One gallery, any encoder

Four multimodal models from four vendors encoded the same 5,000 COCO images. This Space searches one model's gallery using a different model's text encoder, fitting the towers live on CPU from the published states.

Across four vendors, cross-vendor retrieval is statistically indistinguishable from within-vendor retrieval: retention 0.988, 95% CI [0.955, 1.023]. The interval contains 1.0, so the claim is indistinguishability rather than a specific retained fraction. Four of the twelve cross directions score above the within-vendor baseline of the gallery they search.

Hosts: Qwen3-Omni-30B-A3B, gemma-4-31B-it, Mistral-Small-3.1-24B, Aria.

Read this before trusting a number here

Centering is not optional. Raw cosine between unrelated items on these states is +0.869 and raw retrieval sits exactly at chance. Everything shown is per-modality centered.

The Space fits its towers with a plain numpy loop so it can run on a free CPU box. That scores below the torch fit behind the published numbers, so read the cross-versus-within comparison rather than the absolute r@1.

Scope

Images only: two of the four hosts carry no audio or video tower, so the audio evidence elsewhere in this project is single-vendor and video is two-vendor. All four hosts train on overlapping web-scale corpora, so this shows the readable structure survives a change of vendor, architecture and training run. It does not show that independent minds would converge on it, and that reading is not asserted.

• States: RiverRider/srt-omni-crossvendor-states
• Towers: RiverRider/srt-omni-xvendor-towers
• Shared omni tower: RiverRider/srt-omni-shared-tower
• Add your own model: RiverRider/srt-omni-manifest