GWAM_Data — Active-13 current-graph / GWAM corrective successor v1.1.1

Production-recommended code prefix: gwam_active13_current_graph_gwam_v1_1_1_20260818/
Immutable corrective prefix commit: 031f8b4994488c057a74cf369963c5357e26f49a
Authority closure: gwam_active13_canonical_authority_closure_v2_0_2_20260817/
Frozen parent Qwen release: gwam_active13_graph_adapter_qwen_v1_0_1_20260816/
Superseded but immutable predecessor: gwam_active13_current_graph_gwam_v1_1_0_20260817/

v1.1.1 is an additive authority/security correction. It does not modify or overwrite v1.1.0. Use v1.1.1 for production current-graph code; use the v1.1.0 prefix only for historical readback and its retained visual/runtime artifacts.

Corrective authority boundary

The primary current-frame contract remains:

simulator-authoritative frame t
→ authenticated variable-size G_t=(V_t,E_t)
→ frozen semantic + optional pinned Qwen evidence
→ five typed edge-aware relational message-passing blocks
→ CURRENT_GRAPH_T [B,K,384]

History remains optional and model-side. CURRENT_GRAPH_T has no mandatory L axis; HISTORY_GRAPH_T is a temporal-layer output; legacy POSTERIOR_T remains the frozen parent L=8 GRU posterior.

The corrective release closes these production trust seams:

• canonical hashing is a private capability compiled from dependency-verified source bytes, with captured code/bindings and a separate local-law cross-check;
• semantic registry type/origin/content and the exact frozen semantic weights consumed by the encoder are bound together;
• InputPolicy is exact owner authority, not a duck type or subclass;
• production Qwen evidence accepts only the exact issued AuthenticatedLazyEpisodeQwenSidecar; the parent DiskVisualSidecar remains a parent-only POSTERIOR_T accessor;
• frame authority allows arbitrary distinct same-episode frames in caller order; temporal ordering is enforced only by temporal integration;
• tensor/device/padding/mask/history boundaries fail closed.

The v1.1.1 policy authority changed all four input_policy_sha256 receipts. v1.1.0 batch/model receipts are intentionally incompatible and must not be relabeled as v1.1.1.

Audit gate chain

The complete gate chain is explicit:

1. delayed security/spec review: REQUEST_CHANGES;
2. Fable/Codex reconciliation: READY_TO_IMPLEMENT;
3. V2 re-audit: REQUEST_CHANGES after a visual-sidecar seam was found;
4. V3 source re-audit: APPROVE_FOR_STAGING, P0=0, P1=0, P2=0 source findings;
5. staged-prefix readback: APPROVE_FOR_HF_PUBLICATION.

The pinned verification/CORRECTIVE_RELEASE_BOUNDARY_20260818.json inside the v1.1.1 prefix is the 02:14 reopening declaration (status: CORRECTIVE_IMPLEMENTATION_IN_PROGRESS, required_fable_verdict: FINAL_FABLE_CORRECTIVE_VERDICT: PASS); it is superseded for publication authority by FINAL_RELEASE_AUTHORITY.json, the complete gate chain above, and docs/active13/CORRECTIVE_RELEASE_BOUNDARY_V1_1_1_20260818.json.

Authoritative files live under the v1.1.1 prefix:

• FULL_RELEASE_MANIFEST.jsonl;
• FINAL_RELEASE_AUTHORITY.json;
• verification/CORRECTIVE_SOURCE_REVIEW_MANIFEST_V3_20260818.jsonl;
• verification/FABLE_FINAL_CORRECTIVE_V3_REAUDIT_20260818.md;
• verification/FABLE_FINAL_CORRECTIVE_REAUDIT_20260818.md;
• verification/FABLE_CODEX_CORRECTIVE_RECONCILIATION_20260818.md;
• verification/CORRECTIVE_SOURCE_REVIEW_MANIFEST_AUTHORITY_V3_20260818.json.

The staged-prefix readback report FABLE_V1_1_1_STAGED_READBACK_20260818.md remains in the release workspace outside the v1.1.1 prefix.

