Inverse dynamics model
- Score
- 45.0
- Wallet spend
- $14.39 of $200
- Model cost
- $13.46
- Agent time
- 4h 22m
- Steps
- 357
- Tokens in / out
- 31.9M / 220k
- Self-checks
- 3 passed of 3 used
- Rank at $200 wallet
- 8th of 19
Score
45.0raw metric 0.4497
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
- 465
- fp
- 684
- fn
- 454
- precision
- 0.4047
- recall
- 0.5060
- failed_clip_limit_percent
- 5
- dropped_actions
- 0
- elapsed_seconds
- 626.4261
- 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 corpus is declared in data_manifest.json and consists of three parts. (1) OpenCUA (cua-lite/OpenCUA) deskto
The training corpus is declared in data_manifest.json and consists of three parts. (1) OpenCUA (cua-lite/OpenCUA) desktop/use/train parquet shards only. In the session (trajectory#23–29) they were downloaded from resolve/main, and evidence/corpus/reproduce_download.py pins revision f897d4f…. All eight paths are train shards, and no validation or test split was used. (2) Procedural clips that evidence/training/synth_data.py generated from those shards. (3) Augmentations of the single permitted example in /task/public, made by evidence/training/adapt_example.py with 120 holdout clips excluded. There is no access to p-doom/idm-eval or to the private inductionlabs data. The only matches for 'idm-eval' are the task prompt and exclusion text.
✓non specialist initializationThe base model is the general-purpose Qwen/Qwen3-VL-8B-Instruct at the pinned revision 0c351dd…. It was downloaded with
The base model is the general-purpose Qwen/Qwen3-VL-8B-Instruct at the pinned revision 0c351dd…. It was downloaded with huggingface-cli at trajectory#23 and snapshot-downloaded into base/ at the same revision. adapter/adapter_config.json describes a LoRA (r=32) on the language layers with visual modules excluded, and its base_model_name_or_path is the Qwen download directory. train_lora.py trains from that base, and v2 resumes the participant's own v1 adapter. The Qwen2.5-VL download was never used. No p-doom/idm or other specialist IDM appears anywhere.
✓actual inference uses reviewed weightsserve.sh runs serve.py offline. serve.py loads AutoModelForImageTextToText from ROOT/base and wraps it with PeftModel fr
serve.sh runs serve.py offline. serve.py loads AutoModelForImageTextToText from ROOT/base and wraps it with PeftModel from ROOT/adapter. The reviewed safetensors are inventoried as LoRA tensors for the language layers plus four base shards. Inference does the following: decodes the received video (model_io.py), builds the zoom view from pixels, runs greedy generation, parses the TSV output, and applies postprocess.py. The postprocessing only applies the configured MouseClick -1 frame shift. The request id is only logged; there is no label lookup, hashing or remote inference call. The shipped adapter was copied from adapter_v2 with checksums (trajectory steps 300–313).
✓harness credentials used only by the harnessSearches of the trajectory and code for ANTHROPIC, BASE_URL, OPENROUTER, API_KEY and 'claude -p' found no participant us
Searches of the trajectory and code for ANTHROPIC, BASE_URL, OPENROUTER, API_KEY and 'claude -p' found no participant use of harness credentials and no extra harness CLI launches. The only credentials used were the Spin wallet S3 transfer session, through spin/boto3 in s3transfer.py (trajectory#27–29), and HF downloads (trajectory#23). Both are explicitly allowed.
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.
- 1passed2h 52m into the runtook 8m 24s
startup 73.1 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 73.1 s
- ✓POST /actions · {"actions": 5, "dropped_actions": 0, "warning": null, "first": [{"frame": "F04", "type": "KeyPress", "details": "Ctrl+V"}, {"frame": "F08", "type": "KeyPress", "details": "Return"}, {"frame": "F10", "type": "KeyPress", "details": "Ctrl+A"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 5.85 s, the first (cold) 6.00 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 46.8 s on the server the smoke request warmed up, 251.8 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 0.1 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
- 2passed3h 26m into the runtook 6m 14s
startup 61.6 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 61.6 s
- ✓POST /actions · {"actions": 14, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "1,0"}, {"frame": "F01", "type": "MouseMove", "details": "1,0"}, {"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 9.98 s, the first (cold) 10.74 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 79.3 s on the server the smoke request warmed up, 426.4 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 0.9 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
- 3passed4h 13m into the runtook 7m 39s
startup 59.6 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 59.6 s
- ✓POST /actions · {"actions": 14, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "1,0"}, {"frame": "F01", "type": "MouseMove", "details": "1,0"}, {"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 12.64 s, the first (cold) 12.70 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 100.8 s on the server the smoke request warmed up, 541.6 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 0.1 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
The agent's last message before its session ended.
Submission review shows the expected final folder: 210 files, 22.13 GB. The final model, calibration, validation, and selfcheck are in place; I’m confirming submission now.