Inverse dynamics model
- Score
- 44.9
- Wallet spend
- $4.60 of $10
- Model cost
- $29.00
- Agent time
- 56m
- Steps
- 158
- Tokens in / out
- 18.4M / 129k
- Self-checks
- 2 passed of 2 used
- Rank at $10 wallet
- 5th of 14
Score
44.9raw metric 0.4492
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
- 451
- fp
- 638
- fn
- 468
- precision
- 0.4141
- recall
- 0.4908
- failed_clip_limit_percent
- 5
- dropped_actions
- 0
- elapsed_seconds
- 666.0023
- 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 provenanceThe only trained component is CursorNet, and its training data is fully synthetic. training/train_cursor.py and training
The only trained component is CursorNet, and its training data is fully synthetic. training/train_cursor.py and training/make_assets.py build the composites from procedurally generated backgrounds (seed 723), cursor sprites from ful1e5/apple_cursor (logged in evidence/cursor_assets_download.json) and original geometry. data_manifest.json declares every corpus: the permitted /task/public example (with sha256), the synthetic assets, and the frozen Qwen and OCR downloads. The Apple-menu template was extracted only from that permitted example. The trajectory shows no idm-eval access, no load_dataset call and no use of other datasets' val/test splits. Training ran on spin, per trajectory#67 and evidence/train_cursor.log.
✓non specialist initializationThe base model is the general-purpose Qwen/Qwen3-VL-8B-Instruct, pinned to a revision and downloaded via snapshot_downlo
The base model is the general-purpose Qwen/Qwen3-VL-8B-Instruct, pinned to a revision and downloaded via snapshot_download (evidence/download_model.py, evidence/model_download.log). It is used frozen, and its safetensors contain standard Qwen3-VL tensors with no extra adapters. CursorNet starts from random initialization with no pretrained cursor or IDM weights (training/train_cursor.py, evidence/train_cursor.log). The OCR models are general-purpose PP-OCRv5 models from RapidOCR, not IDMs.
✓actual inference uses reviewed weightsserve.sh starts server.py. server.py loads cursor.pt through vision.CursorTracker, ocr_models/*.onnx through ocr.ScreenO
serve.sh starts server.py. server.py loads cursor.pt through vision.CursorTracker, ocr_models/*.onnx through ocr.ScreenOCR, and the local base_model through semantic.SemanticModel (Qwen3VLForConditionalGeneration with local_files_only and offline env vars set). The rest of the pipeline is rule-based: frame sampling, Viterbi cursor tracking, optical-flow scroll detection, OCR diffs and key-action canonicalization. There is no lookup by clip ID or hash, the id field is unused, and there are no remote inference calls.
✓harness credentials used only by the harnessIn the trajectory, the participant used only spin commands (spin s5cmd, spin machines) and a localhost /health curl. Ope
In the trajectory, the participant used only spin commands (spin s5cmd, spin machines) and a localhost /health curl. Openrouter appears only in spin --help output (trajectory#6). Searches found no reads or copies of harness API keys or base URLs, and no launches of extra codex or claude CLIs. Remote training used spin with presigned URLs (trajectory#67).
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.
- 1passed35m 37s into the runtook 7m 33s
startup 94.8 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 94.8 s
- ✓POST /actions · {"actions": 14, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-2,45"}, {"frame": "F01", "type": "MouseMove", "details": "-8,45"}, {"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 6.38 s, the first (cold) 8.27 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 51.5 s on the server the smoke request warmed up, 277.0 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
- 2passed49m 16s into the runtook 7m 55s
startup 101.1 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 101.1 s
- ✓POST /actions · {"actions": 14, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-2,45"}, {"frame": "F01", "type": "MouseMove", "details": "-8,45"}, {"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 8.48 s, the first (cold) 9.52 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 66.8 s on the server the smoke request warmed up, 359.0 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 1.5 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/ via frontier submit.
- Supplied-example action F1: 1.00
- Final offline H100 self-check: passed
- Startup: 101.1 seconds
- Slowest request: 10.5 seconds, with no failures
Includes serve.sh, trained weights, offline dependencies, and training/provenance records.