Prefix authority:

source manifest: 372d974e84193411edc03909e3926878bba47cf85312deacbdaa67e6e78140a5
release manifest: c0763d256c7a14d253bb38ca4fafe6a119c51db00d55d54b0f9df0d83df4f702
release authority: 504cb6f8548a2cf6218b26659bffcfd9d1441d908adcbc37da0c4023bf228837
source tests: 286 passed, 1 explicit live-RoboCasa opt-in skip
heldout_opened: 0
optimizer_steps: 0
checkpoints_written: 0
trained: false

Download and verify v1.1.1

from huggingface_hub import snapshot_download

repo = "ChangChrisLiu/GWAM_Data"
prefix = "gwam_active13_current_graph_gwam_v1_1_1_20260818"
revision = "031f8b4994488c057a74cf369963c5357e26f49a"
root = snapshot_download(
    repo_id=repo,
    repo_type="dataset",
    revision=revision,
    local_dir="GWAM_Data",
    allow_patterns=[f"{prefix}/**"],
)
print(f"{root}/{prefix}")

Erratum E1 — downloaded verifier mode

Hugging Face transport normalizes downloaded file modes and does not preserve the executable bit. The staged 0555 mode recorded in FINAL_RELEASE_AUTHORITY.json is local staging hygiene, not a transported byte or content property. The chmod 0555 step below applies only to the local downloaded verifier. Prefix bytes are unchanged; immutable readback re-verified release manifest c0763d256c7a14d253bb38ca4fafe6a119c51db00d55d54b0f9df0d83df4f702 and release authority 504cb6f8548a2cf6218b26659bffcfd9d1441d908adcbc37da0c4023bf228837.

Verify the downloaded prefix by first authenticating the downloaded verifier bytes, then setting its local staging-compatible mode, then invoking it with isolated Python:

GWAM_V111="$PWD/GWAM_Data/gwam_active13_current_graph_gwam_v1_1_1_20260818"
GWAM_V111_VERIFIER="$GWAM_V111/release_tools/assemble_successor_release_v1_1_1.py"
printf '%s  %s\n' \
  '588532de5acc432fa41d77aace787f0ba5ce19d52de97ba580a954715bdf0921' \
  "$GWAM_V111_VERIFIER" \
  | sha256sum --check --strict - \
  && chmod 0555 "$GWAM_V111_VERIFIER" \
  && python -I -B "$GWAM_V111_VERIFIER" --verify "$GWAM_V111" \
  && python -m pip install -e "$GWAM_V111"

The shipped tests preserve the audited workspace layout. For immutable readback, run them from a workspace-layout checkout or use the composed-root/examples flow in HF_DIRECT_USE.md; running the flat prefix's test files in place is not the supported path.

Historical v1.1.0 assets

v1.1.1 is intentionally minimal and does not reissue visual guide/deck or standalone runtime bundles as if they had new authority. Historical artifacts remain at:

• gwam_active13_current_graph_gwam_v1_1_0_20260817/artifacts/;
• gwam_active13_current_graph_gwam_v1_1_0_20260817/runtime/.

They are reference material, not v1.1.1 production authority. The prior root README is archived byte-for-byte at docs/archive/README_ACTIVE13_CURRENT_GRAPH_V1_1_0_ROOT_20260818.md. The pre-Erratum-E1 v1.1.1 root README is archived byte-for-byte at docs/archive/README_ACTIVE13_CURRENT_GRAPH_V1_1_1_ROOT_20260818.md.

Scientific boundaries

No optimizer step, checkpoint, training run, held-out evaluation, EGNN claim, graph-benefit claim, policy-performance claim, or K=16-optimality claim is introduced. Actions remain separate from graph tokens. The zero-gated joint-GWAM seam remains removable and initialization-identical to baseline.

Publication details: docs/active13/PUBLICATION_ADDENDUM_V1_1_1_20260818.json and docs/active13/CORRECTIVE_RELEASE_BOUNDARY_V1_1_1_20260818.json.

---

Parent release v1.0.1 complete guide

GWAM_Data — Active-13 canonical graph adapter + Qwen visual sidecars v1.0.1

Parent release prefix: gwam_active13_graph_adapter_qwen_v1_0_1_20260816/

Unless a path is explicitly repository-level, paths in the parent guide below are relative to the parent release prefix.

