Screenshot generation

Nothing in the submission is trained. architecture.json and evidence/initialization.json say there is no training or fine-tuning. Development and selection used only author-written synthetic cases (tools/synthetic_cases.py, evidence/synthetic_cases.jsonl), plus the allowed public /task/public examples for interface and latency checks. Those checks ran with --predictions-only after the code was frozen (evidence/quality_freeze.json). The trajectory shows no download of a benchmark dataset or the held-out split. The other assets come from declared sources: lucide icons from npm, Liberation and DejaVu fonts from Debian, and the PP-OCR ONNX model from ModelScope with a sha256 check. Their download scripts and manifests are kept.

Every model is a frozen general-purpose model used as published. The planner is Qwen3-Coder-30B-A3B-Instruct-FP8; models/coder/config.json shows Qwen3MoeForCausalLM with FP8 quantization. The image model is Z-Image-Turbo, and the text detector is the PaddleOCR PP-OCRv5 detector in ONNX form. tools/download_models.py pins the Qwen and Z-Image revisions, evidence/model_download.log confirms both downloads, and tools/download_text_detector.py checks the detector's sha256. The weight inventories match these models. No screenshot-specialist checkpoint was loaded or trained. The Qwen3-4B model used in early development was replaced and is not shipped.

Each image is built at request time. renderer.py turns the LLM-planned SVG into pixels with resvg, using only whitelisted tags and with external, data: and @import references blocked. It embeds only the photos generated for that request and draws the requested text with the bundled fonts. server.py runs offline with no network calls. evidence/artifact_audit.log reports no stored PNG screenshots, and the trajectory shows no code that copies or retrieves reference images.

The serving chain is serve.sh, which installs the local wheels offline and runs server.py. server.py calls model.py: the Planner loads vLLM on models/coder and the Imagery class loads ZImagePipeline on models/imagery with local_files_only. photo_cleaner.py then runs models/text_detector.onnx through ONNX Runtime with cv2 inpainting, and renderer.py rasterizes the scene and draws the text. All of these weights are in the inventories, and there is no remote service or decoy path. The offline H100 self-check passed, starting in 192.1 s (trajectory#195).

The only credential the participant used was their own Spin OpenRouter key, created with `spin openrouter key create` (trajectory#12) and passed to the synthetic scorer as --judge-key-file. This is allowed. The env listing at trajectory#10 showed variable names only, and no harness key value was used. The harness auth.json was only seen in a directory listing (trajectory#16), never read or used. The participant also scanned the output for leaked hf_ and sk- tokens before submitting (trajectory#187).