Add JEVision model card

README.md

@@ -0,0 +1,88 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen3.5-0.8B-Base
+base_model_relation: adapter
+language:
+  - en
+tags:
+  - multimodal
+  - vision
+  - decision-model
+  - long-context
+  - qwen3.5
+  - lora
+  - typesafe
+---
+
+# JEVision
+
+### Give the KEV-style decision interface a pair of eyes.
+
+**Images in. Structured decisions out. A 77K-token proof point.**
+
+JEVision is an experimental multimodal extension of the KEV-style decision workflow. It brings image input into a typed, machine-readable decision path built on Qwen3.5-0.8B-Base. Instead of asking for an open-ended paragraph, provide a state, an image, and a concrete question—and get a structured answer through the familiar **/v1/systemone** route.
+
+## Why JEVision
+
+- **Vision meets decisions.** Image and text go through the same KEV-shaped request/response path.
+- **Built for answers software can use.** The tested image path returns a structured Choice result, not free-form chat.
+- **Long context, demonstrated.** Image-plus-text requests passed at **76,998 tokens** in the POC.
+- **A practical first step.** This release publishes the visual adapter and pointer head needed to extend the public KEV text checkpoint.
+
+## What we verified
+
+The milestone tests exercised the real **/v1/systemone** handler, not a model-only shortcut:
+
+| Check | Result |
+|---|---|
+| Text-only request | 76,999 input tokens; accepted with exact preflight token accounting |
+| Image + text request | 76,998 input tokens; accepted with the same KEV response envelope |
+| Image-content smoke | At 77,003 tokens, swapping only a generated red image for a generated blue image flipped the selected answer from red to blue |
+| Oversized request | Inputs above the 80,000-token service cap were rejected rather than silently truncated |
+
+That is an end-to-end proof that the image path can affect a structured decision at long context. It is an intentionally small functional smoke test—not a claim of broad visual understanding.
+
+## How the package fits together
+
+JEVision is a **two-checkpoint composition**, not a merged standalone model:
+
+1. **Visual base:** [Qwen3.5-0.8B-Base](https://huggingface.co/Qwen/Qwen3.5-0.8B-Base), pinned to revision `dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68`.
+2. **KEV text checkpoint:** [jaredpalmer/kev-0.8b](https://huggingface.co/jaredpalmer/kev-0.8b/tree/54f4f8777356cd5bbbb6c6919c657f26e6f2f6d8), pinned to revision `54f4f8777356cd5bbbb6c6919c657f26e6f2f6d8`.
+3. **This repository:** a visual LoRA adapter plus a small pointer head.
+4. **Serving layer:** compatible KEV runtime composes the checkpoints and exposes **/v1/systemone**.
+
+The base-model weights and KEV text checkpoint are separate downloads; they are not duplicated here. The manifest records the pinned components and artifact hashes.
+
+## Request shape
+
+With the compatible multimodal KEV runtime running, an image request uses a base64 data URL:
+
+    POST /v1/systemone
+    {
+      "model": "kev-latest",
+      "state": "Which option matches the attached image?",
+      "images": [{"data_url": "data:image/png;base64,<BASE64_PNG>"}],
+      "questions": {
+        "answer": {
+          "type": "choice",
+          "instructions": "Which color is shown?",
+          "criteria": {"red": "The image is red", "blue": "The image is blue"}
+        }
+      }
+    }
+
+## What this release is—and is not
+
+The Hub repo contains the visual sidecar artifacts. It is **not** a full Qwen checkpoint, a merged model, or a drop-in **AutoModel.from_pretrained** package. Running the POC requires the matching image-enabled KEV serving code as well as the two referenced checkpoints; turnkey public runtime packaging is a separate step.
+
+The current evidence is deliberately narrow:
+
+- The 80,000-token service cap was tested near 77K; **128K context is not claimed**.
+- Image sensitivity was checked with two generated 32×32 solid-color images; **real-photo accuracy and broad image understanding have not been measured**.
+- **No JevBench comparison, accuracy parity, latency advantage, or cost advantage is claimed yet.**
+
+The next evaluation milestone is a JevBench comparison. Accuracy work and inference optimization follow after that baseline.
+
+## License
+
+Apache-2.0. The referenced Qwen base is also listed under Apache-2.0.