmap | `docs/research_roadmap.html`, `RESEARCH_ROADMAP.md`, `docs/data/research_roadmap.json`, `docs/data/research_roadmap_interactive.json` | interactive and machine-readable path from public-sample task development to multi-episode and larger omni-model work |
| Data windows | `results/episode_task_suite/windows.csv`, `shared_windows.npz`, `summary_report.json` | one public sample episode |
| Feature contract | `results/episode_task_suite/feature_manifest.json`, `available_modalities.json` | 8,378 current features; audio documented but not featurized |
| Evaluation protocol | `EVALUATION_PROTOCOL.md`, `docs/data/evaluation_protocol.json` | windowing, chronological split, leakage controls, and task metrics |
| Research Takeaways | `RESEARCH_TAKEAWAYS.md`, `docs/data/research_takeaways.json`, `scripts/build_research_takeaways.py` | generated interpretation of the committed metrics and scale-up status |
| 12-task suite | per-task `metrics.json`, predictions, confusion matrices | chronological single-episode split |
| Neural heads | `results/episode_task_suite/neural_mlp/` | compact MLP heads, not a foundation model |
| Research directions | `research_direction_taxonomy.json`, extension probe results | direct/proxy/diagnostic evidence, not full solutions |
| Task surface integrity | `docs/data/task_surface_integrity.json`, `scripts/validate_task_surface.py` | public cards use human-readable names, modality thumbnails, and the walkthrough/player data contract |
| Rendered website check | `RENDERED_SITE_CHECK.md`, `docs/data/rendered_site_check.json`, `scripts/build_rendered_site_check.py` | browser-level load, tab, walkthrough deep-link, control-click, and console-health check |
| Qwen3-Omni | `DATA_ACCESS_STATUS.md`, `MULTI_EPISODE_ACCESS_STATUS.md` | setup-stage until 32 valid episodes are available |
| Multi-episode pilot status | `docs/data/scope_claims_audit.json`, `scripts/validate_scope_claims.py` | setup-stage `32ep` artifacts are recorded separately from completed held-out-episode metrics |
| Mirror parity | `docs/data/mirror_parity.json`, `scripts/validate_mirror_parity.py` | prepared repo/HF mirrors carry matching critical data, figures, website HTML, and validator files |
| Public bundle contents | `docs/data/publication_audit.json`, `scripts/validate_publication_package.py` | public files and HF bundles, with current-card checks |
| Public project surface | `PUBLIC_SURFACE_QA.md`, `docs/data/public_surface_qa.json`, `scripts/build_public_surface_qa.py` | repo, website, and Hugging Face cards preserve SEO/social metadata, accessible tab semantics, public links, and reader-facing copy |
| Website integrity | `docs/data/website_integrity.json`, `scripts/validate_website_integrity.py` | local links, anchors, JSON bundles, and referenced images only |
| Release checks | `QUALITY_GATES.md`, `docs/data/quality_gates.json`, `scripts/build_quality_gates.py` | automated release checks plus live post-publish verification |
| Live publication | `docs/data/live_publication_status.json`, `scripts/verify_live_publication.py` | last public GitHub/HF URL verification after upload |
| Official dataset card alignment | `XPERIENCE10M_DATASET_CARD_ALIGNMENT.md`, `docs/data/xperience10m_dataset_card_alignment.json` | official source scope, public sample card, HF API listing, gated access, modality coverage, scale, and this repo's single-episode scope |
| Source alignment | `SOURCE_ALIGNMENT_AUDIT.md`, `docs/data/source_alignment_audit.json`, `scripts/validate_source_alignment.py` | validates full-dataset facts, sample-card facts, API-listing notes, and project coverage |
| Brand assets | `assets/brand/`, `docs/assets/brand/`, `scripts/build_brand_assets.py` | Generated project logo system packaged for favicon, header, card, README, and social preview use |
| Figure index | `FIGURE_INDEX.md`, `docs/data/figure_index.json`, `scripts/build_figure_index.py` | public figures, charts, modality thumbnails, dimensions, hashes, and generation provenance |
| Artifact index | `docs/data/artifact_index.json`, `scripts/build_artifact_index.py` | compact project-artifact catalog with stable hashes |
| Citation metadata | `PROJECT_README.md`, `docs/data/project_manifest.json`, GitHub `CITATION.cff` | code and dataset license terms remain explicit |

