Preserve per-episode task outcome labels and provenance

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+metadata/episode_outcomes.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -137,6 +137,30 @@ The archive members contain the exported video and state/action bytes. Episode m
 reduced to public data fields. `dataset_info.json` summarizes the release and `shards.jsonl`
 lists archive sizes, episode counts, and SHA-256 checksums.
 
+## Episode outcomes
+
+`metadata/episode_outcomes.jsonl` provides one metadata record per episode, joined by
+`episode_id`. The fields are `episode_outcome` (`success`, `failure`, or `unknown`)
+and `episode_outcome_source` (label provenance). A schema is provided alongside it.
+
+All **134,537** retained episodes have `episode_outcome: "success"`, with provenance
+`source_directory:success_episodes`. Every retained episode maps to an upstream
+`success_episodes` directory; no source-labelled failure episodes occur in this subset.
+This preserves the source task-outcome label. It is not independent visual verification
+or a certification of pose, action, or video quality. Processing failures and alignment
+review flags are separate from task outcomes.
+
+The existing TAR episode JSON members predate this metadata addition. Merge the overlay
+into each episode's metadata when loading; video and binary payloads are unchanged:
+
+```python
+with open("metadata/episode_outcomes.jsonl", encoding="utf-8") as stream:
+    outcome_by_id = {row["episode_id"]: row for row in map(json.loads, stream)}
+metadata.update(outcome_by_id[metadata["episode_id"]])
+```
+
+`dataset_info.json` records the overlay counts and SHA-256 checksum.
+
 ## Source and license
 
 Upstream sources are [X-Humanoid/RoboMIND2.0-Agilex](https://modelscope.cn/datasets/X-Humanoid/RoboMIND2.0-Agilex),

dataset_info.json

@@ -47,5 +47,21 @@
       "hours": 33.24141068607515
     }
   ],
-  "format": "WebDataset tar shards; one .mp4, .bin and .json per episode"
+  "format": "WebDataset tar shards; one .mp4, .bin and .json per episode",
+  "episode_outcomes": {
+    "file": "metadata/episode_outcomes.jsonl",
+    "schema": "metadata/episode_outcomes.schema.json",
+    "join_key": "episode_id",
+    "field": "episode_outcome",
+    "provenance_field": "episode_outcome_source",
+    "episodes": 134537,
+    "counts": {
+      "success": 134537,
+      "failure": 0,
+      "unknown": 0
+    },
+    "sha256": "3a996aab4c28ca0dfc82a2ad5806bee01ff378adac38ebf9bb323109cde4edf2",
+    "note": "All retained episodes are labelled successful by the upstream success_episodes directory. This preserves the source label and is not independent visual verification.",
+    "archive_members_updated": false
+  }
 }

metadata/episode_outcomes.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a996aab4c28ca0dfc82a2ad5806bee01ff378adac38ebf9bb323109cde4edf2
+size 24082123

metadata/episode_outcomes.schema.json

@@ -0,0 +1,29 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "title": "Per-episode task outcome metadata overlay",
+  "type": "object",
+  "additionalProperties": false,
+  "required": [
+    "episode_id",
+    "episode_outcome",
+    "episode_outcome_source"
+  ],
+  "properties": {
+    "episode_id": {
+      "type": "string",
+      "pattern": "^episode-[0-9a-f]{64}$"
+    },
+    "episode_outcome": {
+      "type": "string",
+      "enum": [
+        "success",
+        "failure",
+        "unknown"
+      ]
+    },
+    "episode_outcome_source": {
+      "type": "string"
+    }
+  },
+  "description": "Join to the archive episode JSON by episode_id. Source labels describe task outcomes, not processing or action-quality checks. Unknown means no explicit source outcome was provided; do not treat it as success or failure."
+}