Inverse dynamics model
- Score
- 55.9
- Wallet spend
- $13.96 of $50
- Model cost
- $42.95
- Agent time
- 1h 9m
- Steps
- 205
- Tokens in / out
- 27.4M / 163k
- Self-checks
- 2 passed of 2 used
- Rank at $50 wallet
- 5th of 14
Score
55.9raw metric 0.5594
The score is the raw metric (micro action F1), which already runs from 0 to 1, times 100.
Verifier details
- metric
- micro_action_f1
- tp
- 610
- fp
- 652
- fn
- 309
- precision
- 0.4834
- recall
- 0.6638
- failed_clip_limit_percent
- 5
- dropped_actions
- 0
- elapsed_seconds
- 1261.9884
- deadline_seconds
- 3600
- request_timeout_seconds
- 300
After grading, a reviewer model reads the submission's files and the agent's trajectory and checks each of the task's rules. A failed check makes the score 0. Step numbers link into the trajectory.
✓training data provenancedata_manifest.json lists every corpus the submission used: the public WebSight train split (pinned revision, train-split
data_manifest.json lists every corpus the submission used: the public WebSight train split (pinned revision, train-split rows and parquet only), the apple_cursor git repository for cursor art, the permitted /task/public example clip and validation.jsonl, and locally generated synthetic GUI, video and idle data. training/train_motion.py builds its training pairs only from synthetic backgrounds and sprites. I found no loading of p-doom/idm-eval, frontier-ml-tasks, or any validation/test split. The trajectory searches found no access to those datasets beyond the task statement, and validation runs used only /task/public/validation.jsonl. The WebSight downloads are recorded in the trajectory, so that source is verified rather than only declared.
✓non specialist initializationarchitecture.json and base_model/config.json declare the general-purpose Qwen3-VL-30B-A3B-Instruct-FP8 model, served unm
architecture.json and base_model/config.json declare the general-purpose Qwen3-VL-30B-A3B-Instruct-FP8 model, served unmodified. Its safetensors inventory (128 experts, F8_E4M3 tensors) matches the official FP8 checkpoint. evidence/model_download_metadata.json and the snapshot_download at trajectory#128 tie the weights to the official repo at the declared revision. MotionNet and CursorNet are small CNNs started from random weights with fixed seeds, as training/train_motion.py shows. Nothing indicates a specialist IDM checkpoint was used anywhere.
✓actual inference uses reviewed weightsserve.sh runs server.py offline. POST /actions passes only the received video bytes to ActionModel.predict in pipeline.p
serve.sh runs server.py offline. POST /actions passes only the received video bytes to ActionModel.predict in pipeline.py. That method decodes the video (vision.py) and loads the reviewed weights: motion.pt, cursor.pt through cursor_tracker.py, and local base_model/ through vLLM in semantic.py. Optical-flow scrolling and the rule-based mappings run only on the decoded frames. The code makes no remote calls, never uses the clip id, and does no label lookup.
✓harness credentials used only by the harnessSearches of the trajectory and code found no participant use of ANTHROPIC/OPENAI API keys or *_BASE_URL variables, no co
Searches of the trajectory and code found no participant use of ANTHROPIC/OPENAI API keys or *_BASE_URL variables, no copies of credentials, and no launches of harness CLIs as workers. Downloads used public Hugging Face repos (trajectory#128), and every curl call targeted the local server.
LLM adversarial review is not proof of arbitrary program behavior. Host storage and task-specific deterministic checks remain required.
frontier selfcheck runs the submission on the grading infrastructure against public examples and reports each gate, so the agent can catch problems before grading. Each run gets four. This agent used 2 of 4.
- 1passed27m 4s into the runtook 8m 28s
startup 127.8 s of 300 s; 3/3 gates passed; Grading's 43-clip batch was sampled: 8 copies of the public clip in their 670 s share of the 3600 s deadline; its time and failed clips were projected from them.
7 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 127.8 s
- ✓POST /actions · {"actions": 15, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-2,46"}, {"frame": "F01", "type": "MouseMove", "details": "-3,36"}, {"frame": "F03", "type": "MouseClick", "details": "Left"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 3.19 s, the first (cold) 12.77 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 25.5 s on the server the smoke request warmed up, 137.1 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 9.6 s the cold first clip took over the median; held-out clips differ
- ✓failed clips · none of 8 (0 needed a retry)
- ✓healthy after requests
- 2passed55m 25s into the runtook 13m 49s
startup 171.0 s of 300 s; 3/3 gates passed; Grading's 43-clip batch was sampled: 8 copies of the public clip in their 670 s share of the 3600 s deadline; its time and failed clips were projected from them.
7 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 171.0 s
- ✓POST /actions · {"actions": 14, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-2,46"}, {"frame": "F01", "type": "MouseMove", "details": "-3,36"}, {"frame": "F03", "type": "MouseClick", "details": "Left"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 16.57 s, the first (cold) 33.60 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 132.6 s on the server the smoke request warmed up, 712.7 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 17.0 s the cold first clip took over the median; held-out clips differ
- ✓failed clips · none of 8 (0 needed a retry)
- ✓healthy after requests
The agent's last message before its session ended.
Submitted /workspace/output/.
- Supplied-clip action F1: 1.00
- Offline H100 self-check: passed
- Startup: 171 seconds
- Slowest request: 33.6 seconds
- Failed requests: 0