Prepare public jump/backflip release with public source links and fresh behavior evidence

README.md

@@ -36,7 +36,7 @@ uses the same collision box as the visible mat.
 
 ## Reproduce in simulation
 
-[Source and instructions](https://github.com/Vottivott/microduck-secret-playground/tree/release/parkour/experiments/parkour/backflip)
+[Source and instructions](https://github.com/Vottivott/microduck-playground/tree/b6aed572587b52e41a452a9b2be517c95d2b3d96/experiments/parkour/backflip)
 
 ```bash
 uv sync --locked
@@ -52,9 +52,12 @@ parity and a small fresh diagnostic of this checkpoint, `manifest.json` for
 provenance and `SHA256SUMS` for integrity. Load the pickle checkpoint only if
 you trust its source.
 
-Related: [long jump](https://huggingface.co/HannesVonEssen/microduck-long-jump)
-· [pillar jumps](https://huggingface.co/HannesVonEssen/microduck-parkour).
+Related: [long jump](https://huggingface.co/HannesVonEssen/microduck-long-jump).
 
 ## Continue training
 
 [TRAINING.md](TRAINING.md) documents the full-checkpoint continuation recipe, distinct from render profiles. A clean locked installation passed a 64-environment/five-update resume, exact learning-state restoration, finite-step checks and normalized ONNX parity. Recipe provenance and reconstruction limits are explicit; this is not a claim of bit-identical historical replay or improved behavior. Original policy, checkpoint and media are unchanged.
+
+## Fresh release check — 2026-09-29
+
+Completes the backward rotation and recovers to standing on the simulated mat, but later drifts off and falls at 7.62 seconds. **This is not a stable idle policy.** One CPU rollout at the documented default seed (28), using the exact packaged checkpoint; not a success-rate estimate or hardware validation. See [source verification](https://github.com/Vottivott/microduck-playground/blob/b6aed572587b52e41a452a9b2be517c95d2b3d96/experiments/parkour/VERIFICATION.md) and `eval/render-20260929.json`.

SHA256SUMS

@@ -1,20 +1,21 @@
 2d3fd9ffffa70a4307498ec630b14a6b4731e010f8ddda92686eaf65733c2dfe  LICENSE
-0e710c240920284a1afae0ef449f475ed165354db46f688b9e7ce2492cf0feac  README.md
-1202ec927bd5828316461e58e1e66dbc3d749e833b0e659a203334ae86e4dfb9  TRAINING.md
+2a64ad6f7d86cb958cd3430f59dec3afed8dbf9fb77f8ede0ab675ea9d8bc77b  README.md
+f75e93c537ac4368d947ca3335f94cbe7cefc07fd04900e4e9b2dfcbf59c2573  TRAINING.md
 f7e523c19de5cd5e2ced430f3d9bb46bef1ed69ad0003cb45955d56169929cdd  checkpoint.pt
-0f5574c1b044ece6df04567e8d469690cb9e83a46ddb96eeae3ef5c75c20e9e3  config.json
+b4ef411ee0c1357763390596beb8cd8c4cd64d0cbe98c11b9d7daf8e0371727b  config.json
 e5b5bfb4e5fe8320584a20d12d729081a8c29980919be9540e8256d0a3d0fee5  eval/backflip.json
 d2478a39f6d03018d830be78c9d72a7322e755932f21e0c307305220ab890396  eval/onnx-parity.json
+8e3b3fc4530883b403c5f6bec48371e04985b6a3c3957a02bcedd99896c38ba3  eval/render-20260929.json
 a0c92f27287602cd7983d16957890584ce8b9d323a10ffc1aad1b1a60283a52a  eval/summary.json
-cf3b8d9a65d079142c218c1c4c597280805e22b0484dfe0346aa66e1dc53fd0b  manifest.json
+c3de8687b897e3dc50944c76d72f30c1597937bb81d1aab0563d9ef9f4cbeca2  manifest.json
 e0b3e1e17f68151dff05dc019999c0053bf6820cad6e44b2e80e305d799a36bd  media/preview.mp4
 841e4eebfbe9946145f88cf01ce3f8f8a501b6b9312b0f037f2b4e0a5f91a48c  media/social-preview.jpg
 b4d780ee2a2bf7aab5b33358124fd079b17a329412d496574296919dbc5b6ee0  media/social-preview.png
 59dd61ea1e8566a3bad9f8b24dd18530664906c33964b1c05e6ed6e821e16a24  policy.onnx
 880cbe85329c8dd1524997eb9049afaec361ad90514ba71e18290f97db22e4d8  training/provenance/agent.yaml
 ba13ebeb26b8437a862970aceec88f5370d94eb493c1bc6594716ebde3141085  training/provenance/env.yaml
-2fc78f00e9199c2b206b6bcd8d1a7f7bbb95012176b55c9bdddec64036c679b8  training/resume.json
-bc414ea910af3f0e0e623af26a1c1db51e28f5f8c2469a1e6c0b42a60de1d145  training/source.json
+309e9211d115a9d6b3f7ed58e85135538e7e7f165570003d52583dbe5e7c3b57  training/resume.json
+cf4a426315f06e2db9908078a8a1b63a50ec029bf42c3835956017e1ea3f7556  training/source.json
 d013ef788d7789af07f3fce4d2344c25c074f0a2b187367881fba6461fd12d38  training/validation/params/agent.yaml
 c3de77796cdac808fce58bdfce75c2a9c8c8d4ff5bac438a1c8a3021e33f7503  training/validation/params/env.yaml
 e51de08bb6c0cfc8b87ce386b1451a94f504d4d574779446ddd277a7cf7a76d8  training/validation/params/events-after-restore.yaml

TRAINING.md

@@ -1,6 +1,6 @@
 # Continue the released parkour policies
 
-All three HF repos contain original full PPO checkpoints: actor, critic,
+Both HF repos contain original full PPO checkpoints: actor, critic,
 observation normalizers, populated Adam state and environment step counter.
 Use the source revision in each repo's `training/source.json`. Install with
 `uv sync --locked` on Linux/CUDA. MuJoCo also needs system EGL or OSMesa libraries;
@@ -10,12 +10,10 @@ Only load trusted checkpoints.
 | Profile | HF repo | Stored iteration |
 | --- | --- | ---: |
 | `long-jump` | `HannesVonEssen/microduck-long-jump` | 15,250 |
-| `pillar-jumps` | `HannesVonEssen/microduck-parkour` | 34,000 |
 | `backflip` | `HannesVonEssen/microduck-backflip` | 12,000 |
 
