Clarify Bagel-Robotics-1 pruning and data-selection contributions

README.md

@@ -9,7 +9,14 @@ tags:
 
 # Bagel-Robotics-1 — checkpoint 1000
 
-Inference release of the **exact checkpoint 1000** used in the Bagel-Robotics-1 evaluation campaign. **The repository is currently private.** Evaluator access is being arranged using a dedicated fine-grained token with read access only to this repository, delivered through an agreed private channel. Credential provisioning and evaluator assignment remain pending. Public release is prepared for the later verified-leaderboard milestone and requires a bageldotcom administrator. This is distinct from Bagel-Robotics-1-Preview.
+Bagel-Robotics-1 is Bagel Labs' approach to inference-friendly compact world-action
+models (WAMs). Our method combines mask-based structured pruning and recovery
+training with effective selection of critically important segments from each task,
+using a focused subset of the training dataset. VPP2 is the base model used to
+validate this approach; reused implementation components are identified in the
+source notices below.
+
+Inference release of the **exact checkpoint 1000** used in the Bagel-Robotics-1 evaluation campaign. **The repository is currently private.** Evaluator access is being arranged using a dedicated fine-grained token with read access only to this repository, delivered through an agreed private channel. The dedicated repository-only read credential is verified; assigned evaluator handoff remains pending. Public release is prepared for the later verified-leaderboard milestone and requires a bageldotcom administrator.
 
 The video expert has **12,468,409,664 parameters**, 23.95% fewer than its recorded parent. Learned masks were converted into smaller FFN and cross-attention layers. Self-attention heads and the action expert are retained. No teacher or learned-mask module is required for inference. Parameter reduction is not a measured inference-speedup claim.
 
@@ -54,7 +61,7 @@ Follow the included [policy README](inference/policy/bagel_robotics_1/README.md)
 
 The adapter uses `policy_name=bagel_robotics_1`, `ckpt_name=cp1000`, ARX X5 end-effector actions, and one environment per policy process. Fixed inference settings: BF16; predict 32 actions and execute 16; ten Euler steps; shift 1; 90% new / 10% previous temporal blending. Qualification used a 96 GB RTX PRO 6000 Blackwell, Python 3.10.21 and PyTorch 2.11.0+cu130; a lower minimum GPU capacity has not been qualified.
 
-These are compact-aware VPP2 artifacts, not Transformers `AutoModel` checkpoints. The XPolicyLab PR has not been created; publishing this inference release does not imply committee acceptance.
+Use the Bagel-Robotics-1 compact checkpoint loader supplied in the inference package. The implementation and current reviewer instructions are available in [XPolicyLab PR #234](https://github.com/XPolicyLab/XPolicyLab/pull/234).
 
 ## Validation and release scope
 
@@ -64,4 +71,4 @@ The technical report and training/pruning/export code are planned for **November
 
 ## Attribution and terms
 
-Derived from [VPP2](https://github.com/roboterax/video-prediction-policy-2) and Wan2.1 components. See [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) and the adapter's retained MIT/Apache-2.0 notices. Model and auxiliary-asset upstream terms continue to apply; an inference-code license is not a new blanket license for all weights or datasets.
+VPP2 is the base model used to validate the Bagel-Robotics-1 pruning, recovery-training and data-selection approach. Reused base-model inference components come from [VPP2](https://github.com/roboterax/video-prediction-policy-2)/FastWAM and Wan2.1/DiffSynth. See [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) and the adapter's retained MIT/Apache-2.0 notices. Model and auxiliary-asset upstream terms continue to apply; an inference-code license is not a new blanket license for all weights or datasets.