This is the complete train-only Active-13 graph/Qwen release for 13 RoboCasa task families. Start here if you want to:

1. inspect or train with the precomputed Qwen sidecars;
2. join Qwen evidence to simulator-authoritative canonical graphs;
3. run the typed graph adapter and obtain [B,16,384] graph tokens; or
4. extract a current physical graph plus synchronized Qwen evidence during a live, real-time RoboCasa rollout.

The release contains the whole authorized train-derived Qwen dataset as 3,943 episode packages, the exact Qwen model used to create them, the frozen semantic registry, canonical graph and adapter code, live runtime code, examples, schemas, and verification authorities.

Important: Qwen is visual evidence, not graph structure. Simulator state defines entities and relations. Qwen cannot add, delete, rename, or override a physical node or edge.

Start here: choose your workflow

Goal	Entry point	Output	Trust domain
Inspect one packaged episode	examples/inspect_episode.py	counts and validated array shapes	authenticated offline train sidecar
Load one frame's Qwen rows	examples/load_frame_qwen_rows.py	(slot, camera, bbox, embedding) rows	authenticated offline train sidecar
Load semantic vectors	examples/load_semantic_registry.py	family/category/relation/task matrices	frozen offline registry
Embed a new RGB crop	examples/load_bundled_qwen.py	unit-normalized [1,256] vector	authenticated model, new visual evidence
Produce model graph tokens	code/release_scripts/run_production_adapter_smoke_v1_0_1.py	[B,16,384], POSTERIOR_T	authenticated offline production path
Extract during execution	runtime/examples/realtime_qwen_graph.py	variable-size physical graph + camera rows	live current observation, not offline receipt

The offline and live paths are intentionally separate. Never relabel a live output as an authenticated offline training packet.

Release facts

model split:                train only
tasks:                      13
episodes:                   3,943
frames:                     985,959
legal L=8 windows:          926,814
Qwen camera/node rows:      42,661,225
unique RGB crops:           36,181,363
max canonical nodes:        254
max canonical edges:        206
held-out payloads opened:   0
adapter trained:            no
held-out evaluated:         no

Task roster

Task	Episodes	Frames	Qwen rows
CoffeeSetupMug	392	92,659	4,539,678
OpenDrawer	409	108,479	5,585,429
PickPlaceCabinetToCounter	85	16,301	632,672
PickPlaceCounterToCabinet	488	123,523	5,458,861
PickPlaceCounterToMicrowave	88	33,594	1,484,164
PickPlaceCounterToSink	86	17,820	588,569
PickPlaceCounterToStove	487	122,856	5,423,643
PickPlaceMicrowaveToCounter	90	31,351	1,390,903
PickPlaceSinkToCounter	487	177,440	6,612,027
PickPlaceStoveToCounter	87	18,556	772,490
TurnOffStove	406	88,802	3,946,147
TurnOnMicrowave	423	63,088	2,761,397
TurnOnSinkFaucet	415	91,490	3,465,245

What the graph means

A graph is G=(V,E):

• V is the current set of robot, object, fixture, articulated-part, handle, control, surface, and receptacle nodes;
• E contains typed structural or physical relations between nodes;
• node state carries pose, articulation, target state, visibility, and semantic identifiers;
• camera-dependent Qwen vectors are attached as optional per-node/per-camera evidence;
• actions remain a separate model input and are not inserted into graph tokens.

Canonical offline graphs are variable-size. Fixed Nmax=254 and Emax=206 are corpus maxima, not a requirement to treat padding as real nodes. The model performs message passing over all valid nodes and edges before compressing the graph to 16 tokens.