## What Is Included

- `ARTIFACT_GUIDE.md`: human-readable map of project scope, data contract, task evidence, platform mirrors, and scale-up status
- `PROJECT_STATUS.md` and `docs/data/project_status.json`: compact current-state decision table
- `docs/research_roadmap.html`, `RESEARCH_ROADMAP.md`, `docs/data/research_roadmap.json`, and `docs/data/research_roadmap_interactive.json`: staged multi-episode and omni-model roadmap
- `REPRODUCIBILITY.md` and `docs/data/reproducibility_matrix.json`: public commands, expected outputs, exact-match reproduction evidence, and current scale-up requirements
- `EVALUATION_PROTOCOL.md` and `docs/data/evaluation_protocol.json`: generated task protocol, split policy, leakage controls, and current limitations
- `RESEARCH_TAKEAWAYS.md` and `docs/data/research_takeaways.json`: generated metric interpretation and scale-up readout
- `results/**/*.json`: verified metrics and metadata for minimal and neural MLP runs
- `results/**/*.csv`: predictions, confusion matrices, per-class metrics, windows, boundaries
- `results/**/history.json`: neural MLP training traces
- `docs/assets/*.svg` and `docs/assets/*.png`: generated diagrams, charts, and overview figures
- `docs/assets/brand/` and `assets/brand/`: generated project logo mark, favicon variants, apple-touch icon, and social card
- `docs/assets/task_suite_infographic.png`: task-suite infographic with the shared processing contract, all 12 task families, verified metric overlays, and enlarged public-sample modality thumbnails below the task map
- `docs/assets/modalities/` and `docs/data/modality_atlas.json`: small derived sample thumbnails and metadata for the responsive modality atlas
- `docs/data/summary_metrics.json`: dashboard-readable summary bundle
- `docs/data/evidence_contract.json`: machine-readable project scope
- `XPERIENCE10M_DATASET_CARD_ALIGNMENT.md` and `docs/data/xperience10m_dataset_card_alignment.json`: official Xperience-10M dataset-card, public sample-card, and HF API metadata alignment summary
- `SOURCE_ALIGNMENT_AUDIT.md` and `docs/data/source_alignment_audit.json`: generated report that source facts, sample details, API-listing notes, and project coverage are preserved across public surfaces
- `FIGURE_INDEX.md` and `docs/data/figure_index.json`: visual evidence index for public figures, charts, thumbnails, dimensions, hashes, and source scripts
- `docs/data/artifact_index.json`: source-of-truth project-artifact catalog with stable-file hashes
- `docs/data/mirror_parity.json`: prepared Space/artifact/model mirror parity check, including critical website HTML
- `docs/data/scope_claims_audit.json`: machine-readable multi-episode pilot status report for historical `32ep` setup/provenance identifiers
- `docs/data/publication_audit.json`: machine-readable public bundle and public-card freshness report
- `PUBLIC_SURFACE_QA.md` and `docs/data/public_surface_qa.json`: public project-surface report for the repo, website, and Hugging Face cards
- `docs/data/task_surface_integrity.json`: machine-readable task-card and walkthrough-player report
- `RENDERED_SITE_CHECK.md` and `docs/data/rendered_site_check.json`: browser-level website interaction check
- `docs/data/website_integrity.json`: machine-readable website local-reference report
- `QUALITY_GATES.md` and `docs/data/quality_gates.json`: human-readable and machine-readable release checks
- `docs/data/live_publication_status.json`: last live public URL verification after upload
- `docs/single_episode_explorer.html` and `docs/data/single_episode_explorer.json`: static interactive explorer for one sample episode
- `results/single_episode_diagnostics/`: modality ablations, object-label export, timeline overlay, and alignment stress results
- `docs/data/project_manifest.json`: machine-readable public URL and citation metadata
- `docs/data/project_packet.json`: machine-readable project path and scope summary
- `results/omni_finetune/DATA_ACCESS_STATUS.md`: reader-facing multi-episode data requirement for the Qwen3-Omni pilot
- `docs/data/research_directions.json`: generated four-track taxonomy for the website
- `docs/data/research_direction_extensions.json`: four extra data-backed probes, one per research direction
- `docs/data/task_walkthroughs.json`: human-readable task names, modality links, input/process/output contracts, and walkthrough-player data for all 12 tasks
- `results/episode_task_suite/research_directions/`: JSON, CSV, and Markdown task-to-research-track mapping
- `results/episode_task_suite/research_direction_extensions/`: metrics, prediction CSVs, rank CSVs, and Markdown summary for the four extension probes
- `results/episode_task_suite/task_walkthroughs/`: case-study walkthroughs for every task contract
- `scripts/*.py`: reproduction scripts
- `scripts/export_modality_atlas_assets.py`: regenerates the responsive modality-card thumbnails and manifest from the local public sample
- `scripts/build_artifact_index.py`: source-of-truth artifact-index builder
- `scripts/build_research_takeaways.py`: regenerates Research Takeaways from committed metric artifacts
- `scripts/validate_mirror_parity.py`: prepared mirror parity validator
- `scripts/validate_scope_claims.py`: keeps Qwen3-Omni setup status separate from completed 32-episode results
- `scripts/validate_publication_package.py`: public bundle validator
- `scripts/build_public_surface_qa.py`: public project-surface report builder
- `scripts/build_rendered_site_check.py`: rendered website check builder
- `scripts/validate_website_integrity.py`: website local-reference validator
- `notes/*.md`: interpretation and reproducibility notes

