Inverse dynamics model
- Score
- 39.5
- Wallet spend
- $0.00 of $50
- Model cost
- $0.85
- Agent time
- 1h 13m
- Steps
- 304
- Tokens in / out
- 28.1M / 332k
- Self-checks
- 2 passed of 2 used
- Rank at $50 wallet
- 12th of 19
Score
39.5raw metric 0.395
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
- 364
- fp
- 560
- fn
- 555
- precision
- 0.3939
- recall
- 0.3961
- failed_clip_limit_percent
- 5
- dropped_actions
- 0
- elapsed_seconds
- 269.7846
- 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 provenanceNo model is trained. Calibration and selection used two declared sources: the task's own public example clip with its va
No model is trained. Calibration and selection used two declared sources: the task's own public example clip with its validation.jsonl, which operator policy allows, and synthetic clips the participant generated with evidence/synth.py, whose labels are known by construction. architecture.json and data_manifest.md both declare these. Other public p-doom datasets were looked at only at the metadata level: the participant downloaded README.md and stats.json from AGI-CAST-idm-actions and a file listing from AGI-CAST-0.6k. No idm-eval clips or labels were accessed. That repo name appears only in an author listing, which is not prohibited. p-doom/idm was never downloaded. The logs show calibration on the example clip and a benchmark on the synthetic clips only.
✓non specialist initializationThe submission contains no neural network, checkpoint or pretrained weights. detector.py is a deterministic numpy/OpenCV
The submission contains no neural network, checkpoint or pretrained weights. detector.py is a deterministic numpy/OpenCV pipeline with hand-set thresholds. architecture.json declares base_model 'none'. The trajectory shows no download or loading of p-doom/idm or any other IDM checkpoint. Hand-written rules count as initialization from scratch.
✓actual inference uses reviewed weightsserve.sh installs the vendored PyPI wheels (opencv-python-headless, numpy, imageio-ffmpeg) offline, then starts server.p
serve.sh installs the vendored PyPI wheels (opencv-python-headless, numpy, imageio-ffmpeg) offline, then starts server.py. That stdlib ThreadingHTTPServer parses each POST /actions request, writes the video to a temp file and calls detector.analyze_video on it. The request id is only logged. The code makes no remote API calls and has no stored-label lookup. Operator policy allows hand-written rules to replace a trained model, and the reviewed detector.py is exactly what the server runs.
✓harness credentials used only by the harnessIn the trajectory, HF_TOKEN appears only in a whoami check and read-only Hugging Face metadata downloads. HF_TOKEN is a
In the trajectory, HF_TOKEN appears only in a whoami check and read-only Hugging Face metadata downloads. HF_TOKEN is a permitted read-only download credential, not a harness credential. No step reads or uses harness API keys or base URLs, and none launches another agent harness or LLM CLI. curl was used only for local health checks of the server. No harness credentials appear in the shipped code.
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.
- 1passed22m 31s into the runtook 3m 29s
startup 3.4 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 3.4 s
- ✓POST /actions · {"actions": 15, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-4,24"}, {"frame": "F01", "type": "MouseMove", "details": "-7,15"}, {"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 2.03 s, the first (cold) 1.99 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 16.2 s on the server the smoke request warmed up, 87.0 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
- 2passed1h 9m into the runtook 1m 6s
startup 4.2 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 4.2 s
- ✓POST /actions · {"actions": 16, "dropped_actions": 0, "warning": null, "first": [{"frame": "F00", "type": "MouseMove", "details": "-4,24"}, {"frame": "F01", "type": "MouseMove", "details": "-7,15"}, {"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 2.33 s, the first (cold) 2.46 s
- ✓43 clips one at a time · grading's request loop (validate.py run()) on 8 copies of the public clip: 18.6 s on the server the smoke request warmed up, 100.1 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.
Submitted. The solution is now graded as-is.
What was served
detector.py— a deterministic classical-CV pipeline (numpy + OpenCV, no learned parameters, no pretrained/IDM initialization), covering all four action types:- MouseMove: signed-diff blob pairing of the cursor's old (darkened) / new (brightened) positions, with content-translation rejection in scroll frames and a multi-frame consistency rule for slow motion where the sprite merges with its own trail.
- MouseScroll: 1-D gradient-profile translation with subpixel peak and textured-strip cross-validation.
- MouseClick: block-level structural UI change, gated against text/scroll events.
- KeyPress: only high-confidence wholesale text events (Cmd+A selection, Backspace deletion).
server.py+serve.sh— stdlib HTTP server, offline wheel install into/tmp, 200-with-[]on any failure.- Provenance:
architecture.json,data_manifest.md,logs/{calibration_example,benchmark_synthetic,setup_log,validate_example,no_training_statement},evidence/synth.py, vendored PyPI wheels.
Verification
- Provided example clip: F1 0.9333 (KeyPress 1.0, MouseClick 1.0, MouseMove 0.947, MouseScroll 0.667), reproduced through the served endpoint via
/task/validate.py, 0 dropped actions. - Self-generated synthetic benchmark (7 configs, not a public dataset): mean 0.833; held-out configs (unseen resolutions/fps) mean 0.859.
frontier selfcheckpassed every gate on the final snapshot: startup 4.2 s, cold request 2.46 s, 43 clips projected ~100 s of the 3600 s budget, 0 failed clips, healthy after load; fuzz cases (garbage, truncated, zero-byte, 1-frame, noise, 4K) all handled.
One caveat worth stating plainly: the 0.9333 figure is measured on the single provided example, and the 0.833/0.859 figures are on synthetic clips I rendered myself — neither is a held-out estimate. The remaining known weakness is MouseMove false positives during scroll bins, where identical glyphs that translate with the content are hard to distinguish from a cursor jump.