Authority and data flow

RoboCasa simulator state
  └─ canonical graph v1.0.3
       ├─ variable-size nodes and typed edges        structural authority
       ├─ stable slot_id                             join address
       └─ packet_sha256                              frame identity

RGB from right / left / wrist cameras
  └─ visible canonical slot + normalized bbox
       └─ exact crop bytes
            └─ Qwen3-VL-Embedding-2B
                 └─ unit-normalized MRL-256 vector   visual sidecar only

canonical L=8 window
+ frozen semantic registry
+ authenticated Qwen sidecar rows
  └─ full-N typed graph adapter
       └─ five message blocks
            └─ K=16 codec
                 └─ graph_tokens [B,16,384], token_type=POSTERIOR_T

Authority order:

1. canonical simulator graph v1.0.3 defines nodes and relations;
2. canonical source ZIP SHA-256 identifies an episode;
3. packet SHA-256 identifies one canonical frame packet;
4. RGB authority and crop SHA-256 identify visual bytes;
5. sidecar receipt binds episode, packet, slot, camera, bbox, crop, and embedding;
6. semantic receipt binds all text-derived semantic matrices;
7. production authorization binds the complete batch before learned arithmetic.

Directory map

README.md                              this end-to-end guide
SCHEMA.md                              normative file/tensor schema
MANIFEST.jsonl                         one authority row per episode ZIP
PACKAGE_SUMMARY.json                   corpus totals
FULL_RELEASE_MANIFEST.jsonl            full release file ledger
FINAL_RELEASE_AUTHORITY.json           frozen package authority
episodes/                              3,943 sidecar ZIPs
semantic_registry/                     precomputed MRL-256 semantic matrices
models/Qwen3-VL-Embedding-2B/           exact 18-file upstream model closure
models/QWEN3_VL_EMBEDDING_HF_METADATA/ authenticated publishable HF metadata
models/QWEN3_VL_EMBEDDING_MODEL_MANIFEST.json
models/QWEN3_VL_EMBEDDING_APACHE_2_0.txt
code/canonical_graph_v1_0_3/            exact canonical graph code/schema
code/adapter_package/                  typed adapter and strict loaders
code/release_scripts/                  validation and production smoke tools
examples/                              portable offline/model examples
runtime/                               live RoboCasa graph + Qwen runtime
authorities/                           recount, RGB, retention, validation evidence

The canonical package is a lazy source-backed view: canonical source graph ZIPs already present elsewhere in GWAM_Data are not duplicated inside this prefix. Each sidecar manifest row carries canonical_zip_sha256, and production loading requires the source authority closure. The standalone sidecar ZIP is sufficient for inspecting visual evidence but not for reconstructing simulator structure by itself.

1. Download

Install minimal download/inspection dependencies:

python -m pip install "huggingface_hub>=0.34" numpy

Download the complete prefix:

# If you cloned the whole dataset repository, run:
python gwam_active13_graph_adapter_qwen_v1_0_1_20260816/examples/download_release.py --local-dir GWAM_Data

# For snapshot_download users, run the Python block below, then set:
export GWAM_RELEASE="$PWD/GWAM_Data/gwam_active13_graph_adapter_qwen_v1_0_1_20260816"

Equivalent Python:

from huggingface_hub import snapshot_download

repo = "ChangChrisLiu/GWAM_Data"
prefix = "gwam_active13_graph_adapter_qwen_v1_0_1_20260816"
root = snapshot_download(
    repo_id=repo,
    repo_type="dataset",
    revision="main",  # replace with a verified commit for immutable experiments
    local_dir="GWAM_Data",
    allow_patterns=[f"{prefix}/**"],
)
print(f"release: {root}/{prefix}")

The complete release is approximately 28.8 GB, including the 4.27 GB Qwen model. For a small trial, use allow_patterns for the README, examples, semantic registry, and one episode ZIP.

2. Inspect an episode without extracting it

python "$GWAM_RELEASE/examples/inspect_episode.py" "$GWAM_RELEASE" \
  --episode target/TurnOnMicrowave/episode_000256

This command:

1. finds exactly one row in MANIFEST.jsonl;
2. validates the episode ZIP size and SHA-256;
3. reads receipt.json and NumPy arrays directly from the ZIP;
4. checks aligned shapes, finite embeddings, and camera ids;
5. reports heldout_opened=0.

