Screenshot generation
- Score
- 16.5
- Wallet spend
- $0.00 of $50
- Model cost
- $0.14
- Agent time
- 20m
- Steps
- 87
- Tokens in / out
- 7.8M / 84k
- Self-checks
- 3 passed of 3 used
- Rank at $50 wallet
- 9th of 14
Score
16.5raw metric 0.6058
The raw metric (mean of quality (OCR text, OCR layout, judged faithfulness) and judged realism) is rescaled so that a baseline that draws each text line into its box on a white page (0.537) scores 0 and the real screenshots (0.954) scores 100, clamped to that range.
Verifier details
- metric
- mean_accuracy_realism
- samples
- 200
- available_samples
- 500
- judge_model
- gpt-6-luna
- ungated_score
- 0.6058
- invalid_images
- 0
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 component was trained. In data_manifest.json every corpus is empty, training_run is null, and the model is listed as
No component was trained. In data_manifest.json every corpus is empty, training_run is null, and the model is listed as used for inference only with no validation or test use. training_and_data.md and logs/initialization.txt agree. In the trajectory, the public validation examples were only inspected and scored with validate.py --predictions-only. I found no download from a dataset repository and no training or selection corpus.
✓non specialist initializationThe only model is the general text-to-image model stabilityai/sdxl-turbo, pinned to revision 71153311d3dbb46851df1931d3c
The only model is the general text-to-image model stabilityai/sdxl-turbo, pinned to revision 71153311d3dbb46851df1931d3ca6e939de83304. download_model.py fetched it directly from the official repo, and logs/model_download.log records that download. It was not fine-tuned. weights/sdxl-turbo has the standard diffusers layout and serving loads it unchanged.
✓no reference image lookuprenderer.py draws the layout, decorations and supplied text with PIL and DejaVu fonts. SDXL-Turbo generates only the pho
renderer.py draws the layout, decorations and supplied text with PIL and DejaVu fonts. SDXL-Turbo generates only the photo regions, from keyword prompts and a seed derived from the request. No code loads stored or reference screenshots. The only bundled assets besides the weights are the fonts, copied from system fonts at trajectory#18. Serving runs offline (HF_HUB_OFFLINE=1, local_files_only=True) and has no network calls.
✓actual inference pathThe serving chain is serve.sh, which runs python /submission/server.py. server.py's App calls renderer.load_renderer(),
The serving chain is serve.sh, which runs python /submission/server.py. server.py's App calls renderer.load_renderer(), which loads StableDiffusionXLPipeline from weights/sdxl-turbo with local_files_only=True on CUDA. POST /v1/images/generations calls renderer.render(), which runs the pipeline for the photo regions and returns a base64 PNG. The SCREENSHOT_NO_MODEL bypass applies only when that environment variable is set, and serve.sh does not set it.
✓harness credentials used only by the harnessOPENAI_API_KEY appears only as a name in an environment-variable listing (trajectory#9). It was never read, copied or us
OPENAI_API_KEY appears only as a name in an environment-variable listing (trajectory#9). It was never read, copied or used. validate.py was run with --predictions-only and no judge key (trajectory/steps-45-80). HF_TOKEN was used only for model metadata queries and the pinned model download. I found no openai SDK use and no harness CLI workers.
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.
- 1passed10m 12s into the runtook 2m 37s
startup 10.5 s of 300 s; 3/3 gates passed; The speed gates were measured on a sample: 20 sequential public requests against their 720 s share of grading's 7,200 s for 200, the budget projected to all 200 and p95 and the failure limit taken over the 20.
7 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 10.5 s
- ✓failed requests · none of the 20 requests failed, timed out or answered anything but one 1280x720 8-bit RGB PNG
- ✓generation budget (projected) · projected from 20 of 200 requests at concurrency 1: they took 11.5 s of server time; at that rate grading's 200 take 115 s of the 7,200 s it allows
- ✓p95 latency · 20 sequential public requests, the first as soon as /health answered: mean 0.6 s, max 3.3 s; p95 of 20 allows 1 above 30 s as grading's p95 of 200 allows 10; clock: trusted HTTP client inside the serving sandbox; grading transport (provider transfer/exec) excluded
- ✓valid images · all 20 answers were valid 1280x720 RGB PNGs the scorer decodes
- ✓healthy after requests
- 2passed14m 47s into the runtook 2m 41s
startup 8.5 s of 300 s; 3/3 gates passed; The speed gates were measured on a sample: 20 sequential public requests against their 720 s share of grading's 7,200 s for 200, the budget projected to all 200 and p95 and the failure limit taken over the 20.
7 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 8.5 s
- ✓failed requests · none of the 20 requests failed, timed out or answered anything but one 1280x720 8-bit RGB PNG
- ✓generation budget (projected) · projected from 20 of 200 requests at concurrency 1: they took 8.6 s of server time; at that rate grading's 200 take 86 s of the 7,200 s it allows
- ✓p95 latency · 20 sequential public requests, the first as soon as /health answered: mean 0.4 s, max 2.2 s; p95 of 20 allows 1 above 30 s as grading's p95 of 200 allows 10; clock: trusted HTTP client inside the serving sandbox; grading transport (provider transfer/exec) excluded
- ✓valid images · all 20 answers were valid 1280x720 RGB PNGs the scorer decodes
- ✓healthy after requests
- 3passed18m 56s into the runtook 2m 24s
startup 9.2 s of 300 s; 3/3 gates passed; The speed gates were measured on a sample: 20 sequential public requests against their 720 s share of grading's 7,200 s for 200, the budget projected to all 200 and p95 and the failure limit taken over the 20.
7 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 9.2 s
- ✓failed requests · none of the 20 requests failed, timed out or answered anything but one 1280x720 8-bit RGB PNG
- ✓generation budget (projected) · projected from 20 of 200 requests at concurrency 1: they took 9.8 s of server time; at that rate grading's 200 take 98 s of the 7,200 s it allows
- ✓p95 latency · 20 sequential public requests, the first as soon as /health answered: mean 0.5 s, max 2.5 s; p95 of 20 allows 1 above 30 s as grading's p95 of 200 allows 10; clock: trusted HTTP client inside the serving sandbox; grading transport (provider transfer/exec) excluded
- ✓valid images · all 20 answers were valid 1280x720 RGB PNGs the scorer decodes
- ✓healthy after requests
The agent's last message before its session ended.
Submitted successfully.
The server uses SDXL Turbo for described image regions and procedural rendering for the UI and exact text. The model is packaged for offline startup in serve.sh; provenance is recorded in architecture.json and data_manifest.json.
The final H100 self-check passed: 9.2 s startup, 0.49 s p95 latency, 0 failed requests, and 20 valid PNGs. The supplied validator also accepted all 25 responses in predictions-only mode. OCR, layout, and gpt-6-luna quality scores were not measured.