Release Jev-Style-0.8B-Decision-v3

.gitattributes

@@ -33,3 +33,14 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+figures/beyond_laya.png filter=lfs diff=lfs merge=lfs -text
+figures/calibration.png filter=lfs diff=lfs merge=lfs -text
+figures/design_table.png filter=lfs diff=lfs merge=lfs -text
+figures/headline_typed.png filter=lfs diff=lfs merge=lfs -text
+figures/jevbench.png filter=lfs diff=lfs merge=lfs -text
+figures/latency.png filter=lfs diff=lfs merge=lfs -text
+figures/long_context.png filter=lfs diff=lfs merge=lfs -text
+figures/multilingual.png filter=lfs diff=lfs merge=lfs -text
+figures/quantization.png filter=lfs diff=lfs merge=lfs -text
+figures/zeroshot.png filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Alibaba Cloud
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

NOTICE

@@ -0,0 +1,27 @@
+chaoliangUNSW/Jev-Style-0.8B-Decision-v3
+Copyright 2026 chaoliangUNSW. Licensed under the Apache License, Version 2.0 (see LICENSE).
+
+This model is a fine-tuned derivative of Qwen3.5-0.8B (https://huggingface.co/Qwen/Qwen3.5-0.8B,
+revision 2fc06364715b967f1860aea9cf38778875588b17), Copyright 2026 Alibaba Cloud, licensed under the Apache
+License, Version 2.0. The LICENSE file in this repository is the license file distributed with Qwen3.5-0.8B.
+
+Modifications relative to Qwen3.5-0.8B:
+- all text-model weights were fine-tuned (full fine-tuning, bf16 training) to score typed decision questions
+  (choice / score / true-false) with a verdict readout: logit(" yes") - logit(" no") at one " ->" slot per
+  option; calibration temperatures were fitted afterwards (readout_config.json);
+- the vision tower (model.visual.*) and the multi-token-prediction head (mtp.*) were removed; the checkpoint
+  is a text-only Qwen3_5ForCausalLM with tied input/output embeddings;
+- added the runtime script, readout/release configuration files, the integrity manifest and this NOTICE.
+
+The question types (choice / score / noul) follow the typed-decision convention of Laya
+(https://github.com/NandhaKishorM/laya, Apache-2.0) so both models can be evaluated on the same
+inputs. No Laya code or weights are included.
+
+Third generation (v3) of the Jev-Style decision series. Earlier generations: v1 =
+chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision (public GGUF release: chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF) and
+v2 = chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2, both 2B models built on Qwen3.5-2B-Base. v3 is fine-tuned from
+Qwen3.5-0.8B; no weights of v1 or v2 were reused.
+
+Not affiliated with, endorsed by or connected to TypeSafe or Jev. "Jev-Style" only describes the kind
+of model (a small typed-decision model in a similar style); no Jev weights, code or outputs are included.
+Not affiliated with or endorsed by Alibaba Cloud / the Qwen team or the Laya authors.

