Inverse dynamics model
- Score
- 67.2
- Wallet spend
- $20.14 of $200
- Model cost
- $34.22
- Agent time
- 1h 19m
- Steps
- 178
- Tokens in / out
- 21.0M / 147k
- Self-checks
- 2 passed of 2 used
- Rank at $200 wallet
- 3rd of 14
Score
67.2raw metric 0.6723
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
- 631
- fp
- 327
- fn
- 288
- precision
- 0.6587
- recall
- 0.6866
- failed_clip_limit_percent
- 5
- dropped_actions
- 0
- elapsed_seconds
- 493.0350
- 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 provenanceEvery training and calibration source is declared in architecture.json, and the scripts match those declarations. The so
Every training and calibration source is declared in architecture.json, and the scripts match those declarations. The sources are: two named train-split parquet shards of osunlp/Multimodal-Mind2Web at pinned rev 1b4c6a8c…, used only as background screenshots; ful1e5/apple_cursor SVGs, cloned at a pinned revision; procedurally generated cursors and clips; and one I-beam template taken from the permitted /task/public example. The downloads are logged in evidence/downloads.jsonl. Trajectory step 12 shows the actual hf_hub_download of only the train shards and the git clone of apple_cursor. Searching the session found no access to p-doom/idm-eval, inductionlabs/frontier-ml-tasks, or validation/test splits of other datasets.
✓non specialist initializationCursorNet is a small U-Net (about 388K parameters). training/train_cursor.py builds it from random initialization (seed
CursorNet is a small U-Net (about 388K parameters). training/train_cursor.py builds it from random initialization (seed 8102). training/refine_cursor.py only resumes from the participant's own checkpoint (--init). The cursor.pt tensor inventory matches the CursorNet state_dict layout. The semantic component is the general-purpose Qwen3-VL-32B-Instruct, downloaded at a pinned revision (evidence/base32_download.jsonl, training/download_base32_candidate.py). base/config.json matches the Qwen3-VL-32B architecture: 64 layers, hidden size 5120. No p-doom/idm or other specialist IDM checkpoint is downloaded or loaded anywhere in the code or the session.
✓actual inference uses reviewed weightsserve.sh sets HF_HUB_OFFLINE and TRANSFORMERS_OFFLINE, installs only from local wheels, and starts uvicorn server:app. A
serve.sh sets HF_HUB_OFFLINE and TRANSFORMERS_OFFLINE, installs only from local wheels, and starts uvicorn server:app. At startup server.py loads CursorDetector(ROOT/'cursor.pt') and PairSemanticModel(ROOT/'base'). semantic_pairs.py calls Qwen3VLForConditionalGeneration.from_pretrained(path, local_files_only=True) on the shipped base/ safetensors shards. vision.py runs the CursorNet weights plus OpenCV optical flow. The request id and filename are not used as features. There are no remote API calls, no label lookups and no hashing tables in the serving path.
✓harness credentials used only by the harnessThe trajectory shows only allowed tools: spin budget/compute/openrouter help and compute commands, public HF downloads,
The trajectory shows only allowed tools: spin budget/compute/openrouter help and compute commands, public HF downloads, pip and git. Searches for API keys, BASE_URL, printenv, auth.json, ANTHROPIC, codex exec and similar found no participant use or copying of harness credentials, and no extra harness CLI was launched.
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.
- 1passed25m 26s into the runtook 9m 26s
startup 82.3 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 82.3 s
- ✓POST /actions · {"actions": 41, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-5,41"}, {"frame": "F01", "type": "MouseMove", "details": "-2,31"}, {"frame": "F03", "type": "KeyPress", "details": "F"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 3.20 s, the first (cold) 4.94 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 25.6 s on the server the smoke request warmed up, 137.4 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 1.8 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
- 2passed1h 1m into the runtook 17m 50s
startup 83.2 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 83.2 s
- ✓POST /actions · {"actions": 12, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-5,40"}, {"frame": "F01", "type": "MouseMove", "details": "-2,31"}, {"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 6.08 s, the first (cold) 8.00 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 49.0 s on the server the smoke request warmed up, 263.4 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 1.9 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-example F1: 0.923. Final offline self-check passed: 83.2-second startup, 8.0-second slowest request, and zero failures.