Inverse dynamics model
- Score
- 39.9
- Wallet spend
- $13.96 of $200
- Model cost
- $5.16
- Agent time
- 1h 49m
- Steps
- 182
- Tokens in / out
- 16.9M / 213k
- Self-checks
- 2 passed of 2 used
- Rank at $200 wallet
- 7th of 14
Score
39.9raw metric 0.3989
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
- 428
- fp
- 799
- fn
- 491
- precision
- 0.3488
- recall
- 0.4657
- failed_clip_limit_percent
- 5
- dropped_actions
- 0
- elapsed_seconds
- 1049.3802
- 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 corpus that was used is declared. data_manifest.json, architecture.json and evidence/training_commands.txt list fo
Every corpus that was used is declared. data_manifest.json, architecture.json and evidence/training_commands.txt list four inputs: the single permitted /task/public example, procedurally rendered GUI and cursor composites, typing and selection calibration clips, and the public apple_cursor artwork (git clone logged in evidence/cursor_artwork_download.log). training/train_action_head.py renders 5000 procedural pairs and adds 6 hand-labelled pairs from the permitted example. The cursor trainer (training/train_cursor.py) uses only synthetic composites. Validation was run only on the permitted example. In the trajectory, idm-eval appears only in the task prompt and in manifest text. There is no access to frontier-ml-tasks and no dataset downloads; the synthetic calibration clips were made by the participant's own generator.
✓non specialist initializationThe base is the general-purpose Qwen/Qwen3-VL-8B-Instruct at a pinned revision. It was downloaded through training/downl
The base is the general-purpose Qwen/Qwen3-VL-8B-Instruct at a pinned revision. It was downloaded through training/download_model.py, which asserts a Qwen/Qwen3-VL-* repo, and evidence/base_download.log records the download. The shard inventories show standard Qwen3-VL tensors (hidden size 4096), and the model is used frozen. CursorNet is randomly initialized in training/train_cursor.py and records that in its checkpoint (weights/cursor_initial.pt, weights/cursor.pt). The action head starts as a zero tensor. Nothing loads p-doom/idm or any IDM descendant.
✓actual inference uses reviewed weightsserve.sh sets offline mode, installs from local wheels and runs server.py. server.py loads weights/cursor.pt, weights/vl
serve.sh sets offline mode, installs from local wheels and runs server.py. server.py loads weights/cursor.pt, weights/vlm (local_files_only), weights/action_head.json and weights/scroll.json. Each /actions request decodes the uploaded video bytes (video_features.py) and runs the cursor tracker (cursor_tracking.py), the scroll model and the semantic model (semantic_model.py: local Qwen generation, OCR, action head). The request id is ignored. There are no remote inference calls, no hash, ID or order lookup, and no packaged labels.
✓harness credentials used only by the harnessTrajectory searches for API_KEY, BASE_URL, OPENAI, printenv and claude return no participant use. HF_TOKEN appears only
Trajectory searches for API_KEY, BASE_URL, OPENAI, printenv and claude return no participant use. HF_TOKEN appears only in the prompt. The only credentials the participant handled were Spin wallet transfer credentials (`spin transfer credentials`), used by training/storage_transfer.py for S3 transfers. The policy allows these.
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.
- 1passed1h 26m into the runtook 7m 40s
startup 62.5 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 62.5 s
- ✓POST /actions · {"actions": 15, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-2,24"}, {"frame": "F01", "type": "MouseMove", "details": "-10,63"}, {"frame": "F04", "type": "MouseClick", "details": "Left"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 12.97 s, the first (cold) 15.89 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 111.4 s on the server the smoke request warmed up, 599.0 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 3.0 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
- 2passed1h 39m into the runtook 7m 56s
startup 79.1 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 79.1 s
- ✓POST /actions · {"actions": 14, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-2,24"}, {"frame": "F01", "type": "MouseMove", "details": "-10,63"}, {"frame": "F04", "type": "MouseClick", "details": "Left"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 11.85 s, the first (cold) 15.03 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 95.3 s on the server the smoke request warmed up, 512.0 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 3.2 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/.
- Includes
serve.sh, trained weights, offline dependencies, and provenance logs. - Final H100 selfcheck passed all gates; startup took 79.1 seconds.
- Provided-example action F1: 1.00 — a training/calibration result, not a held-out estimate.