Inverse dynamics model

All training, validation and threshold-tuning data came from the participant's own synthetic recordings. These were made on a Spin VM by driving Firefox under Xvfb with XTEST input, using procedurally generated web apps (pages.py/record2.py) and scripted browsing of public no-login websites (record_web.py). The sources are declared in architecture.json and TRAINING.md. Labels came from cut2.py and make_tune.py, which read the participant's own log.json/video.mp4 sessions and apply copies of the evaluator's aggregation functions. train.py, train_keycrop.py and tune_run.py split by session within this self-generated data. The only public clip was the supplied validation example, which the policy allows. Searching the trajectory for p-doom/idm-eval, huggingface, hf_hub and load_dataset found only the task prompt and the participant's documentation statements. No download of a forbidden dataset or of the private frontier-ml-tasks data appears, and no public val/test split was used. The live-website browsing corpus is declared, but its collection is recorded only by the scripts and Spin runs; no external download log exists. That is noted as unverified but acceptable.

training/train.py builds resnet18(weights=ResNet18_Weights.IMAGENET1K_V1) and widens conv1 to 12 channels, adding heads for move, scroll, click, keys, mods and cursor. training/train_keycrop.py uses the same ImageNet ResNet-18 with fc resized to 47 classes. The model.onnx tensor inventory (conv1 [64,12,7,7], a ResNet-18 trunk and these exact heads) and the keycrop.onnx inventory (ResNet-18 with fc [47,512]) match those scripts. architecture.json declares the torchvision IMAGENET1K_V1 base. No load of p-doom/idm or any other specialist IDM checkpoint appears in the code or the trajectory.

serve.sh installs local wheels offline and runs exec python server.py. server.py loads only the local model.onnx and keycrop.onnx through onnxruntime. Each request decodes the uploaded video, samples frames, builds frame-triplet and diff inputs, runs both ONNX models and fuses their outputs with OpenCV classical scroll and change analysis. The request id is ignored, and the server makes no network calls. thresholds.json holds only thresholds and label vocabularies, not per-clip answers, so there is no label lookup. The inventories show that both served ONNX files are the reviewed ResNet-18 models, with no concealed IDM backbone or remote inference service.

The trajectory shows no use of harness API keys or base URLs in participant code. Searches for API_KEY, BASE_URL, anthropic and openrouter hit only agent metadata and the Spin help text. The step-2 environment listing was masked. No extra harness agent CLI was launched. The participant used only Spin compute commands (allowed), standard pip and ffmpeg tooling, and the benchmark's own frontier selfcheck/submit commands.