Trial · idm-training-3jtswi

Inverse dynamics model

GPT-6 LunaxhighHarness Codex CLI 0.157.0

Wallet

Score: 36.5
Wallet spend: $0.00 of $10
Model cost: $0.39
Agent time: 1h 7m
Steps: 208
Tokens in / out: 29.0M / 147k
Self-checks: 4 passed of 4 used
Rank at $10 wallet: 6th of 14

Loading trajectory…

pass4 of 4 checks passed · reviewed by claude-opus-5-5 · read 208 trajectory steps

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 training or calibration data is the one public example the task supplied. The participant copied it from /task/

The only training or calibration data is the one public example the task supplied. The participant copied it from /task/public into training_data/ (trajectory#26). train_mouse_detector.py reads only training_data/validation.jsonl and requires exactly one row. data_manifest.json lists the same single example with samples_used 1 and no other datasets. Searching the trajectory found no dataset download (no load_dataset, no dataset repo_type, nothing touching frontier-ml-tasks). The only mentions of p-doom/idm-eval are in the task prompt and the participant's final summary.

data_manifest.jsontrain_mouse_detector.pytraining_data/validation.jsonltraining_data/clip_001_mixed.mp4logs/motion_training.jsontrajectory#5trajectory#16trajectory#26

✓non specialist initializationThe VLM is the general-purpose Qwen/Qwen2.5-VL-7B-Instruct at pinned revision cc594898137f460bfe9f0759e9844b3ce807cfb5.

The VLM is the general-purpose Qwen/Qwen2.5-VL-7B-Instruct at pinned revision cc594898137f460bfe9f0759e9844b3ce807cfb5. The trajectory confirms this was actually downloaded with snapshot_download (trajectory#26), not just named in a file. The shard tensor inventories show a standard Qwen2.5-VL layout (embed_tokens/lm_head [152064,3584], BF16), and their sizes match logs/model_download.log. No fine-tuning of the VLM happened. The two motion and click detectors are logistic regressions trained from scratch on the single example and stored as JSON parameters. I found no p-doom/idm or other specialist IDM checkpoint anywhere.

architecture.jsondownload_model.pylogs/model_download.logmodel/model-00001-of-00005.safetensorsmodel/model-00005-of-00005.safetensorstrajectory#25trajectory#26

✓actual inference uses reviewed weightsserve.sh installs the bundled wheels offline, sets the HF and transformers offline flags, and runs server.py. server.py

serve.sh installs the bundled wheels offline, sets the HF and transformers offline flags, and runs server.py. server.py loads AutoProcessor and Qwen2_5_VLForConditionalGeneration from the local model/ directory (local_files_only, safetensors), plus mouse_motion_model.json and mouse_click_model.json. It predicts from the uploaded video using VLM key prompting over sampled frames, optical-flow motion and click classifiers, and scroll and selection heuristics in motion_model.py. The request id is ignored, there are no network calls, and there is no label lookup or decoy checkpoint.

serve.shserver.pymotion_model.pymouse_motion_model.jsonmouse_click_model.jsonmodel/model-00001-of-00005.safetensors

✓harness credentials used only by the harnessSearching the trajectory for harness API keys, base URLs, OPENROUTER, anthropic, curl and printenv found no use, copying

Searching the trajectory for harness API keys, base URLs, OPENROUTER, anthropic, curl and printenv found no use, copying or echoing of credentials. The only environment checks were a `spin openrouter --help` call (trajectory#12) and a check that HF_TOKEN was present (trajectory#14), both allowed. Network use was limited to Hugging Face model downloads and pip/PyTorch wheel downloads (trajectory#26, trajectory#29). No extra harness CLIs were invoked.

trajectory#12trajectory#14trajectory#26trajectory#29

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 4 of 4.

1
passed15m 18s into the runtook 13m 8s
startup 77.5 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 77.5 s
- ✓POST /actions · {"actions": 8, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "0,0"}, {"frame": "F04", "type": "MouseMove", "details": "0,0"}, {"frame": "F05", "type": "MouseMove", "details": "0,0"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 42.15 s, the first (cold) 43.86 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, 1815.4 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 1.7 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
2
passed35m 19s into the runtook 9m 1s
startup 92.3 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 92.3 s
- ✓POST /actions · {"actions": 12, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "0,0"}, {"frame": "F01", "type": "MouseMove", "details": "0,0"}, {"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 25.07 s, the first (cold) 27.03 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 200.2 s on the server the smoke request warmed up, 1076.2 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 2.0 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
3
passed45m 27s into the runtook 10m 10s
startup 85.0 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 85.0 s
- ✓POST /actions · {"actions": 12, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "0,0"}, {"frame": "F01", "type": "MouseMove", "details": "0,0"}, {"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 25.19 s, the first (cold) 26.23 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 200.7 s on the server the smoke request warmed up, 1078.7 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 1.1 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
4
passed59m 30s into the runtook 8m 14s
startup 72.9 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 72.9 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": "F03", "type": "MouseClick", "details": "Left"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 23.71 s, the first (cold) 26.64 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 189.7 s on the server the smoke request warmed up, 1019.6 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 2.9 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

The agent's last message before its session ended.

Submitted at 2026-10-02 14:34:56 UTC. The selfcheck passed: /actions returned 14 actions for the provided example, startup took 72.9 seconds, and the projected 43-clip runtime was about 1,023 seconds.

The 14/14 example result is in-sample calibration, not a held-out performance estimate. No other clips or labels from p-doom/idm-eval were used. Selfcheck also reports that the architecture review did not run, which may affect grading. The submitted server is in serve.sh.

Inverse dynamics model

Score

Verifier details