THIRD_PARTY_NOTICES.md

@@ -1,12 +1,17 @@
 # Source and artifact notices
 
-- The retained VPP2/FastWAM-derived runtime carries the original MIT notice in
-  `LICENSE`, copyright 2026 The FastWAM Authors. Bagel-specific packaging does not
-  remove that attribution. See `source-provenance.json` for frozen-source hashes.
+Bagel-Robotics-1's method contribution is the pruning/recovery training procedure
+and effective selection of task-critical data for an inference-friendly compact
+WAM. VPP2 is the base model used to validate that approach. The source notices
+below identify reused implementation components and their licenses.
+
+- Base-model inference components adapted from [VPP2](https://github.com/roboterax/video-prediction-policy-2)/FastWAM retain the MIT notice in
+  [LICENSE](inference/policy/bagel_robotics_1/LICENSE), copyright 2026 The FastWAM Authors. See [source-provenance.json](inference/policy/bagel_robotics_1/source-provenance.json)
+  for the source identities and packaging changes.
 - Wan2.1/DiffSynth-derived model components retain Apache-2.0 terms; see
-  `licenses/Apache-2.0.txt` and upstream Wan2.1/DiffSynth source notices.
+  [licenses/Apache-2.0.txt](inference/policy/bagel_robotics_1/licenses/Apache-2.0.txt) and upstream Wan2.1/DiffSynth source notices.
 - The XPolicyLab interface follows its upstream license, included in
-  `licenses/XPolicyLab.txt`. RoboDojo and simulator assets are not bundled.
+  [licenses/XPolicyLab.txt](inference/policy/bagel_robotics_1/licenses/XPolicyLab.txt). RoboDojo and simulator assets are not bundled.
 - Encoder weights, tokenizer files and model weights have their own upstream
   terms. The code license does not assign a new license to those artifacts.
 - The existing preview PR by Soroush Mehraban was consulted as a reference; its

inference/README.md

@@ -1,5 +1,10 @@
 # Bagel-Robotics-1 inference adapter
 
+Bagel-Robotics-1 combines mask-based structured pruning and recovery training with
+effective selection of task-critical data to produce an inference-friendly compact
+WAM. VPP2 is the base model used to validate this approach; reused implementation
+components are identified in the policy's source notices.
+
 This folder is the standalone XPolicyLab adapter for Bagel-Robotics-1 checkpoint 1000. It contains inference code and evaluation instructions, with no training repository dependency. Model tensors download separately from `bageldotcom/Bagel-Robotics-1`, pinned to `2651631c5d08819088475e8778676f5cf6d9afcf` and verified against `policy/bagel_robotics_1/artifact-manifest.json`.
 
 Install this adapter into an existing XPolicyLab checkout (the installer refuses to overwrite an existing policy):
@@ -10,6 +15,6 @@ python install_adapter.py /absolute/path/to/XPolicyLab
 
 Then follow [Installation](policy/bagel_robotics_1/README.md#installation) for the Python 3.10 environment, exact checkpoint and encoder download, and debug checks. [REEVALUATION.md](policy/bagel_robotics_1/REEVALUATION.md) gives native simulator task/suite/three-seed commands. RoboDojo and its assets are installed separately.
 
-The adapter is prepared against XPolicyLab `6baccd57ff958570d0db1588370289ee36c8bb70`; source and dependency pins are preserved in the policy folder. This inference package is not an XPolicyLab PR or an official leaderboard result. Training/pruning/export source and the technical report are planned for November 20, 2026.
+The adapter is prepared against XPolicyLab `6baccd57ff958570d0db1588370289ee36c8bb70`; source and dependency pins are preserved in the policy folder. The adapter is under review in [XPolicyLab PR #234](https://github.com/XPolicyLab/XPolicyLab/pull/234). Training/pruning/export source and the technical report are planned for November 20, 2026.
 
 The repository is currently private. Evaluators need separately arranged repository-only read access; public release is a later leaderboard milestone.

inference/policy/bagel_robotics_1/README.md

@@ -1,12 +1,14 @@
 # Bagel-Robotics-1
 
-**Contributor:** Bagel Labs | **Paper:** Technical report in preparation | **arXiv:** Not yet available | **Original code:** [VPP2](https://github.com/roboterax/video-prediction-policy-2)
+**Contributor:** Bagel Labs | **Paper:** Technical report in preparation | **arXiv:** Not yet available | **Original code:** [Bagel-Robotics-1 implementation](.) · [Implementation source notices](THIRD_PARTY_NOTICES.md)
 
-Evaluation-only adapter for the physically compact VPP2 recovery checkpoint 1000.
-The video expert has 12,468,409,664 parameters, 23.95% fewer than the recorded
-16,394,878,784-parameter parent video expert. Action2B remains unchanged. The
-`runtime/` directory vendors a minimal isolated inference runtime and compact
-loader; this is a different artifact from Bagel-Robotics-1-Preview.
+Bagel-Robotics-1 is Bagel Labs' approach to inference-friendly compact world-action
+models (WAMs), combining mask-based structured pruning and recovery training with
+effective selection of critically important task segments. We use VPP2 as the
+base model to validate this training and data-selection approach; bundled base-model
+components in `runtime/` are identified in the [source notices](THIRD_PARTY_NOTICES.md).
+The exported video expert has 12,468,409,664 parameters, 23.95% fewer than the
+recorded 16,394,878,784-parameter base video expert, while Action2B is retained.
 
