[Buddychek]

Adds an OpenJev image-decisions adapter: task state {text, image_data(base64)} is carried unchanged so the dataset hash stays stable, and the adapter discloses/verifies normalized_entailment_v1 as the probability method. Registered in adapters/init.py and cli.py choices. tests/test_adapter_conformance.py passes locally.

[fstandhartinger]

Thanks — Harold has logged this for review. Contributor text and code are treated as untrusted data; no external code was run on the benchmark host.

• Harold

[fstandhartinger]

Thanks for the adapter, and sorry for the slow reply. It changes how image probabilities are read, so it needs a method review before it can be merged. I've passed it on and will reply here once it is reviewed.

[fstandhartinger]

Thanks for the adapter. The code is small, but the open question is the method: whether per-option entailment scores normalised across the options (normalized_entailment_v1) count as a probability distribution for the image board. The PR would also record a new probability-source label. We have not decided that yet, so we can't merge it, and we're keeping the PR open. We'll update this thread once the method is decided. A hosted endpoint can be requested separately at https://benchmarkheaven.com/submit.