Verified example:

episode=target/TurnOnMicrowave/episode_000256
frames=90
rows=4,077
embedding_mrl256=[4077,256]
camera_ids=[0,1,2]

3. Load Qwen evidence for one frame

Run the provided frame loader:

python "$GWAM_RELEASE/examples/load_frame_qwen_rows.py" "$GWAM_RELEASE" \
  --episode target/TurnOnMicrowave/episode_000256 \
  --frame 0

Import it in your own code:

import sys
from pathlib import Path

release = Path("GWAM_Data/gwam_active13_graph_adapter_qwen_v1_0_1_20260816")
sys.path.insert(0, str(release / "examples"))
from load_frame_qwen_rows import load_frame

visual = load_frame(
    release,
    episode="target/TurnOnMicrowave/episode_000256",
    frame_index=0,
)

print(visual["canonical_packet_sha256"])
print(visual["slot_id"].shape)                    # [M_frame]
print(visual["camera_name"][:5])                 # right / left / wrist
print(visual["bbox_xyxy_normalized"].shape)      # [M_frame,4]
print(visual["embedding_mrl256"].shape)           # [M_frame,256], float32 view

The canonical join key is:

(canonical_packet_sha256, slot_id, camera_name, bbox_xyxy_normalized)

Use slot_id to attach each visual row to the canonical node with the same stable slot. Preserve camera rows separately. Do not average cameras before the model unless your experimental method explicitly defines that ablation.

ZIP member contract

For F episode frames and M visual rows:

Member	Dtype	Shape	Meaning
receipt.json	JSON	scalar	episode and authority closure
frame_offsets.npy	int64	[F+1]	CSR frame-to-row offsets
frame.npy	int32	[M]	canonical frame index
source_frame.npy	int32	[M]	source RGB frame index
camera_id.npy	uint8	[M]	0=right, 1=left, 2=wrist
slot_id.npy	int16	[M]	canonical node slot
bbox_xyxy_normalized.npy	float32	[M,4]	normalized crop box
embedding_mrl256.npy	float16	[M,256]	unit-normalized Qwen vector
packet_sha256.npy	S32	[F]	canonical packet digest bytes
crop_rgb_sha256.npy	S32	[M]	exact crop digest bytes

For frame index t, rows are:

start = frame_offsets[t]
end = frame_offsets[t + 1]
frame_rows = slice(start, end)

Absence of a row means there is no valid visual measurement for that node/camera at that frame. It does not mean the physical node is absent.

4. Load the frozen semantic registry

Runtime text prompts are not used to recreate semantic vectors in the authenticated offline path. Load the frozen matrices:

python "$GWAM_RELEASE/examples/load_semantic_registry.py" "$GWAM_RELEASE"

Actual matrix shapes:

families_mrl256.npy     [11,256]  sparse canonical family-id table
categories_mrl256.npy   [171,256]
relations_mrl256.npy    [7,256]
tasks_mrl256.npy        [13,256]

There are 8 populated family phrases mapped into the 11-row canonical family-id table. Reserved/unpopulated ids stay present for index stability. Modeled relations are:

part_of, has_part,
controls, controlled_by,
component_of, has_component,
contact

Canonical visibility_change is retained as audit-only metadata and is not a default adapter relation.

5. Use the bundled Qwen model on new images

The model directory is the exact 18-file closure of:

model_id: Qwen/Qwen3-VL-Embedding-2B
revision: 9f2f7e710d6d81056aa5c0a4f04764fec6bb7bda
upstream bytes: 4,271,068,726
output width: 256 after MRL projection
instruction: Represent this RoboCasa graph image region.

Install the tested model stack:

python -m pip install numpy Pillow torch \
  "transformers==4.57.6" "qwen-vl-utils==0.0.14"

Embed an image:

python "$GWAM_RELEASE/examples/load_bundled_qwen.py" \
  "$GWAM_RELEASE" --image your_rgb.png

The loader authenticates 18 published model files and 18 HF metadata records, hydrates the model-local metadata cache, rejects path traversal/symlinks/special files/substitutions/unauthorized model scripts, and then returns a finite unit-normalized [1,256] embedding.