 Shared conventions — argument meanings, checkpoint naming, split-machine deployment, `EVAL_ENV_TYPE` — are documented in the [XPolicyLab README](https://github.com/XPolicyLab/XPolicyLab/blob/main/README.md). Official results: [RoboDojo LeaderBoard](https://robodojo-benchmark.com/LeaderBoard).
 
@@ -17,6 +19,19 @@ workspace. `env_cfg/` belongs next to XPolicyLab, not inside the policy. Install
 RoboDojo and its simulator/assets separately using the
 [official installation guide](https://robodojo-benchmark.com/doc/usage/install-and-download/).
 
+Until this adapter is merged, fetch [PR #234](https://github.com/XPolicyLab/XPolicyLab/pull/234)
+into a dedicated reviewer checkout of XPolicyLab in that workspace. Replace the
+example path, and keep the printed commit SHA with the evaluation records:
+
+```bash
+git -C /absolute/path/to/XPolicyLab fetch https://github.com/XPolicyLab/XPolicyLab.git refs/pull/234/head
+git -C /absolute/path/to/XPolicyLab checkout --detach FETCH_HEAD
+git -C /absolute/path/to/XPolicyLab rev-parse HEAD
+```
+
+For an evaluation application that specifies a commit, check out that exact SHA
+after fetching. The upstream base commit alone does not contain this adapter.
+
 **1. Set paths and install the policy environment.** Replace the example checkout
 path below. Run the following sections in the same shell so these variables remain
 available. Use **Python 3.10**, the version used for the GPU qualification; set
@@ -42,7 +57,8 @@ host RAM and eight CPUs; these are tested allocations, not measured minima.
 prepared at [bageldotcom/Bagel-Robotics-1](https://huggingface.co/bageldotcom/Bagel-Robotics-1).
 The repository is currently private. Evaluators need a dedicated fine-grained
 token granting read access only to this repository, supplied separately through
-an agreed private channel. Credential creation, validation and handoff are pending.
+an agreed private channel. The dedicated credential and repository-only read scope are verified; assigned
+evaluator handoff remains pending.
 Set `HF_TOKEN` securely before downloading (for example, `read -r -s HF_TOKEN`
 followed by `export HF_TOKEN`; use `set +x` first and `unset HF_TOKEN` afterward).
 Do not place the token in scripts, the PR, logs, or the application text.
@@ -62,15 +78,20 @@ files, follow [Model Assets](#model-assets) instead. Continue to
 
 ## Data Processing
 
-No conversion is needed for inference. The research recovery run used a curated
-30,000-window RoboDojo EE16 export; training reproduction is outside this initial
-adapter. `process_data.sh` is intentionally omitted under the proposed eval-only
-exception. Inputs are decoded RGB arrays from XPolicyLab's shared server.
+No conversion is needed for inference. A key component of our approach is to use
+a dataset subset containing critically important segments from each task instead
+of the complete training dataset. The research recovery run used a curated,
+task-balanced 30,000-window RoboDojo EE16 subset.
+
+Training reproduction is outside this initial adapter. `process_data.sh` is
+intentionally omitted under the proposed eval-only exception. Inputs are decoded
+RGB arrays from XPolicyLab's shared server. The segment-selection procedure and
+recovery recipe will be described in the technical report.
 
 ## Training
 
 `train.sh` is intentionally omitted. Training, mask learning and export are not included in this adapter; the
-standalone review package contains a method note. The technical report and training,
+technical report and training,
 pruning, and export code are planned for release on **November 20, 2026**.
 The proposed eval-only integration and release schedule remain subject to maintainer
 coordination; this preparation does not claim that the exception is approved.
@@ -197,11 +218,20 @@ See [VALIDATION.md](VALIDATION.md) for capture/replay scope.
 
 ## Notes
 
-Mask-based structured pruning and recovery produce physically smaller,
-nonuniform FFN and cross-attention layers. The method is intended for compatible
-WAM backbones; the evidence here is limited to VPP2. The technical report will
-provide the detailed algorithm and recovery recipe.
+**Method contribution:** our training procedure combines structured pruning,
+recovery of the compact model, and effective selection of task-critical training
+data to produce an inference-friendly compact WAM. VPP2 serves as the validation
+base model for this approach.
+
+The approach combines two key components:
+
+- **Mask-based structured pruning and recovery:** produce physically smaller,
+  nonuniform FFN and cross-attention layers.
+- **Task-focused data selection:** use a subset containing critically important
+  segments from each task for recovery, instead of the complete training dataset.
 
