Screenshot diffing

training/data_manifest.json declares every training corpus. Each one is original procedural output from shipped generators (generate.py, browser_pairs.py, combine.py), listed with its seed and command, and external_dataset is null for all of them. The corpora are training_corpus with seed 271828 and 1600 examples, plus the stage-2 and stage-3 mixtures built from the varied, browser and desktop sets on their own seeds. In the stage training logs, the example counts and the stage-3 manifest sha match the declared mixtures. Development corpora use disjoint seeds. The public validation file was used only through /task/validate.py and selfcheck. I found no training on it and no download of any external or held-out dataset.

architecture.json declares a general-purpose VLM as the base: Qwen/Qwen3-VL-8B-Instruct, pinned to revision 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b. The trajectory shows it fetched with training/download_model.py through snapshot_download, using that repo and revision. In training/train.py and training/train_stage2.py, Qwen3VLForConditionalGeneration is loaded from the base directory and LoRA adapters are applied (r16, alpha 32). The adapters are chained stage1→stage2→stage3 and then merged. The initialization events in the training logs show base /workspace/base_model with these adapters. model/config.json (36 layers, hidden 4096, vision depth 27) and the tensor inventory match standard Qwen3-VL-8B. There is no specialist or task-tuned checkpoint.

server.py decodes only the two base64 images in the request. It runs the deterministic pixel-diff views from vision.py (region_boxes, make_views) and builds static prompts with messages_for. It then generates with temperature 0 via vLLM. I found no answer tables, caches, stored descriptions, retrieval or network access. serve.sh sets HF_HUB_OFFLINE and TRANSFORMERS_OFFLINE and installs only local wheels.

serve.sh runs `exec python server.py`. server.py loads AutoProcessor and a vLLM AsyncLLM, with dynamic FP8 quantization, from CHECKPOINT, which defaults to ROOT/model. The model/ directory holds the merged stage-3 weights: 5 safetensors shards and 750 tensors, consistent with evidence/model_manifest.json and the reviewed weight inventory. The trajectory shows the merged stage-3 output moved into output/model. The trusted binary audits show every wheel matches an approved upstream PyPI digest, so I found no altered loading code.

The only model-API credential the participant used was a scoped OpenRouter key they created with `spin openrouter key create`, which the operator permits. training/judge_observations.py and training/judge_synthetic.py used it to run the Gemini judge on original synthetic development data only. I found no use of harness-reserved credentials such as OPENAI_API_KEY or the harness base URL, and no spawned agent sub-sessions. The BASE_URL matches were only help text.