Using the Python class directly:

import sys
from pathlib import Path
from PIL import Image

release = Path("GWAM_Data/gwam_active13_graph_adapter_qwen_v1_0_1_20260816").resolve()
sys.path[:0] = [str(release), str(release / "code" / "adapter_package")]

from runtime.realtime.packaged_qwen import ensure_bundled_qwen_metadata
from graph_adapter_qwen.qwen_embedder import QwenEmbedder

ensure_bundled_qwen_metadata(release)
embedder = QwenEmbedder(
    repo_root=release,
    model_path="models/Qwen3-VL-Embedding-2B",
    token_budget=65536,
)
z = embedder.embed_images([Image.open("your_rgb.png").convert("RGB")])
assert z.shape == (1, 256)

When reproducing sidecar crops, use graph_adapter_qwen.rgb.crop_rgb; it implements the release crop policy, including normalized xyxy, 12% context, boundary clipping/padding, and a minimum 16-pixel side.

6. Build offline graph tokens for training

Production path

The production adapter is deliberately fail-closed. It requires:

• the canonical source graph/authority closure;
• an authenticated L=8 canonical window;
• the frozen semantic registry;
• the complete sidecar receipt population;
• RGB, model, migration, retention, recount, and validation authorities.

The exact executable reference is:

code/release_scripts/run_production_adapter_smoke_v1_0_1.py

Its core implementation is:

from graph_adapter_qwen.authority import CanonicalWindowAuthority
from graph_adapter_qwen.batch import collate_model_graph_windows
from graph_adapter_qwen.storage import DiskVisualSidecar, load_semantic_registry
from graph_adapter_qwen.model import ModelGraphPacketEncoder, K16GraphCodec
from graph_adapter_qwen.conditioner import GWAMConditioner

# These constructors require the full authenticated source/authority closure.
authenticated = CanonicalWindowAuthority(repo_root).window(
    "target/TurnOnMicrowave/episode_000256", start=0
)
batch = collate_model_graph_windows([authenticated], registry, sidecar)
packet = ModelGraphPacketEncoder(registry).eval()(batch)
graph = K16GraphCodec().eval()(packet)

assert graph.graph_tokens.shape == (1, 16, 384)
assert graph.graph_mask.shape == (1, 16)
assert graph.token_type == "POSTERIOR_T"

# Actions are supplied separately to the world model.
conditioned_hidden = GWAMConditioner(hidden_width=768)(
    hidden_states, graph, enabled=True
)

Production batch tensors include:

node_valid          bool     [B,8,N]
family_id           integer  [B,N]
category_id         integer  [B,N]
target_state                 [B,N]
pose                float    [B,8,N,9]
articulation        float    [B,8,N]
view8               float    [B,8,N,3,8]
image_mrl256        float    [B,8,N,3,256]
edge_sender         integer  [B,8,E]
edge_receiver       integer  [B,8,E]
edge_relation       integer  [B,8,E]
task_z              float    [B,256]

All continuous fields have explicit validity masks. Categorical ids are embedding indices, not continuous scalar features. Padding is not a real node.

Debug path

collate_model_graph_windows_debug() exists for shape tests and controlled ablations. It produces DEBUG_ABLATION_T, not POSTERIOR_T, and GWAMConditioner rejects it. Never rename or cast debug tokens to bypass authorization.

Training boundary

The packaged encoder, codec, and conditioner are newly initialized and untrained. The release proves shapes, authority, mutation rejection, and runtime execution—not learned quality, convergence, policy success, or held-out performance. Train your own model/checkpoint and publish that checkpoint separately with its training provenance.

7. Extract the graph during RoboCasa execution with Qwen

The live runtime performs this sequence without stepping the simulator between graph and camera capture:

current MuJoCo/RoboCasa state
  ├─ deterministic node inventory
  ├─ pose, velocity, articulation, visibility
  ├─ static part/control edges
  ├─ current contact edges
  └─ current camera segmentation/bboxes
          + synchronized right/left/wrist RGB
              + bundled Qwen crop encoder
                  → physical_graph + qwen_evidence

Install

bash "$GWAM_RELEASE/runtime/setup_realtime_qwen_env.sh" gwam-active13-live

Pinned source commits:

robosuite  5ce6643f3092639d08f7b0f90ed1c6a84f50552c
RoboCasa   b4684e6ee37d377cc392e98302a6b916d588b415
transformers 4.57.6
qwen-vl-utils 0.0.14

The setup downloads upstream RoboCasa kitchen assets. They are required and are not redistributed. Missing assets fail loudly instead of degrading to an incomplete graph.

