Inverse dynamics model
- Score
- 68.7
- Wallet spend
- $36.16 of $50
- Model cost
- $36.62
- Agent time
- 4h 37m
- Steps
- 577
- Tokens in / out
- — / —
- Self-checks
- 2 passed of 2 used
- Rank at $50 wallet
- 2nd of 19
Score
68.7raw metric 0.6873
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
- 246
- fn
- 309
- precision
- 0.7126
- recall
- 0.6638
- failed_clip_limit_percent
- 5
- dropped_actions
- 0
- elapsed_seconds
- 363.7476
- 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 training data is declared in architecture.json and is downloaded by src/prepare_agicast_data.py through hf_hub_downl
The training data is declared in architecture.json and is downloaded by src/prepare_agicast_data.py through hf_hub_download. It is the public train splits of p-doom/AGI-CAST-0.6k (videos, CC0) and p-doom/AGI-CAST-idm-actions (labels, Apache-2.0). The provenance is recorded in data_manifest.json and logs/prepare_data.log. The labels are pseudo-labels that p-doom's IDM generated. They are a public training corpus, not idm-eval, and no specialist weights were obtained to make them. For selection, the submission holds out three AGI-CAST train-split sessions, kept session-disjoint. src/rebalance_data.py also adds windows from the public /task/public validation example to training, which the policy allows. In the trajectory, 'idm-eval' appears only in the task prompt and in the participant's exclusion filters (e.g. `if 'idm-eval' not in id`). There is no download of p-doom/idm-eval or of inductionlabs/frontier-ml-tasks, and no other benchmark's val/test data.
✓non specialist initializationThe base model is Qwen/Qwen3-VL-8B-Instruct, a general VLM, at the revision declared in architecture.json. The trajector
The base model is Qwen/Qwen3-VL-8B-Instruct, a general VLM, at the revision declared in architecture.json. The trajectory shows it fetched with snapshot_download. src/train_idm.py loads it with AutoModelForImageTextToText.from_pretrained and creates a new LoRA with get_peft_model. The optional --init-adapter was used only for the participant's own stage-1 adapter during stage-2 refinement. In lora_adapter/adapter_config.json, the base is the downloaded Qwen and the class is Qwen3VLForConditionalGeneration. No p-doom/idm checkpoint or other specialist IDM checkpoint was downloaded or loaded.
✓actual inference uses reviewed weightsserve.sh sets HF_HUB_OFFLINE and TRANSFORMERS_OFFLINE and runs server.py. server.py loads the local base_model shards, m
serve.sh sets HF_HUB_OFFLINE and TRANSFORMERS_OFFLINE and runs server.py. server.py loads the local base_model shards, merges lora_adapter/adapter_model.safetensors with PeftModel merge_and_unload, and decodes the uploaded video with ffmpeg. It then runs sliding-window generate calls and applies the hand-written pixel-diff postprocessing in src/actions_utils.py. The clip id is used only for logging. There are no remote inference calls, no hidden IDM backbone and no stored-label lookup.
✓harness credentials used only by the harnessThe trajectory has no matches for harness API keys, base URLs or printenv-style credential access, and the participant l
The trajectory has no matches for harness API keys, base URLs or printenv-style credential access, and the participant launched no extra harness CLIs. The only credentials used were Spin's own: Spin compute, plus Spin S3 transfer credentials from `spin transfer credentials`, used in src/upload_s3.py and src/auto_sync_s3.py. Those are allowed. HF access was used only for downloads.
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.
- 1passed2h 10m into the runtook 7m 11s
startup 66.1 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 66.1 s
- ✓POST /actions · {"actions": 17, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "100,106"}, {"frame": "F01", "type": "MouseMove", "details": "10,10"}, {"frame": "F02", "type": "MouseMove", "details": "1,0"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 7.80 s, the first (cold) 7.74 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 62.4 s on the server the smoke request warmed up, 335.6 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
- 2passed4h 28m into the runtook 7m 56s
startup 90.3 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 90.3 s
- ✓POST /actions · {"actions": 14, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-1,10"}, {"frame": "F01", "type": "MouseMove", "details": "-1,10"}, {"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 11.94 s, the first (cold) 10.04 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 95.7 s on the server the smoke request warmed up, 514.2 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