Screenshot diffing

Every training corpus has a declared public source: - GUI-World (ONE-Lab/GUI-World): only the Annotation/train/{software,website}.jsonl files and the train videos they reference were downloaded. The Annotation/benchmark files were listed while exploring the repo but were never used. - showlab/ShowUI-desktop: downloaded with snapshot_download. - Synthetic edits: generated by synth_edits.py, with labels emitted from the edit parameters. - Teacher labels: from qwen/qwen3-vl-235b-a22b-instruct via the participant's Spin openrouter key. The policy allows this. The download commands are kept in download_guiworld.py, download_all.sh and data_manifest.json. The self-built eval set is split by video id from the train data. The task's single public validation pair was used only as a diagnostic score; nothing was trained on it. There is no access to the pinned inductionlabs/frontier-ml-tasks repository and no use of the held-out data.

The base model is the general VLM Qwen/Qwen3-VL-8B-Instruct, downloaded with snapshot_download; the participant verified it against revision 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b. train.py loads that base and applies LoRA (r=64). Stage 2 continues from the stage-1 adapter (INIT_FROM=ckpt/run1/lora). merge_lora.py merges the adapter into the same base. The shipped weight inventory has the Qwen3-VL-8B layout (hidden size 4096, 151936 vocab, BF16), and config.json reports Qwen3VLForConditionalGeneration.

server_wrap.py only starts vLLM on the shipped model directory, warms it up with synthetic PIL-drawn images, and proxies requests to the internal vLLM port, clamping max_tokens. There is no lookup table, no hashing of inputs, no stored reference descriptions and no external network calls at serving time. serve.sh runs fully offline (HF_HUB_OFFLINE, TRANSFORMERS_OFFLINE, pip --no-index).

serve.sh installs the vendored wheels offline and execs server_wrap.py, which runs `vllm serve $SUB/model` with fp8 dynamic quantization. model/ holds the four reviewed safetensors shards, and their inventory matches the merged Qwen3-VL-8B. The trajectory shows these merged weights being staged, uploaded and copied into /workspace/output/model (steps 308-311), and the selfcheck passed using this stack.

The teacher and judge calls used a scoped key the participant created with `spin openrouter key create` (trajectory steps 12 and 14), stored in spin_key.txt, with the Spin base URL. The policy allows Spin openrouter calls. HF_TOKEN was copied to the GPU box only for Hugging Face downloads, which the policy allows. Searches found no OPENROUTER_API_KEY or other harness credentials, and the participant launched no extra harness CLIs (claude, codex, etc.).