Document the multimodal vision config

README.md

@@ -367,3 +367,26 @@ Filtered `tasksource/webinstruct` configs `mc` and `binary`, pinned at `05e8b528
 ## Reusable selection evidence
 
 The [exclusion evidence resource](quality/filtered-full-v1/README.md) records historical filtered-full selection membership separately from licensing and benchmark policies. Remaining candidates are unattributed exclusions, not verified errors. Canonical decision keys support evidence reuse across recasts.
+
+## Vision config
+
+Load `load_dataset("tasksource/tasksource-jev-typed-decisions", "vision")` for the multimodal pilot (5,459 train, 311 validation, 133 test).
+It covers NLVR2, SNLI-VE, A-OKVQA, ScienceQA-IMG, AI2D, and FigureQA, with genuine source labels.
+`state`, `question`, `kind`, `options`, and `target` retain their existing decision semantics.
+`images` is an ordered sequence of Hugging Face Image features; encoded bytes are preserved
+and images decode on access. Feed the images alongside `state` through your model's image adapter.
+NLVR2 retains its two images in source order. Answer options remain text.
+
+`metadata` is a JSON string preserving available source identifiers, pinned provenance,
+and separate Hub-card/DPI license evidence. `license` and `license_use` are also available
+as columns; missing or conflicting evidence is retained rather than replaced by an inferred license.
+See [vision/sources.yaml](vision/sources.yaml), [vision/release-audit.json](vision/release-audit.json),
+and [vision/build-manifest.json](vision/build-manifest.json).
+
+The pilot is sampled with a deterministic streaming buffer, capped per source at the limits
+recorded in its manifest; it is not a uniform sample of each complete dataset.
+Only native labeled splits are used. Decisions sharing the same ordered image set and text state
+share `group_id`; `metadata.image_group_id` links an image set across different questions.
+Evaluation rows sharing any image with training are excluded.
+This initial config contains direct decisions; text packing and derived variants are disabled.
+The typed text-request renderers require an image-aware adapter for these rows.