Inverse dynamics model
- Score
- 35.0
- Wallet spend
- $5.40 of $10
- Model cost
- $20.72
- Agent time
- 5h 14m
- Steps
- 502
- Tokens in / out
- 48.1M / 368k
- Self-checks
- 2 passed of 2 used
- Rank at $10 wallet
- 11th of 19
Score
35.0raw metric 0.3501
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
- 268
- fp
- 344
- fn
- 651
- precision
- 0.4379
- recall
- 0.2916
- failed_clip_limit_percent
- 5
- dropped_actions
- 0
- elapsed_seconds
- 172.3769
- 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 provenancearchitecture.json and PROVENANCE.md list three public training corpora, each pinned to a revision: p-doom/AGI-CAST-0.6k
architecture.json and PROVENANCE.md list three public training corpora, each pinned to a revision: p-doom/AGI-CAST-0.6k (CC0 screencasts), p-doom/AGI-CAST-idm-actions (weak labels) and the PSAI train split only. The final run used 2048 windows (1536 PSAI and 512 AGI-CAST). Model selection used the /task/public example plus 64 held-back PSAI train windows. No validation or test split was used. download_train_data.py downloads only from an allowlist with pinned revisions and contains an explicit assert that excludes p-doom/idm-eval. The download receipts are kept in the evidence manifests. In the trajectory, 'idm-eval' appears only in instructions and in exclusion asserts or prompts that forbid it. There is no download of idm-eval and no access to inductionlabs/frontier-ml-tasks. A subagent read about other datasets, such as originlab, during research, but none of them was used.
✓non specialist initializationTraining starts from the general VLM Qwen/Qwen2.5-VL-3B-Instruct at revision 66285546 (evidence/base_info.json, architec
Training starts from the general VLM Qwen/Qwen2.5-VL-3B-Instruct at revision 66285546 (evidence/base_info.json, architecture.json). logs/train.log shows the two base shards loading, a LoRA with 61.7M of 3.82B parameters trainable, 256 steps, and then a merge into the general base. model/merge_metadata.json records the base revision and the selected adapter checkpoint-256. The safetensors inventory shapes (embedding 151936x2048, MLP 11008) match Qwen2.5-VL-3B. The AGI-CAST labels were generated by p-doom/idm but were used only as weak training targets. No p-doom/idm or other specialist IDM weights were loaded as initialization.
✓actual inference uses reviewed weightsserve.sh installs from local wheels offline and runs `python -I serve.py`. serve.py's /actions handler drops the clip id
serve.sh installs from local wheels offline and runs `python -I serve.py`. serve.py's /actions handler drops the clip id, writes the uploaded video to a temp file and calls ActionModel.predict. model_engine.py decodes the frames at 5 fps, loads the local model/ directory with Qwen2_5_VLForConditionalGeneration (local_files_only) and runs greedy generation over frame windows. Its output is combined with the generic classical-CV rules in cv_postprocess.py (connected components for cursor moves, phase correlation for scrolls). There is no remote service, no label lookup by id, order or hash, and no hidden IDM backbone.
✓harness credentials used only by the harnessIn the trajectory, no step reads, prints, copies or uses ANTHROPIC_API_KEY or any *_BASE_URL harness credential, and no
In the trajectory, no step reads, prints, copies or uses ANTHROPIC_API_KEY or any *_BASE_URL harness credential, and no extra harness CLI (claude -p, codex, etc.) is launched. The only external model tool used is the permitted spin openrouter CLI (trajectory#17). The Agent tool calls are the harness's own subagents. None of the shipped code references 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 2 of 4.
- 1passed4h 55m into the runtook 4m 11s
startup 61.1 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 61.1 s
- ✓POST /actions · {"actions": 9, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "0,0"}, {"frame": "F01", "type": "MouseMove", "details": "0,0"}, {"frame": "F04", "type": "MouseMove", "details": "0,0"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 3.38 s, the first (cold) 4.18 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 34.3 s on the server the smoke request warmed up, 184.6 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 0.8 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
- 2passed5h 10m into the runtook 3m 46s
startup 60.7 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 60.7 s
- ✓POST /actions · {"actions": 9, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "0,0"}, {"frame": "F01", "type": "MouseMove", "details": "0,0"}, {"frame": "F04", "type": "MouseMove", "details": "0,0"}]}
- ✓slowest request · 9 clips one at a time, each from its first attempt to its last (retries included), median 2.40 s, the first (cold) 3.35 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 19.2 s on the server the smoke request warmed up, 103.1 s for 43 (projected from 8 of 43 requests at concurrency 1), plus 1.0 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.
The exact final folder passed the H100 self-check: startup 60.7 s, slowest request 3.35 s, projected 43-clip batch 104.0 s, and zero failed clips. I’m handing in the audited artifact now.