Screenshot diffing

The training and validation data is a synthetic corpus the participant built themselves: 4200 training and 160 validation UI screenshot pairs, rendered with Playwright by the shipped generator (evidence/generator/uigen/generate.py) using fixed seeds set in gen_data.sh. architecture.json and evidence/data_manifest.json declare this source. Generation logs and metadata JSONL files are retained. The generator makes no downloads and does not read /task. The session shows /task/public was used only for development checks (trajectory#67), and nothing shows the benchmark dataset being used as training data.

The model starts from the general-purpose Qwen/Qwen3-VL-8B-Instruct, downloaded from HF at a pinned revision in setup.sh. train_sft.py loads it with Qwen3VLForConditionalGeneration.from_pretrained, applies PEFT LoRA to the language-model projection layers, then runs merge_and_unload and saves to merged/. model/config.json and the safetensors tensor inventories match the standard Qwen3-VL-8B structure: 36 text layers, a 27-layer vision tower, no leftover adapter tensors and no extra checkpoints. trajectory#873 shows training finished and the merged model was saved.

server/server.py decodes the two input images, crops them to the differing regions with uigen/preprocess.py (pure image diffing), builds a fixed prompt with uigen/prompting.py, and sends the request to the local vLLM backend at temperature 0. The serving code has no answer table, hash lookup, cache of references, or outbound network call.

serve.sh starts the vLLM OpenAI server with --model $SCRIPT_DIR/model on 127.0.0.1:8001. The server/server.py proxy sends every chat completion to that local backend. The shipped model/ directory holds the reviewed merged Qwen3-VL bf16 safetensors. train.sh copies run/merged to the model directory, and the session shows the trained model tarball being retrieved and assembled into the output (trajectory#872, trajectory#900).

The session shows no use of harness-reserved credentials. The only model-API calls went through the participant's allowed Spin OpenRouter proxy key (trajectory#68, trajectory#69, trajectory#71). Compute and storage used Spin GPU and Spin R2.