 ```bash
 hf download HannesVonEssen/microduck-long-jump checkpoint.pt --local-dir policies/long-jump
-hf download HannesVonEssen/microduck-parkour checkpoint.pt --local-dir policies/pillar-jumps
 hf download HannesVonEssen/microduck-backflip checkpoint.pt --local-dir policies/backflip
 
 # From the source root, without inherited MICRODUCK_* environment settings:
@@ -23,10 +21,6 @@ uv run python scripts/resume_release.py \
   --recipe experiments/parkour/resume.json --profile long-jump \
   --checkpoint policies/long-jump/checkpoint.pt \
   --output logs/long-jump-smoke --num-envs 64 --iterations 5
-uv run python scripts/resume_release.py \
-  --recipe experiments/parkour/resume.json --profile pillar-jumps \
-  --checkpoint policies/pillar-jumps/checkpoint.pt \
-  --output logs/pillar-jumps-smoke --num-envs 64 --iterations 5
 uv run python scripts/resume_release.py \
   --recipe experiments/parkour/resume.json --profile backflip \
   --checkpoint policies/backflip/checkpoint.pt \
@@ -65,14 +59,6 @@ performance, robustness or hardware-safety result.
   resumes at the saved counter (late range: gap 0.15–0.40 m, drop 0.15–0.35 m).
   This source includes later research changes, so it is a supported continuation
   rather than an exact reconstruction of p1's original implementation.
-- **Pillar jumps:** the c29 recipe is reconstructed, not a recovered launch.
-  It explicitly enables **informed course observations** (`MICRODUCK_PK_BLIND=0`),
-  five pillars, 12 cm gap, 14 cm drop, and a 75%-of-standing-height hold gate.
-  Seed42 and remaining pinned values are documented continuation choices.
-  The bare task defaults to blind observations and a disabled height gate;
-  using it unconfigured would silently change the problem. Fixed gap/drop
-  disable per-environment difficulty promotion, whose old state was not saved.
-  Nine pillars remain a transfer/evaluation profile, not another checkpoint.
 - **Backflip:** f21 launcher and recorded YAML were recovered: 0.7–0.9 m drop,
   entropy0.003, takeoff weight45, minimum takeoff vz0, curriculum shift3000.
   The initial midflip probability0.25 becomes **0.15** at the restored mature
@@ -91,4 +77,4 @@ command slots and action bounds, not reward totals alone. The original montage's
 exact checkpoint/seed mapping remains unverified; training documentation does
 not change that limitation.
 
