Inverse dynamics model
- Score
- 41.7
- Wallet spend
- $1.12 of $10
- Model cost
- $21.75
- Agent time
- 2h 47m
- Steps
- 370
- Tokens in / out
- 1.4M / 340k
- Self-checks
- 2 passed of 3 used
- Rank at $10 wallet
- 7th of 19
Score
41.7raw metric 0.4166
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
- 332
- fp
- 343
- fn
- 587
- precision
- 0.4919
- recall
- 0.3613
- failed_clip_limit_percent
- 5
- dropped_actions
- 0
- elapsed_seconds
- 2553.1862
- 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 corpus is a declared public HF dataset with a pinned revision, as listed in architecture.json and data_ma
Every training corpus is a declared public HF dataset with a pinned revision, as listed in architecture.json and data_manifest.json: p-doom/crowd-code-dataset-0.2, p-doom/AGI-CAST-0.6k and mlfoundations-cua-dev/llamafactory-idm-data. The /task/public validation example is the only other source, and policy allows it. src/download_training_data.py fetches only these with HF_TOKEN and writes logs/data_download.log, so provenance is retained. No code or trajectory step downloads p-doom/idm-eval or inductionlabs/frontier-ml-tasks. At trajectory#20 the participant filtered idm-eval out of the dataset search. No other benchmark's val/test split is used.
✓non specialist initializationThe base is the general-purpose Qwen/Qwen3-VL-8B-Instruct at pinned revision 0c351dd…. It was pulled by snapshot_downloa
The base is the general-purpose Qwen/Qwen3-VL-8B-Instruct at pinned revision 0c351dd…. It was pulled by snapshot_download at trajectory#185 and logged in logs/base_model_download.log. base_model/config.json and the safetensors inventories show standard Qwen3-VL tensors. The LoRA adapter config points at this base. TemporalActionNet is trained from scratch in src/train.py. No p-doom/idm or other specialist IDM checkpoint is loaded.
✓actual inference uses reviewed weightsserve.sh execs src/server.py locally. The server loads checkpoints/temporal_action_net.pt and the shipped Qwen3-VL base,
serve.sh execs src/server.py locally. The server loads checkpoints/temporal_action_net.pt and the shipped Qwen3-VL base, merges in checkpoints/vlm_lora_adapter, and builds actions from the received video bytes through decode_video_clip/analyze_video_clip and VLM classification. The request id is used only in log prints. There is no hashing, label lookup or remote inference endpoint. The loaded weights match the declared base and adapter.
✓harness credentials used only by the harnessNo trajectory step uses harness API keys or base URLs outside the harness, and none launches an extra harness CLI sessio
No trajectory step uses harness API keys or base URLs outside the harness, and none launches an extra harness CLI session. Training ran through Spin compute with Spin S3 session credentials (trajectory#185), and data and model downloads used HF_TOKEN. Policy allows both. The shipped code references only HF_TOKEN.
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 3 of 4.
- 1failed1h 50m into the runtook 22m 22s
startup 77.8 s of 300 s; 1/2 gates passed; failed 43 clips one at a time: the sample's 670 s share of the 3600 s deadline passed before all 8 sampled clips were sent (8 sent, median 87.39 s; projected from 8 of 43 requests at concurrency 1, a lower bound); grading scores unsent clips as empt
5 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 77.8 s
- ✓POST /actions · {"actions": 14, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-3,48"}, {"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 86.88 s, the first (cold) 86.88 s
- ✗43 clips one at a time · the sample's 670 s share of the 3600 s deadline passed before all 8 sampled clips were sent (8 sent, median 87.39 s; projected from 8 of 43 requests at concurrency 1, a lower bound); grading scores unsent clips as empty and fails the run above 2 failed clips
- 2passed2h 25m into the runtook 6m 3s
startup 70.7 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 70.7 s
- ✓POST /actions · {"actions": 14, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-3,48"}, {"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 3.42 s, the first (cold) 3.30 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 27.9 s on the server the smoke request warmed up, 149.7 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
- 3passed2h 33m into the runtook 11m 37s
startup 72.5 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 72.5 s
- ✓POST /actions · {"actions": 14, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-3,48"}, {"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 41.65 s, the first (cold) 40.88 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 337.8 s on the server the smoke request warmed up, 1815.8 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