Document Pyramids, Worm, and Walker release-23 datasets

README.md

@@ -28,6 +28,9 @@ and discrete action spaces for testing Causal GPT-RL across space types.
 | `unity/soccer-twos/expert-selfplay-v0` | **`r23 / r23 / r23`** | 14,588 | 1,000,816 | ego `Dict` wrapping `Tuple(Box(264), Box(72))` | ego `Dict` wrapping `MultiDiscrete([3, 3, 3])` |
 | `unity/dungeon-escape/expert-poca-v0` | **`r23 / r23 / r23`** | 35,787 | 1,003,523 | ego `Dict` wrapping `Tuple(Box(10), Box(360), Box(1))` | ego `Dict` wrapping `Discrete(7)` |
 | `unity/3dball-hard/expert-v0` | `r20 / r20 / r20` | 1,022 | 1,000,008 | `Tuple(Box(27), Box(18))` | `Box(2)` |
+| `unity/pyramids/expert-v0` | **`r23 / r23 / r23`** | 5,348 | 1,000,000 | `Tuple(Box(56), Box(56), Box(56), Box(4))` | `Discrete(5)` |
+| `unity/worm/expert-v0` | **`r23 / r23 / r23`** | 1,000 | 1,000,000 | `Box(64)` | `Box(9, [-1, 1])` |
+| `unity/walker/expert-v0` | **`r23 / r23 / r23`** | 1,458 | 1,000,010 | `Box(243)` | `Box(39, [-1, 1])` |
 
 ### ML-Agents release boundary
 
@@ -39,8 +42,8 @@ The table reports three independent provenance axes in this order:
 - **Model** — ML-Agents release from which the behavior policy model originated.
 
 At this repository revision, all three axes are aligned for every dataset.
-Crawler, PushBlock, SoccerTwos, and DungeonEscape expert datasets are aligned
-on `release_23`; 3DBallHard remains on `release_20`. The migration to `release_23` will be incremental,
+Crawler, PushBlock, Pyramids, Worm, Walker, SoccerTwos, and DungeonEscape
+expert datasets are aligned on `release_23`; 3DBallHard remains on `release_20`. The migration to `release_23` will be incremental,
 and future rows may intentionally contain mixed values. Consumers must inspect
 all three fields instead of inferring compatibility from the dataset name or
 the shared `v0` suffix. See `release_manifest.json` for the machine-readable
@@ -160,6 +163,22 @@ NaN/Inf/shape/action-range errors. Every match links exactly three episodes.
    and 99 truncated at the recording boundary), mean return 4.972084, and mean
    episode length 18.9390.
 
+## Pyramids, Worm, and Walker release_23
+
+Use each model-removed release-23 build with its matching unmodified stock
+release-23 ONNX policy. Record each physical scene agent independently and
+stop at one million transitions. The published schemas come directly from
+the live ML-Agents behavior specs.
+
+- **Pyramids:** 5,348 episodes, 1,000,000 transitions,
+  mean return 1.795309, mean length 186.9858,
+  5,298 terminated and 50 truncated episodes.
+- **Worm:** 1,000 episodes, 1,000,000 transitions,
+  mean return 1044.106317, mean length 1000.0000,
+  0 terminated and 1,000 truncated episodes.
+- **Walker:** 1,458 episodes, 1,000,010 transitions,
+  mean return 1363.645402, mean length 685.8779,
+  747 terminated and 711 truncated episodes.
 The recipe ends at the Minari dataset — a portable, env-less trajectory set.
 
 ## Provenance & attribution

release_manifest.json

@@ -37,6 +37,24 @@
       "dataset_release": "release_20",
       "policy_model_release": "release_20",
       "aligned": true
+    },
+    "unity/pyramids/expert-v0": {
+      "environment_release": "release_23",
+      "dataset_release": "release_23",
+      "policy_model_release": "release_23",
+      "aligned": true
+    },
+    "unity/worm/expert-v0": {
+      "environment_release": "release_23",
+      "dataset_release": "release_23",
+      "policy_model_release": "release_23",
+      "aligned": true
+    },
+    "unity/walker/expert-v0": {
+      "environment_release": "release_23",
+      "dataset_release": "release_23",
+      "policy_model_release": "release_23",
+      "aligned": true
     }
   }
 }