One-state command

MUJOCO_GL=egl conda run -n gwam-active13-live \
  python "$GWAM_RELEASE/runtime/examples/realtime_qwen_graph.py" \
  --task OpenDrawer --robots PandaOmron --steps 1

Node and edge counts vary with the sampled scene/reset. Validate invariants instead of expecting one exact count:

N_real > 1
physical_graph.x.shape == [N_real,33]
physical_graph.edge_index.shape == [2,E]
physical_graph.edge_attr.shape == [E,8]
qwen_evidence.embedding_mrl256.shape == [M_live,256]
offline_production_authorized == false

The verified live smoke produced a valid OpenDrawer graph and Qwen evidence. Different verified resets produced different valid N, E, and visible-row counts, which is expected.

Use inside an evaluation or policy loop

from runtime.realtime.qwen_realtime_graph import RealtimeQwenGraphExtractor

obs = env.reset()
extractor = RealtimeQwenGraphExtractor.from_env(
    env,
    release_root=GWAM_RELEASE,
    model_path="models/Qwen3-VL-Embedding-2B",
)

for step in range(max_steps):
    result = extractor.extract()  # does not advance env

    physical = result["physical_graph"]
    evidence = result["qwen_evidence"]

    x = physical["x"]                    # [N_real,33]
    edge_index = physical["edge_index"]  # [2,E]
    edge_attr = physical["edge_attr"]    # [E,8]

    slot_id = evidence["slot_id"]
    view_id = evidence["view_id"]
    bbox = evidence["bbox_xyxy_normalized"]
    qwen = evidence["embedding_mrl256"]  # [M_live,256]

    # Your live model must preserve this camera-specific mapping.
    model_input = {
        "physical_graph": physical,
        "qwen_evidence": evidence,
        "observation": obs,
    }
    action = policy(model_input)
    obs, reward, done, info = env.step(action)
    if done:
        break

Call extractor.reset_temporal_state() after a hard environment reset if you reuse the extractor object. Prefer rebuilding the extractor after a reset when episode metadata or object inventory may change.

Live output schema

result
├─ schema = gwam.realtime.graph_plus_qwen.v1
├─ t
├─ physical_graph
│  ├─ x                 [N_real,33]
│  ├─ edge_index        [2,E]
│  └─ edge_attr         [E,8]
├─ physical_snapshot    slot-aligned current-state tensors and masks
├─ rgb_frames           {0:right RGB, 1:left RGB, 2:wrist RGB}
├─ rgb_frame_cameras
├─ qwen_evidence
│  ├─ trust_domain
│  ├─ offline_production_authorized = false
│  ├─ active_slot_id    [N_real]
│  ├─ slot_id           [M_live]
│  ├─ view_id           [M_live]
│  ├─ camera_alias      [M_live]
│  ├─ bbox_xyxy_normalized [M_live,4]
│  ├─ crop_rgb_sha256   [M_live]
│  ├─ embedding_mrl256  [M_live,256]
│  ├─ node_mean_mrl256_non_authoritative [N_real,256]
│  └─ node_mean_valid   [N_real]
└─ offline_production_authorized = false

node_mean_mrl256_non_authoritative is only a convenience view. The camera-specific rows are the evidence authority.

What the live extractor does and does not claim

It does:

• read the current simulator state;
• construct a current physical graph;
• render synchronized camera frames;
• crop currently visible node regions;
• run the authenticated bundled Qwen model;
• preserve slot, view, bbox, and crop-hash provenance.

It does not:

• infer physical relations with Qwen;
• create an offline canonical receipt;
• authorize POSTERIOR_T training tokens;
• include future state, future contacts, or future actions;
• prove that the untrained adapter improves a policy.