The companion model repo stores the lightweight model checkpoints and mirrors
the binary arrays (`model.npz`, `model.pt`, and compact neural prediction
arrays). This artifact dataset stays focused on inspectable CSV/JSON/Markdown,
scripts, notes, and visual assets:

https://huggingface.co/cy0307/ropedia-xperience-10m-task-baselines

## Links

| Resource | URL |
| --- | --- |
| Hugging Face Space | https://huggingface.co/spaces/cy0307/ropedia-xperience-10m-task-suite |
| Live Hugging Face app | https://cy0307-ropedia-xperience-10m-task-suite.static.hf.space/ |
| Hugging Face collection | https://huggingface.co/collections/cy0307/ropedia-xperience-10m-task-suite |
| Minimal and neural task baseline repo | https://huggingface.co/cy0307/ropedia-xperience-10m-task-baselines |
| GitHub repo | https://github.com/ChaoYue0307/ropedia-xperience-10m-task-suite |
| GitHub Pages dashboard | https://chaoyue0307.github.io/ropedia-xperience-10m-task-suite/ |
| Xperience-10M website | https://ropedia.com/dataset |
| Xperience-10M release page | https://ropedia.com/blog/20260316_xperience_10m |
| Ropedia GitHub organization | https://github.com/Ropedia |
| HOMIE Toolkit | https://github.com/Ropedia/HOMIE-toolkit |
| Xperience-10M dataset | https://huggingface.co/datasets/ropedia-ai/xperience-10m |
| Xperience-10M sample | https://huggingface.co/datasets/ropedia-ai/xperience-10m-sample |

## Scope

The artifacts validate one public sample episode:

- 5,821 aligned frames
- 1,161 sliding windows
- 8,378 current feature dimensions, with audio documented but not featurized
- 12 human-readable supervised/self-supervised task cards
- minimal linear/ridge baselines and neural MLP heads for all 12 tasks
- four direction-extension probes with minimal and neural MLP baselines
- chronological 70/30 split

For research conclusions, rerun the same scripts over many episodes and evaluate on held-out episodes.