+The approach is intended for compatible WAM backbones. The technical report will
+describe the pruning, segment-selection and recovery-training methodology in detail.
 
 The recorded research results cover all five suites:
 
@@ -217,9 +247,8 @@ The recorded research results cover all five suites:
 Memory includes one repaired variant, Long Horizon two labeled replacements, and Open one labeled replacement. Generalization combines standard and randomized episodes. Memory and nine Generalization tasks use user-supplied summaries; see [EVALUATION_RESULTS.md](EVALUATION_RESULTS.md) for complete counts, task results, source limitations and original-only figures. These are research-runner results, not official leaderboard scores or proof of packaged-adapter equivalence.
 
 The learned masks have been exported into nonuniform per-layer FFN/cross-attention
-reductions. Action2B and all video self-attention heads are retained. Eligibility
-as a new model and the evaluation-only exception remain subject to committee review.
+reductions. Action2B and all video self-attention heads are retained.
 
 Package validation so far: 15 CPU tests, real compact-video strict tensor loading, all checkpoint/encoder hashes, static checks, isolated wheel processor loading, and clean-checkout installation. Both real-checkpoint GPU debug transports passed in job 53609. Four native simulator episodes completed with valid results, and all eight action/history/reset/backend comparison checks passed with zero recorded differences. See [VALIDATION.md](VALIDATION.md) for the measured task scores and test scope.
-Reviewer access remains pending. The technical report and training/pruning/export code are planned for **November 20, 2026**. Full-suite
-coverage is a separate leaderboard-verification milestone. See [validation evidence](VALIDATION.md) and [historical task results](EVALUATION_RESULTS.md). The standalone review package contains the fuller internal method note.
+The repository-only read credential is verified; evaluator assignment and private handoff remain pending. The technical report and training/pruning/export code are planned for **November 20, 2026**. Full-suite
+coverage is a separate leaderboard-verification milestone. See [validation evidence](VALIDATION.md) and [historical task results](EVALUATION_RESULTS.md).

inference/policy/bagel_robotics_1/REEVALUATION.md

@@ -10,7 +10,27 @@ The only policy checkpoint is `cp1000`, from
 revision `2651631c5d08819088475e8778676f5cf6d9afcf`. The downloader and loader reject
 files that do not match `artifact-manifest.json`. The VAE, image/text encoders and
 tokenizer are downloaded automatically at their separately pinned public revision.
-The checkpoint is prepared for public access, pending an organization administrator visibility change. Once public, it and the required encoder assets download without authentication.
+The checkpoint is currently private. Obtain the dedicated repository-only read
+credential through the agreed private channel and set `HF_TOKEN` securely as
+shown in the README before running the downloader. Public release is pending an
+organization administrator visibility change; anonymous checkpoint access will
+work only after that change. The required encoder assets are public.
+
+## What a reviewer can reproduce
+
+The adapter loads the pinned checkpoint and implements the recorded inference
+settings. Its completed qualification consists of both real-checkpoint debug
+transports, four native simulator episodes and eight exact comparison checks on
+two recorded observation sequences. The full five-suite research score table has
+not been rerun end to end through this packaged adapter.
+
+The commands below start a fresh official evaluation using supplied original
+layouts. They do not recreate the research run's three replacement scenes or one
+repaired Memory variant. Compare episode selection and seed coverage before
+comparing score totals; retain startup failures for committee review. The public
+comparison receipts describe the tested scope, but the original research runner
+and raw comparison fixtures are not included in this PR. They are not required to
+run a fresh evaluation with the packaged model.
 
 ## Set the local paths
 
@@ -120,7 +140,7 @@ by these commands. Official verification and acceptance remain with RoboDojo.
 
 | Symptom | Check |
 |---|---|
-| HTTP 401/403 or repository not found | Use the canonical public repository and pinned revision; retry without a stale cached credential (`HF_HUB_DISABLE_IMPLICIT_TOKEN=1`). |
+| HTTP 401/403 or repository not found | The checkpoint repository is currently private. Set the supplied repository-only read token as `HF_TOKEN` and use the canonical repository and pinned revision. Removing authentication will not fix private access. After public release, anonymous access can be checked with the token unset. |
 | Hash/size mismatch | Check the pinned revision and local file completeness; do not disable verification. |
 | Missing encoder or tokenizer | Run the full downloader; `--metadata-only` does not fetch tensors or encoders. |
 | Missing `env_cfg` or simulator | Complete the official workspace setup; `env_cfg/` belongs beside XPolicyLab. |

