News: add Jev-Style v3 (0.8B full fine-tune, 25,600-token input)

Adds one entry to the `trained` section of `ITEMS` in `news.html`: **Jev-Style-0.8B-Decision-v3**, a full fine-tune of Qwen3.5-0.8B (Apache-2.0) released 2026-09-24.

- One verdict slot per option (no letter cap; 77 options scored in one pass), up to 25,600 input tokens, many choice / score / noul questions per state in one call.
- Builds: transformers, GGUF (Q4_K_M 0.53 GB) and MLX; a ZeroGPU Space.
- Links: model, GGUF repo, Space, GitHub mirror, the Hub post and jevstyle.com. There is no post on X, so `by.handle` is null.
- Metrics as of 2026-09-25: 0 GitHub stars, 0 likes on the main Hub repo.
- The one number in the description (79.2% vs Laya typed 76.6% on the 2,000 typed decisions) is from the model card and says that both were trained on that dataset's train split.

Only this entry is added; nothing else in the file changes (checked by parsing `ITEMS` with node: 64 entries). Not on the Decision Index yet. Happy to change the wording or drop fields.

Not affiliated with TypeSafe AI.

news.html

@@ -434,6 +434,11 @@ const ITEMS = [
     m: { stars: 0 }, date: '2026-09-22', tags: 'jobe frozen qwen3.5 4b letter logits semif readout no training rtx 3080 local decision index 0.2' },
 
   /* ---------------- trained ---------------- */
+  { cat: 'trained', title: 'Jev-Style v3: a 0.8B full fine-tune with a 25,600-token input', by: { name: 'Chaoliang Yan', handle: null },
+    desc: 'A full fine-tune of Qwen3.5-0.8B that scores one verdict slot per option, so there is no letter cap (77 options in one pass). Takes up to 25,600 tokens of input and answers many choice / score / noul questions about one state in a single call. Ships as transformers, GGUF (0.53 GB Q4_K_M) and MLX, with a ZeroGPU Space. On the 2,000 typed decisions: 79.2% vs Laya typed 76.6% (both trained on that train split).',
+    quote: 'Give it any text and a question. It returns a calibrated probability for every option in one forward pass.',
+    links: { hf: 'https://huggingface.co/chaoliangUNSW/Jev-Style-0.8B-Decision-v3', hf2: 'https://huggingface.co/chaoliangUNSW/Jev-Style-0.8B-Decision-v3-GGUF', space: 'https://huggingface.co/spaces/chaoliangUNSW/jev-style-v3', gh: 'https://github.com/lawrence3699/Jev-Style-0.8B-Decision-v3', blog: 'https://huggingface.co/posts/chaoliangUNSW/170962406339651', site: 'https://jevstyle.com' },
+    m: { stars: 0, hfLikes: 0 }, date: '2026-09-24', tags: 'qwen3.5 0.8b full fine-tune verdict slot no letter cap 25k long context long document multilingual chinese 51 languages gguf llama.cpp mlx apple silicon zerogpu space calibrated apache-2.0 jevstyle' },
   { cat: 'trained', title: 'Verdict: calibrated bi-encoder with a conformal abstain set', by: { name: 'Manav Arya Singh', handle: 'Manavarya09' },
     desc: 'A 118M multilingual bi-encoder fine-tuned on a typed-decision mix. Answers choice, score and noul over the /v1/systemone wire format with temperature-scaled probabilities and a conformal abstain set (coverage guarantee). Fits on your own labels in seconds on a CPU; runs in the browser via ONNX. README lists the benchmarks it loses too.',
     quote: 'Every answer carries a probability that is honest and an abstain flag with a coverage guarantee.',