PROTOCOL.md

Frozen model protocol and environment

protocol/prompt.json holds the exact user prompt template and requested JSON schema. Present one image and the unchanged source question independently; group states only after inference. All six reported configurations used temperature 0 and a 512-token output cap through vLLM-based serving. The saved analysis treats them as serving configurations; it does not equate API and local checkpoint revisions that lack matching provenance.

Scoring key	Paper name	Model identifier
qwen38-flash-next	Qwen 3.8 Flash Next	Qwen/Qwen3.8-Flash-Next
gemma4-31b-it	Gemma 4 31B	google/gemma-4-31B-it
qwen	Qwen 3.8 27B	Qwen/Qwen3.8-27B-FP8
gemma4-26b-a4b-it	Gemma 4 26B A4B	google/gemma-4-26B-A4B-it
glm	GLM 5.3 Flash	GLM-5.3-Flash
molmo2-8b	Molmo2-8B	allenai/Molmo2-8B

In inputs/crossdomain/, qwen is named qwen3.8-27b-fp8 and glm is named GLM-5.3-Flash. Each configuration contributes 5,000 chart and 7,000 scene responses. The 11 terminal invalids comprise eight chart outputs from Molmo2 and one CLEVR output each from Qwen Next, Qwen 27B and Gemma 26B. All GQA final responses are schema-valid.

Qwen 27B scene outputs combine 2,425 preserved bridge responses and 4,575 local FP8 fallback responses. The local fallback's model revision and serving settings are in inputs/crossdomain/qwen-backend-provenance.json. The bridge checkpoint revision is unknown. Mixed-backend question groups are retained and reported; do not interpret a backend difference as a causal model/domain effect.

The verified CPU reproduction environment is pinned in requirements.txt and protocol/verified-runtime.json. It reproduces statistics and validates the packaged evidence; it is not claimed to be every model's inference environment. Available historical inference revisions/versions and decoding settings are recorded in INFERENCE_ENVIRONMENTS.json; unknown values stay unknown. No model weights, API credentials or generation traces are included, and preparing this release ran no new inference.

For future GPU replication, start with a small preflight and record GPU memory, utilization and throughput. The original local runners used vLLM where supported, gpu_memory_utilization=0.90, greedy decoding, and model-native image processing. Batching/parallelism and backend-specific compatibility affect the runtime; the frozen results can be reproduced from the supplied scoring records without access to any original endpoint.

Strict scene answers normalize rational numbers, Unicode NFKC, case and spaces only. Invalid outputs fail full-denominator scores and all-sibling group success. All post-hoc strata, class-balanced metrics and bootstrap settings retain their descriptive status. The finite-program construction records are independent of the six-configuration model response denominator.