-Pinned source: [release/parkour@58d56df0](https://github.com/Vottivott/microduck-secret-playground/tree/58d56df0e863df552fac7caf409aaa968cfde34d).
+Pinned public source: [microduck-playground@b6aed572](https://github.com/Vottivott/microduck-playground/tree/b6aed572587b52e41a452a9b2be517c95d2b3d96).

config.json

@@ -87,9 +87,9 @@
   "format": "onnx",
   "policy_file": "policy.onnx",
   "checkpoint_file": "checkpoint.pt",
-  "source_repository": "Vottivott/microduck-secret-playground",
-  "source_branch": "release/parkour",
-  "source_revision": "26557d0a59537ce9d99f8fe1ca8f7e9bd8ebb6cb",
+  "source_repository": "Vottivott/microduck-playground",
+  "source_branch": "main",
+  "source_revision": "b6aed572587b52e41a452a9b2be517c95d2b3d96",
   "observation_layout": {
     "base_ang_vel": [
       0,

eval/render-20260929.json

@@ -0,0 +1,20 @@
+{
+  "task": "Mjlab-Flip-MicroDuck",
+  "profile": "backflip",
+  "seed": 28,
+  "num_envs": 1,
+  "duration_s": 12.0,
+  "first_episode_end_events": [
+    {
+      "env": 0,
+      "time_s": 7.62,
+      "terms": [
+        "fell"
+      ]
+    }
+  ],
+  "best_link_lower_bound": [
+    0.0
+  ],
+  "note": "Small fixed-profile diagnostic, not a hardware or robustness validation. Terminal-step state may reset before the link diagnostic is read."
+}

manifest.json

@@ -9,10 +9,10 @@
   "policy_sha256": "59dd61ea1e8566a3bad9f8b24dd18530664906c33964b1c05e6ed6e821e16a24",
   "checkpoint_sha256": "f7e523c19de5cd5e2ced430f3d9bb46bef1ed69ad0003cb45955d56169929cdd",
   "source": {
-    "repository": "Vottivott/microduck-secret-playground",
-    "branch": "release/parkour",
-    "revision": "58d56df0e863df552fac7caf409aaa968cfde34d",
-    "public_base": "c594e4457983ed1c1c36193e3df60e16de66b819"
+    "repository": "Vottivott/microduck-playground",
+    "branch": "main",
+    "revision": "b6aed572587b52e41a452a9b2be517c95d2b3d96",
+    "public_base": "4289c236864ea7d8f9a72951667b31c4fcd4b935"
   },
   "profiles": {
     "backflip": {
@@ -43,12 +43,12 @@
   "additional_policy_switching": false,
   "onnx_export_path": "scripts/export.py",
   "simulation_contact_capacity": 200,
-  "source_note": "Scoped import of retained research modules; shared public dependencies retained.",
+  "source_note": "Public two-policy split; original task physics and per-policy continuation settings unchanged from validated preparation. Pillar release excluded.",
   "warning": "Landing may damage or break the real robot. No hardware safety validation.",
   "training_continuation": {
-    "repository": "Vottivott/microduck-secret-playground",
-    "branch": "release/parkour",
-    "revision": "58d56df0e863df552fac7caf409aaa968cfde34d",
+    "repository": "Vottivott/microduck-playground",
+    "branch": "main",
+    "revision": "b6aed572587b52e41a452a9b2be517c95d2b3d96",
     "profile": "backflip",
     "recipe": "training/resume.json",
     "original_release_binaries_unchanged": true,

