Inverse dynamics model
- Score
- 45.1
- Wallet spend
- $38.39 of $50
- Model cost
- $11.31
- Agent time
- 2h 52m
- Steps
- 356
- Tokens in / out
- 33.9M / 254k
- Self-checks
- 1 passed of 1 used
- Rank at $50 wallet
- 10th of 19
Score
45.1raw metric 0.4513
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
- 345
- fp
- 265
- fn
- 574
- precision
- 0.5656
- recall
- 0.3754
- failed_clip_limit_percent
- 5
- dropped_actions
- 0
- elapsed_seconds
- 201.5112
- 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 provenanceTraining data comes only from the public videos in p-doom/AGI-CAST-0.6k and the weak labels in p-doom/AGI-CAST-idm-actio
Training data comes only from the public videos in p-doom/AGI-CAST-0.6k and the weak labels in p-doom/AGI-CAST-idm-actions, both pinned to a revision. The code (prepare_training.py) and data_manifest.json both say so. Models were chosen on 3 whole AGI-CAST videos held out from training, plus the /task/public example, which the policy allows. p-doom/idm-eval shows up only as a name in a list of datasets (trajectory#12). Nothing was downloaded from it, and neither idm-eval nor frontier-ml-tasks data was used.
✓non specialist initializationThe served models are scikit-learn HistGradientBoosting classifiers trained from scratch on hand-built visual features.
The served models are scikit-learn HistGradientBoosting classifiers trained from scratch on hand-built visual features. architecture.json declares no pretrained base. The only pretrained model downloaded was Qwen3-VL-4B-Instruct, a general vision-language model that the policy allows. It was only an experiment and is marked served=false. No p-doom/idm weights or their descendants were downloaded.
✓actual inference uses reviewed weightsserve.sh installs local wheels with no network access, turns on offline mode and runs server.py. server.py reads frames
serve.sh installs local wheels with no network access, turns on offline mode and runs server.py. server.py reads frames from the uploaded video, computes features (frame_features.py, direction_features.py), runs the shipped joblib classifiers and decodes events with action_utils.py. It makes no remote calls, does not look actions up by clip ID or hash, and stores no evaluation labels. inference_config.json disables the VLM path.
✓harness credentials used only by the harnessIn the session, no participant step read, copied, echoed or used the harness API key or base URL, and none started extra
In the session, no participant step read, copied, echoed or used the harness API key or base URL, and none started extra harness CLIs. A ps listing only showed that the harness processes were running (trajectory#61). The presigned S3 URL came from the participant's own Spin transfer session (trajectory#24), not from harness credentials.
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 1 of 4.
- 1passed2h 52m into the runtook 1m 6s
startup 5.3 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 5.3 s
- ✓POST /actions · {"actions": 14, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "0,0"}, {"frame": "F01", "type": "MouseMove", "details": "0,0"}, {"frame": "F02", "type": "MouseClick", "details": "Left"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 3.28 s, the first (cold) 3.13 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 26.8 s on the server the smoke request warmed up, 144.2 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 0.0 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.
frontier selfcheck passed on the exact frozen submission (digest ee6e822a…). Startup was 5.3 seconds, the slowest request was 3.64 seconds, projected 43-clip serving was 144.2 seconds, and there were zero failed clips or retries. The architecture review remains separate, as the report notes. I’m proceeding with the requested two-step submission.