inference/policy/bagel_robotics_1/THIRD_PARTY_NOTICES.md

@@ -1,8 +1,13 @@
 # Source and artifact notices
 
-- The retained VPP2/FastWAM-derived runtime carries the original MIT notice in
-  `LICENSE`, copyright 2026 The FastWAM Authors. Bagel-specific packaging does not
-  remove that attribution. See `source-provenance.json` for frozen-source hashes.
+Bagel-Robotics-1's method contribution is the pruning/recovery training procedure
+and effective selection of task-critical data for an inference-friendly compact
+WAM. VPP2 is the base model used to validate that approach. The source notices
+below identify reused implementation components and their licenses.
+
+- Base-model inference components adapted from [VPP2](https://github.com/roboterax/video-prediction-policy-2)/FastWAM retain the MIT notice in
+  `LICENSE`, copyright 2026 The FastWAM Authors. See `source-provenance.json`
+  for the source identities and packaging changes.
 - Wan2.1/DiffSynth-derived model components retain Apache-2.0 terms; see
   `licenses/Apache-2.0.txt` and upstream Wan2.1/DiffSynth source notices.
 - The XPolicyLab interface follows its upstream license, included in

inference/policy/bagel_robotics_1/VALIDATION.md

@@ -1,6 +1,6 @@
 # Adapter validation
 
-Status: **Technical qualification passed: 15 CPU tests, both real-checkpoint debug transports, four valid native simulator episodes, and all eight exact replay/backend checks. The technical report and training/pruning/export code are planned for November 20, 2026; evaluator access and user review remain before submission.**
+Status: **Technical qualification passed: 15 CPU tests, both real-checkpoint debug transports, four valid native simulator episodes, and all eight exact replay/backend checks. [PR #234](https://github.com/XPolicyLab/XPolicyLab/pull/234) is open. The dedicated private read credential is verified; delivery to an assigned evaluator remains a separate handoff. The technical report and training/pruning/export code are planned for November 20, 2026. Full official benchmark verification remains pending.**
 
 | Check | Evidence and scope |
 |---|---|
@@ -22,6 +22,26 @@ Two real recordings of 209 observations / 14 action chunks each were replayed in
 
 The original Store laptop/headphones layout 1 failed with 0.2/1 partial and is retained. Debug completion is an interface check; four simulator smoke episodes do not establish full-suite performance. Historical research scores are separately documented in EVALUATION_RESULTS.md.
 
+## Reviewer workflow audit (November 7, 2026)
+
+A fresh regular-file snapshot of submitted commit
+`a08a40fb01664cfe5411a02d723f9e0af7233683` passed all 15 CPU contract tests with
+zero skips after the published metadata downloader fetched and verified the
+four pinned JSON files. This check used only the dedicated reviewer credential
+and an empty Hugging Face cache; no owner credential fallback was used.
+
+Authenticated metadata access also passed for both checkpoint tensors and all
+nine public dependency files, matching pinned sizes and the appropriate remote
+LFS SHA-256 or Git blob identity. Large tensors were not downloaded again in this
+audit. CPU checks used Python 3.12.7 / PyTorch 2.11.0+cpu with existing test
+dependencies; they do not replace the earlier fresh Python 3.10.21 / CUDA 13.0
+installation and GPU qualification. Inference code and settings are unchanged.
+
+The reviewer guide now explicitly fetches the PR source and explains private
+authentication. A complete fresh-machine simulator/assets installation, different
+GPU types, and a full five-suite packaged-adapter run have not been qualified by
+this audit.
+
 ## Reproduce the transport checks
 
 Use the policy README's debug command with the actual checkpoint. Repeat it with `DEBUG_OBS_ENCODED=1`. Both runs must finish normally and print the evaluation-finished marker.

inference/policy/bagel_robotics_1/runtime/THIRD_PARTY_NOTICES.md

@@ -1,6 +1,11 @@
 # Runtime source notices
 
-The VPP2/FastWAM-derived source retains the MIT notice in LICENSE.
+This runtime contains base-model inference components adapted from
+[VPP2](https://github.com/roboterax/video-prediction-policy-2)/FastWAM for validation
+of the Bagel-Robotics-1 approach. These reused source components retain the MIT
+notice in LICENSE. The method contribution combines structured pruning and
+recovery training with effective selection of task-critical data, as described
+in the policy README.
 Wan2.1/DiffSynth-derived components retain Apache-2.0 terms, included in
 licenses/Apache-2.0.txt. Source attribution is retained in individual modules
 where present. Source hashes and packaging changes accompany the adapter in