training/resume.json

@@ -1,38 +1,6 @@
 {
   "schema_version": 1,
   "profiles": {
-    "long-jump": {
-      "task": "Mjlab-PlatformJump-MicroDuck",
-      "checkpoint_sha256": "114ab38724b1bea789b0d3f12a1a18a5981b40572863117e212a2a0601f9ed4a",
-      "seed": 42,
-      "provenance": "p1_v1 stored iteration 15250. Original launch_p1.sh and recorded agent/env YAML recovered. Entropy 0.005; no evaluation gap/drop pin. Uses the published platform-jump source, which contains later research changes, not the original p1 implementation.",
-      "limitations": "Supported continuation from original PPO state under explicit current-source settings, not bit-exact p1 replay. Stage schedule is applied at the restored counter; new episode/RNG state. No external reset bank required.",
-      "environment": {
-        "MICRODUCK_PJ_ENTROPY": "0.005",
-        "MICRODUCK_PJ_CURRICULUM_SHIFT": "0",
-        "MICRODUCK_PJ_FLIGHT_PROB": "0.30"
-      }
-    },
-    "pillar-jumps": {
-      "task": "Mjlab-Parkour-MicroDuck",
-      "checkpoint_sha256": "549ff41b9d892c7244252b7238bb0bd127fd3ec1665e811fb89435f099919828",
-      "seed": 42,
-      "provenance": "c29 stored iteration 34000. Reconstructed continuation from the retained research log: informed actor, five pillars, 12 cm pinned gap and hold requiring 75% standing height. Original c29 launcher/complete resolved run configuration not recovered; seed 42 and remaining pinned settings are explicit continuation choices.",
-      "limitations": "Not exact historical replay. Per-environment promotion state was not checkpointed; fixed gap/drop deliberately disable promotion. Nine pillars are evaluation/transfer, not a separate checkpoint or the default continuation curriculum. No external reset bank required.",
-      "environment": {
-        "MICRODUCK_PK_BLIND": "0",
-        "MICRODUCK_PK_PILLARS": "5",
-        "MICRODUCK_PK_BASE_TOP": "1.0",
-        "MICRODUCK_PK_DEPTH": "0.15",
-        "MICRODUCK_PK_GAP": "0.12",
-        "MICRODUCK_PK_DROP": "0.14",
-        "MICRODUCK_PK_EPISODE_S": "11",
-        "MICRODUCK_PK_HOLD_STAND_FRAC": "0.75",
-        "MICRODUCK_PK_MAX_SPAWN_LINK": "3",
-        "MICRODUCK_PK_ENTROPY": "0.004",
-        "MICRODUCK_PK_CURRICULUM_SHIFT": "0"
-      }
-    },
     "backflip": {
       "task": "Mjlab-Flip-MicroDuck",
       "checkpoint_sha256": "f7e523c19de5cd5e2ced430f3d9bb46bef1ed69ad0003cb45955d56169929cdd",

training/source.json

@@ -1,7 +1,7 @@
 {
-  "repository": "Vottivott/microduck-secret-playground",
-  "branch": "release/parkour",
-  "revision": "58d56df0e863df552fac7caf409aaa968cfde34d",
+  "repository": "Vottivott/microduck-playground",
+  "branch": "main",
+  "revision": "b6aed572587b52e41a452a9b2be517c95d2b3d96",
   "profile": "backflip",
   "recipe": "training/resume.json",
   "original_release_binaries_unchanged": true