## Neural MLP Result Snapshot

These are single-episode chronological-split metrics. They are useful for checking task definitions and input contracts; cross-episode conclusions require held-out episodes.

| Task | Neural metric | Minimal metric |
| --- | ---: | ---: |
| Action Recognition macro-F1 | 0.0263 | 0.0500 |
| Procedure Step Recognition macro-F1 | 0.0175 | 0.0495 |
| Action Boundary Detection macro-F1 | 0.6485 | 0.6552 |
| Next-Action Prediction macro-F1 | 0.0235 | 0.0593 |
| Hand Trajectory Forecasting MPJPE, lower is better | 0.1116 | 0.8223 |
| Contact State Prediction macro-F1 | 1.0000 | 1.0000 |
| Object Relevance Prediction micro-F1 | 0.1798 | 0.1839 |
| Language Grounding MRR | 0.0178 | 0.0172 |
| Cross-Modal Retrieval MRR | 0.1530 | 0.2634 |
| Cross-Modal Reconstruction R2 | -0.0102 | -0.0160 |
| Temporal Order Verification F1 | 0.8718 | 0.5487 |
| Multimodal Synchronization Detection F1 | 0.7335 | 0.4866 |

Primary NN artifact path:

`results/episode_task_suite/neural_mlp/<task>/`

## Four Research Directions

The current 12 tasks are organized into the four Ropedia research directions
with two baselines per task: minimal interpretable heads and neural MLP heads.

| Direction | Current status | Evidence |
| --- | --- | --- |
| A. Human Modeling & Motion Understanding | partially implemented | hand trajectory and contact are direct; action/object tasks are proxies |
| B. 3D/4D Reconstruction & Neural Rendering | proxy tasks only | retrieval, reconstruction, and misalignment diagnose prerequisites |
| C. Egocentric Vision & Interaction | strongest implemented track | 6 direct tasks plus order/alignment diagnostics |
| D. Scene Reconstruction & World Modeling | early proxy tasks | state, object, retrieval, reconstruction, and temporal probes |

Primary taxonomy artifact:

`results/episode_task_suite/research_directions/research_direction_taxonomy.json`

## Four Direction-Extension Probes

The artifact bundle also includes one extra coded probe for each Ropedia
research direction. These are still single-episode diagnostics, but they make
the four-direction roadmap concrete.

| Direction | Extension task | Minimal | Neural MLP |
| --- | --- | ---: | ---: |
| A. Human Modeling & Motion Understanding | Body and Hand Motion Intensity | 0.7827 macro-F1 | 0.7986 macro-F1 |
| B. 3D/4D Reconstruction & Neural Rendering | Multi-View Consistency Retrieval | 0.5534 MRR | 0.3469 MRR |
| C. Egocentric Vision & Interaction | Action Phase Progress Estimation | 0.3416 MAE | 0.3038 MAE |
| D. Scene Reconstruction & World Modeling | Short-Horizon Ego-Motion Forecasting | 0.1989 MAE | 0.0989 MAE |

Primary extension artifact:

`results/episode_task_suite/research_direction_extensions/research_direction_extension_results.json`

## Task Walkthroughs

Each task has a human-readable research name, task card, case study, input
contract, middle process modules, output contract, modality list, metric, and
current limitation. The website mirrors these records as an interactive
scrub/play walkthrough storyboard for onboarding a junior researcher or
engineer:

`results/episode_task_suite/task_walkthroughs/TASK_WALKTHROUGHS.md`

## Pending 32-Episode Pilot

| Item | Value |
| --- | --- |
| Selection strategy | stratified round-robin across top-level session UUIDs |
| Candidate scan | first 64 top-level session UUIDs |
| Valid complete candidates | 680 |
| Selected pilot episodes | 32 from 32 session UUIDs |
| Estimated raw subset | about 72.0 GB |
| Excluded file type | `visualization.rrd` |
| Current access status | HF gated dataset approval pending |

![Verified episode pipeline](assets/pipeline_diagram.png)

![Minimal task architectures](assets/task_architectures.png)