8. Verification commands

Run portable examples:

python "$GWAM_RELEASE/examples/inspect_episode.py" "$GWAM_RELEASE" \
  --episode target/TurnOnMicrowave/episode_000256
python "$GWAM_RELEASE/examples/load_frame_qwen_rows.py" "$GWAM_RELEASE" \
  --episode target/TurnOnMicrowave/episode_000256 --frame 0
python "$GWAM_RELEASE/examples/load_semantic_registry.py" "$GWAM_RELEASE"
python "$GWAM_RELEASE/examples/load_bundled_qwen.py" "$GWAM_RELEASE"

Run focused tests:

PYTHONPATH="$GWAM_RELEASE/runtime:$GWAM_RELEASE/code/adapter_package" \
  python -m pytest -q "$GWAM_RELEASE/runtime/tests/test_qwen_realtime_graph.py"

PYTHONPATH="$GWAM_RELEASE/code/canonical_graph_v1_0_3:$GWAM_RELEASE/code/adapter_package" \
  python -m pytest -q "$GWAM_RELEASE/code/tests/test_adapter_model.py"

Run the full manifest verifier:

python "$GWAM_RELEASE/code/release_scripts/verify_expanded_qwen_release_v1_0_1.py" \
  --release "$GWAM_RELEASE"

Release verification covers all episode identities, array contracts, semantic matrices, exact Qwen closure, code fingerprints, symlink/special-file rejection, held-out non-access, and remote Git/LFS object identity.

9. Common errors

EPISODE_MATCH_COUNT:0

The episode key is wrong or the corresponding ZIP was not downloaded. Use keys from MANIFEST.jsonl, without the leading episodes/ or trailing .zip.

MODEL_*_MISMATCH or QWEN_*_HASH

A model or metadata file is missing/changed. Redownload from a verified revision. Do not bypass the check or enable remote model code.

RoboCasa asset XML is missing

Run runtime/setup_realtime_qwen_env.sh and complete upstream asset installation. The runtime intentionally does not substitute fake assets.

Robot-only or unexpectedly tiny live graph

Episode metadata was unavailable or the RoboCasa/robosuite versions are incompatible. Pass the base env, verify the pinned commits, and ensure object_cfgs and fixtures are populated after reset.

CUDA out of memory

Reduce Qwen token_budget, run fewer crops per call, or use a larger GPU. Do not silently drop rows without a validity/missingness record.

Live tokens are rejected by GWAMConditioner

This is expected. Live outputs are not offline authenticated POSTERIOR_T. Implement and validate a separately named live adapter/checkpoint rather than forging production authorization.

10. Scientific and safety boundaries

MODEL_TRAINED=false
HELDOUT_EVALUATED=false
HELDOUT_OPENED=0
LIVE_OUTPUT_IS_OFFLINE_CANONICAL=false
QWEN_CAN_MODIFY_STRUCTURE=false
ACTIONS_INSIDE_GRAPH_TOKENS=false

Use past/current observations only for causal rollout evaluation. Future simulator state, future robot nodes, future contacts, and goal-derived target edges are oracle/teacher-forcing information unless explicitly labeled as such.

Licensing and attribution

• RoboCasa-derived dataset content follows repository-level LICENSE_PENDING.md terms.
• Bundled Qwen/Qwen3-VL-Embedding-2B files are Apache-2.0; see models/QWEN3_VL_EMBEDDING_APACHE_2_0.txt and the upstream model card.
• RoboCasa kitchen assets are not included; obtain them upstream and follow their license terms.

Please cite RoboCasa and Qwen3-VL-Embedding when using their simulator data or model. Preserve the manifest, receipts, model revision, and exact Hugging Face revision with experimental artifacts.

Previous repository documentation

The README that preceded this Active-13 dataset card remains preserved byte-for-byte at:

docs/archive/README_GWAM_V3_0_7_BEFORE_ACTIVE13_QWEN_20260817.md

SHA-256: 2617e4559396cf9a7da214e5d9d66309579f68a951b9804d91a1cd86cf913e61

Legacy namespaces and immutable revisions remain available; this card documents the current additive Active-13 release.