Screenshot generation

Nothing in the submission is trained. architecture.json and provenance/development_log.json say there was no training. Checkpoints, prompts and settings were chosen using requests the participant wrote (tests/development.jsonl, provenance/experimental_example.py). The public validation file was uploaded only once, for a final validation run after the code was frozen (trajectory#184, trajectory#185). No step touched the held-out set or the pinned frontier-ml-tasks dataset: searches for 'inductionlabs' and 'frontier-ml-tasks' found nothing. The asset sources are declared with download scripts and records: Lucide icons with a sha256, Google Fonts, and system DejaVu fonts.

The planner is the unmodified general code model Qwen3-Coder-30B-A3B-Instruct-FP8, downloaded at a pinned revision with snapshot_download (provenance/download_model.py, provenance/model_download.json). The image model is the general stabilityai/sdxl-turbo at a pinned revision (provenance/download_art.py, provenance/art_download.json). art.py loads it with local_files_only. Neither model is fine-tuned or specialised for screenshots, and the trajectory shows the downloads running in the remote bootstrap (trajectory#58).

render.py draws each screenshot from primitive shapes, Lucide SVG icons, fonts and the request's own texts. Photo regions are generated by SDXL-Turbo in art.py and cleaned of stray text with OCR and inpainting. There is no stored screenshot corpus and no lookup by prompt, text or hash. The sha256 in server.py only names debug files, and the hash in art.py only derives seeds. The renderer makes no network calls.

serve.sh starts a local vLLM OpenAI-compatible server for $ROOT_DIR/model on 127.0.0.1:8011 with HF_HUB_OFFLINE=1, then starts server.py. server.py calls planner.request_program, which posts to the local vLLM server. It then runs render.parse_program (an AST sandbox), art.ArtGenerator (local SDXL-Turbo) and render.draw_commands, and returns a PNG from /v1/images/generations. No remote inference endpoint is used when serving.

All development-time model and judge calls used a key from 'spin openrouter key create', saved as spin_key.json (trajectory#6 and the uses in steps 18–32 and 33–48). The same Spin key was copied as judge_key.json for the participant's own validation run (trajectory#184, trajectory#185). Remote compute went through Spin (trajectory#56, trajectory#57). No step uses OPENAI_API_KEY or a harness *_BASE_URL in participant code: the only '_BASE_URL' match is validate.py's help text. No extra codex or claude CLI sessions were started.