# CogWAM — RoboDojo, 25-step horizon, 50k steps

Released weights for the CogWAM recipe
`cogwam_robodojo_causal_dino_mot_h25_eventmem_dino_multilayer_50k_v1`,
taken at optimizer step 50,000.

## Contents

| File | Size | |
|---|---|---|
| `model.safetensors` | 7.80 GiB | 2272 BF16 tensors: the framework state dict |
| `checkpoint_keys.json` | 215 KiB | `key -> [dtype, shape]`, for validating without loading weights |
| `dataset_statistics.json` | 5 KiB | action/state normalisation — **the policy server cannot un-normalise actions without it** |
| `inference_config.yaml` | | the resolved configuration this checkpoint was trained with |
| `artifact_manifest.json` | | sha256 of each file, model geometry, backbone requirements |

## What is NOT here

The two backbones are referenced, not bundled, because one of them may not be
redistributed:

| Backbone | Where to get it | Licence |
|---|---|---|
| RynnBrain1.1-2B | `Alibaba-DAMO-Academy` on Hugging Face | Apache-2.0 |
| DINOv3 ViT-B/16 | `facebook/dinov3-vitb16-pretrain-lvd1689m` | **Meta `dinov3-license`, gated** — accept Meta's terms yourself |

The DINO teacher is also genuinely absent from the state dict: it is attached
outside the module tree and its normalisation buffers are non-persistent, so it
never enters `state_dict()`.

## Loading

```bash
export COGWAM_BASE_VLM=/models/rynnbrain1.1-2B
export COGWAM_DINO_MODEL=/models/dinov3-vitb16

python tools/verify_checkpoint.py --artifact <this directory> --build
```

`--build` constructs the framework from `inference_config.yaml` and loads with
`strict=True`. No key remapping is needed or performed: checkpoint key prefixes
come from `nn.Module` attribute names, not from the framework class name.

To serve it:

```bash
COGWAM_ARTIFACT_DIR=<this directory> NUM_SERVERS=8 bash scripts/serve_policy.sh
```

## Geometry

| | |
|---|---|
| Tensors / dtype | 2272 / BF16 |
| Prefixes | `action_model.` 1652, `qwen_vl_interface.` 618, `world_plan_queries.embedding`, `action_plan_queries.embedding` |
| Action horizon / dim | 25 / 14 |
| World queries | 16 |
| Planner hidden | 2048 |
| Vocabulary | 248081 (includes `<WORLD_PLAN>`, `<ACTION_PLAN>`, `<KEEP>`, `<UPDATE>`) |
| MoT | 30 layers; world 512, action 1024 |
| DINO embedding | 768 |

Two digests are recorded in the manifest and they answer different questions:
`model.key_inventory_sha256` covers names/shapes/dtypes, so two runs of the same
recipe share it; only `files["model.safetensors"].sha256` identifies these
particular weights.

## Evaluation

Run against RoboDojo 0.2.0 at commit
`9b4cc885e8f530ed3ab14a30a312ae70242771c4`. Protocol, chunk/replan semantics and
the event-memory state machine are documented in the repository under
`docs/evaluation.md`. Published results are not included with this release.

## Licence

MIT for the code; these weights are released under the same terms. The two
backbone models carry their own licences, listed above.