README.md

@@ -0,0 +1,448 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen3.5-0.8B
+base_model_relation: finetune
+library_name: transformers
+pipeline_tag: text-classification
+language:
+  - en
+  - zh
+  - ar
+  - bg
+  - de
+  - el
+  - es
+  - fr
+  - hi
+  - ja
+  - ko
+  - pt
+  - ru
+  - sw
+  - ta
+  - th
+  - tr
+  - ur
+  - vi
+tags:
+  - decision-model
+  - jev-style
+  - system-one
+  - calibration
+  - classification
+  - long-context
+  - multilingual
+  - qwen3.5
+---
+
+# Jev-Style-0.8B-Decision-v3
+
+**Jev-Style decision series:** [v1 · 2B](https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF) → [v2 · 2B](https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2) → **v3 · 0.8B (this model)** · **Website:** [jevstyle.com](https://jevstyle.com)
+
+**One state. One pass. Every option scored.**
+
+Jev-Style v3 is a 0.8B decision model. It takes inputs of up to 25,600 tokens, reads the state once and returns a calibrated
+probability for every option of every question you ask about it. There is no letter cap on options, and it was
+evaluated in 51 languages.
+
+| **0.8B** | **25,600 tokens** | **77 options** | **51 languages** | **0.53 GB** |
+|:---:|:---:|:---:|:---:|:---:|
+| parameters, full fine-tune | input, preregistered 25K claim passed | scored in one pass (largest tested) | evaluated on MASSIVE intent | 4-bit Q4_K_M; same top-1 as FP32 on 240/240 parity rows |
+
+| Build | Size | Runtime |
+|---|---:|---|
+| **Transformers safetensors (bf16) · this repository** | 1.50 GB | PyTorch on CUDA, Apple MPS or CPU (`jev_style_decision.py`) |
+| [GGUF F16 / Q8_0 / Q4_K_M](https://huggingface.co/chaoliangUNSW/Jev-Style-0.8B-Decision-v3-GGUF) | 1.52 / 0.81 / 0.53 GB | llama.cpp + the bundled `jev-score` scorer |
+| [MLX bf16](https://huggingface.co/chaoliangUNSW/Jev-Style-0.8B-Decision-v3-MLX-bf16) | 1.50 GB | Apple silicon, mlx-lm |
+| [MLX 8-bit](https://huggingface.co/chaoliangUNSW/Jev-Style-0.8B-Decision-v3-MLX-8bit) | 0.80 GB | Apple silicon, mlx-lm |
+
+## Highlights
+
+- **79.2% on 2,000 typed decisions: +6.4 points over Jev, +5.7 over our 2B v2, +2.6 over Laya's typed
+  checkpoint.** Its Brier score is **3.2× lower than Jev's** (0.046 vs 0.148). v3 and Laya typed trained on
+  this dataset's train split; Jev's number is zero-shot, from the dataset card.
+- **Up to +30.3 points over the best official Laya checkpoint on five decision tasks**, on identical rows: +30.3 on model routing,
+  +29.5 macro-F1 on toxicity, +29.4 on the 37 locales held out of MASSIVE training, +19.0 on 77-way Banking77
+  and +7.2 balanced accuracy on jailbreak detection. Every paired 95% CI excludes zero.
+- **Ahead of Laya multilingual in 51 of 51 languages.** MASSIVE intent macro accuracy is 71.7% vs 40.1% (+31.7 points),
+  and v3 stays above 3× chance in every language.
+- **4.5× lower NLL, 3.0× lower Brier and 5.5× lower ECE** than the best Laya checkpoint as deployed (with its shipped
+  temperatures), across 49 suites and 17,416 paired rows.
+- **25,600-token inputs, with the preregistered 25K claim passed.** At 24K tokens v3 answers **98.3%** of 1,280
+  real items correctly. The question-only and state-swap controls stay at chance, and the preregistered controlled
+  accuracy is within 1.1 points of the 2K–4K reference.
+- **64.1% zero-shot on JevBench v1.4.1**, ahead of Laya (58.4%) and every Qwen3.5-0.8B-based system on the board
+  (point estimates on 231 public items). On two topic sets it never trained on, v3 leads English Laya by +12.3 and
+  +12.5 points, and it comes within 4 points of Jev on tweet_topic.
+- **Up to 4.6× faster than a Laya-architecture engine when 10 questions share one 4K-token state** (1,381 ms vs
+  6,364 ms with the GGUF runtime's `many_mode="batched"`; the engine is our round-1 MacLaya-4K, one call per question,
+  not an official Laya checkpoint). The
+  **0.53 GB Q4_K_M file matched full precision on 240 of 240 parity rows** (plus 6 of 6 at 16K and 25.6K tokens).
+
+Each result below states its protocol and source under the figure.
+
+## Results
+
+### Typed decisions: 0.8B beats the 2B models and Jev
+
+![Typed-decisions accuracy and Brier score: Jev-Style 0.8B v3 vs Jev, Laya typed and the 2B v1/v2](figures/headline_typed.png)
+
+At 0.8B parameters, v3 scores **79.2%** on the 2,000 typed decisions. That is **+6.4 points over Jev**, +5.7 over
+our 2B v2 and +2.6 over Laya's typed checkpoint, and the Brier score is **3.2× lower than Jev's** (0.046 vs 0.148).
+
+<sub>Typed-decisions test set (LocalLLaMA/typed-decisions), 2,000 decisions from 400 states. In-domain for v3 and Laya typed (both trained on its train split); zero-shot for Jev (numbers from the dataset card, measured through the Jev API on all 2,000 decisions). Laya: official typed-decisions checkpoint re-run by us on identical rows with its shipped temperature. 2B v1/v2: teacher agreement as reported on the v2 card (same 2,000 decisions, scored by that card's harness; v1 was not trained on typed decisions, v2's training pool included typed workflow decisions). Jev's accuracy is published as 0.727, so the gap is 6.40–6.50 points. v3: 1,583 / 2,000 correct, 95% CI 77.3–80.9% (Wilson); v3 minus Laya typed, paired bootstrap 95% CI +1.0 to +4.2 points.</sub>
+
+**Head-to-head against Laya's typed checkpoint.** Both models trained on this dataset's train split, and v3 wins on all four
+metrics, each with a paired 95% CI that excludes zero:
