[KartikP]

Add the recorded offline preprocessing reproduction to the input-resolution and input-format evidence for all seven CvT models. This provides a follow-up publication input after #2567 merged without the required independent approval. The diff adds seven lines per file and preserves model facts, confidence, existing source protections and loaders.

Validation passed for schema, table conversion and evidence coverage against the approved core revision. A local first-publication rehearsal and retry preserved all 181,755 score rows, displayed leaderboard ranks and existing legacy metadata. Evidence and validation.

Before merge, another maintainer must review the complete seven metadata documents, approve this exact head and apply metadata-source-override. Use this PR's number for initial dev publication instead of #2567. No shared-dev publication has been made.

[KartikP]

The current follow-up head is d378ba074169fb29a96b38d9a783defa2074a549. The metadata workflow's validation and publication jobs are currently skipped because deployment enablement is off. The offline schema/evidence checks and local publication rehearsal passed; that is not a live source-policy approval.

After independent current-head approval and metadata-source-override, an operator can run the source-policy validator from a trusted environment with core pinned to 30318623bae9846ca9d2dd7ae4573ad8d99ded9f and an authorized token in GITHUB_TOKEN:

python -m brainscore_core.metadata.review \
  --repository brain-score/vision --pr 2571 \
  --head-sha d378ba074169fb29a96b38d9a783defa2074a549 \
  --domain vision --model-root brainscore_vision/models

Revalidate and reapprove if this head changes. After the approved PR merges and dev credentials/cache are configured, publish with --domain vision --pr 2571 --settings=web.metadata_publisher_settings. Keep automatic publication disabled until its deployment setup is ready.