docs: correct pinned loading and follow-up scope

README.md

@@ -1,9 +1,8 @@
 ---
 pretty_name: V-JEPA Reproduce Recipe
 license: other
-task_categories:
-  - video-classification
-  - feature-extraction
+license_name: mixed-upstream-terms
+license_link: https://huggingface.co/datasets/BootsofLagrangian/V-JEPA-Reproduce-Recipe/blob/main/README.md#license-and-access
 tags:
   - video
   - metadata
@@ -19,7 +18,7 @@ V-JEPA-style observation-pretraining mixture. This repository does **not**
 redistribute source videos and does not grant rights beyond each upstream
 dataset's terms.
 
-## Current release
+## Release being published
 
 - Manifest ID: `YTWS-CANONICAL-MANIFEST-V1`
 - Run ID: `823483f781094a668cf6682828723d16`
@@ -31,11 +30,11 @@ dataset's terms.
 - Parquet payload: `41,557,148,598` bytes (38.70 GiB)
 
 The manifest combines the 11 currently materialized metadata sources: COIN,
-DROID, EPIC-KITCHENS-100, HD-VILA-100M, HowTo100M ID recovery, InternVid,
-Kinetics-710, OpenVid-1M, Panda-70M, YouNiverse, and YT-Temporal-180M.
-Provenance quality is recorded per source and ranges from official releases to
-explicitly marked mirrors. Dataset membership is not a statement that every
-media URL remains reachable.
+DROID, EPIC-KITCHENS-100, HD-VILA-100M, HowTo100M ID recovery (unattested ID
+mirror), InternVid, Kinetics-710, OpenVid-1M, Panda-70M, YouNiverse (structural
+HF mirror), and YT-Temporal-180M (unverified mirror). Provenance quality is
+recorded per source. Dataset membership is not a statement that every media URL
+remains reachable.
 
 ## Layout
 
@@ -53,26 +52,35 @@ runs/<run_id>/source_coverage.parquet
 - `samples`: exact `(platform, video_id, millisecond interval)` identities.
 - `assets`: source video or trajectory identities without interval duplication.
 - `memberships`: lossless lineage back to each source metadata row.
-- `dataset_roles`: overlapping semantic views such as `broad_in_the_wild`,
-  `instructional`, and `ego_and_synchronized_observation`.
+- `dataset_roles`: the overlapping `broad_in_the_wild`, `instructional`,
+  `curated_temporal_action`, `ego_and_synchronized_observation`, and
+  `hosted_quality_video` views.
 - `dataset_stats`: per-source row, dedup, reservation, and resolvability counts.
 - `source_coverage`: included and excluded catalog sources with provenance or
   exclusion reasons.
 
-Captions, action labels, and other large source-specific payloads are not
-copied into every canonical row. They join through `dataset_id`,
-`source_row_id`, and `metadata_revision` in `memberships`.
+Captions, action labels, and other source-specific payload tables are not part
+of this repository. Membership rows retain `dataset_id`, `source_row_id`, and
+`metadata_revision` as lineage keys for consumers that separately acquired the
+corresponding upstream payload. The HowTo100M recovery contains IDs only; the
+canonical captions/tasks archive was unavailable and cannot be reconstructed
+from this release.
 
 ## Load with Hugging Face Datasets
 
-Read one of the 256 hash buckets per worker instead of materializing all 470M
-sample rows in one process:
+Authentication is required because the repository is private; export a token
+with read access as `HF_TOKEN`. Bare `load_dataset(repo_id)` is unsupported
+because the repository intentionally contains tables with different schemas.
+Always select explicit `data_files`.
+
+Assign the 256 hash buckets across workers so each worker opens only its own
+shards:
 
 ```python
 from datasets import load_dataset
 
 repo_id = "BootsofLagrangian/V-JEPA-Reproduce-Recipe"
-revision = "main"  # replace with the pinned publication commit for training
+revision = "<PINNED_HUB_COMMIT>"
 run_id = "823483f781094a668cf6682828723d16"
 bucket = "00"
 
@@ -80,13 +88,12 @@ samples = load_dataset(
     "parquet",
     data_files={
         "train": (
-            f"hf://datasets/{repo_id}/runs/{run_id}/samples/"
+            f"hf://datasets/{repo_id}@{revision}/runs/{run_id}/samples/"
             f"dedup_bucket={bucket}/*.parquet"
         )
     },
     split="train",
     streaming=True,
-    revision=revision,
 )
 
 web_train = samples.filter(
@@ -106,13 +113,12 @@ memberships = load_dataset(
     "parquet",
     data_files={
         "train": (
-            f"hf://datasets/{repo_id}/runs/{run_id}/memberships/"
+            f"hf://datasets/{repo_id}@{revision}/runs/{run_id}/memberships/"
             f"dataset_id={dataset_id}/**/*.parquet"
         )
     },
     split="train",
     streaming=True,
-    revision=revision,
 )
 ```
 
@@ -146,6 +152,10 @@ fraction or dataset quality.
 - Retry the same failed YouTube smoke IDs with an official yt-dlp EJS runtime
   and a bounded native-download-then-local-trim fallback.
 - Add a range-aware EPIC media adapter and repeat the same selected ID.
+- Run a separately registered, bounded, stratified random-asset reachability
+  measurement for every included source. Preserve the selected asset as the
+  inferential unit, prohibit candidate replacement, and report per-source
+  success fractions with the egress, timestamp, adapter, and uncertainty.
 - Recover or explicitly retain exclusions for unavailable, gated, or legally
   withheld sources, including YT-Temporal-1B, LVD-142M, SSv2, Ego4D,
   Ego-Exo4D, and WebVid.