+
+| Metric (2,000 decisions) | Laya typed-decisions checkpoint | **Jev-Style 0.8B v3** | Difference, paired 95% CI |
+|---|---:|---:|---|
+| Accuracy ↑ | 76.6% | **79.2%** | +2.6 pts [+1.0, +4.2] |
+| Soft accuracy ↑ | 47.1% | **52.4%** | +5.4 pts [+5.0, +5.7] |
+| Brier vs soft labels ↓ | 0.061 | **0.046** | −0.016 [−0.019, −0.012] |
+| Score-question MAE ↓ | 0.242 | **0.195** | −0.047 [−0.060, −0.035] |
+
+<sub>Both in-domain; v3 also trained on 27,300 synthetic typed items from other workflows. Laya: official checkpoint re-run by us on identical rows with its shipped temperature. Paired case-cluster bootstrap within suites, 2,000 resamples.</sub>
+
+### Beyond Laya: up to +30 points
+
+![v3 vs the best official Laya checkpoint on five decision tasks](figures/beyond_laya.png)
+
+**On five decision tasks scored on identical rows, the 0.8B v3 beats the best official Laya checkpoint on every
+one:** +19.0 points on 77-way Banking77, +7.2 balanced accuracy on jailbreak detection, +29.5 macro-F1 on
+toxicity, +30.3 on model routing and +29.4 across the 37 locales held out of MASSIVE training.
+
+<sub>Laya numbers: official checkpoints (English, typed-decisions, multilingual) re-run by us on identical rows with their shipped temperatures and default token budgets; the best of the three is shown per task. v3 trained on tasks of the same kind from other datasets, never on these evaluation rows: intent (CLINC150/HWU64; Banking77 never trained), jailbreak (other permissive sets plus teacher data), toxicity (civil_comments plus teacher data; toxic-chat is evaluation-only), routing (teacher-written; the gsm8k/mbpp/AG rows are evaluation-only), MASSIVE in 14 other locales (no MASSIVE rows in these 37). n = 400 / 400 / 400 / 399 / 3,700 (37 × 100). Every gap's paired 95% bootstrap CI excludes zero.</sub>
+
+### 51 languages, 51 wins
+
+![Per-language MASSIVE intent accuracy, v3 vs Laya multilingual, 51 languages](figures/multilingual.png)
+
+**One 0.8B model, 51 languages, 51 wins over Laya.** On MASSIVE intent (20 options per question) v3 averages
+**71.7%** across 51 languages, against 40.1% for the official Laya multilingual checkpoint (+31.7 points). It
+beats Laya multilingual in every one of the 51 languages, by at least 11 points, and stays above 3× chance in all
+of them. That includes the 37 locales held out of MASSIVE training (65.5% vs 36.1%), 32 of them outside the 19
+fine-tuning languages.
+
+<sub>MASSIVE intent (mteb/amazon_massive_intent) test rows, 100 per language, 20 candidate intents per row (chance 5%, 3× chance 15%); accuracy = top-scored option. v3: in-domain for the 14 trained locales, held out for the other 37 (vi/th/el/ur had about 1.3K translated-NLI training rows each; zh-TW shares Chinese with zh-CN; a 69-row multilingual jailbreak set in training may include a few prompts in other held-out languages). Laya: official multilingual checkpoint re-run by us on identical rows with its shipped temperature and default token budget (held-out for Laya). v3 is also ahead of the best of the three official Laya checkpoints in all 51 languages (per-language point estimates on 100 rows each, smallest gap 10 points). Paired 95% CI of the 51-language macro difference: +30.1 to +33.1 points.</sub>
+
+### Probabilities you can act on
+
+![Macro NLL, Brier and ECE over 49 suites: v3 vs the best Laya checkpoint](figures/calibration.png)
+
+**Across 49 suites and 17,416 identical rows, v3's probabilities beat the best official Laya checkpoint on all three
+probability-quality metrics:** **4.5× lower NLL** (0.493 vs 2.213), **3.0× lower Brier** (0.239 vs 0.712) and
+**5.5× lower ECE** (0.054 vs 0.299).
+
+<sub>Macro average over 49 suites, 17,416 identical rows for both models. Mixed protocol for both v3 and Laya (in-domain on some suites, held-out on others). As deployed: v3 with its supplied temperatures; Laya with the shipped temperatures of its official checkpoints, re-run by us on identical rows. Best Laya = best of the three official checkpoints per metric (multilingual on all three). Paired case-cluster bootstrap 95% CIs (2,000 resamples) of every difference exclude zero.</sub>
+
+**Stable under option shuffling.** When the options are presented in a different order, v3 changed its answer on
+**1 of 200** option-order pairs (0.5%), against 22 of 200 (11.0%) for the best Laya checkpoint.
+
+<sub>MASSIVE intent English, 200 option-permutation pairs, identical rows. In-domain for v3, held-out for Laya (typed-decisions checkpoint, the best of the three here). Paired 95% CI of the difference: −15.0 to −6.5 points.</sub>
+
+### Long context: flat from 1K to 24K tokens
+
+![Controlled accuracy by input length, 1K to 24K tokens](figures/long_context.png)
+
+**v3 reads documents far past Laya's 512 / 1,024-token default budgets.** Controlled accuracy stays within 3.8
+points across all seven length bins. At 24K it is 55.3%, 1.1 points from the 2K–4K reference (56.4%) and well
+inside the preregistered ±5-point limit, so **the 25K claim passed**. In plain accuracy, v3 answers **98.3%**
+(1,258 of 1,280) of the real 24K-token items correctly and at least 96.9% in every length bin. The same questions
+with the state removed or swapped for another item's state fall to chance (28.9% and 28.5%, against 28.4% chance
+at 24K), so the answers cannot be recovered from the question alone.
+
+<sub>v3 only. Laya's default input budget is 512 tokens (English) / 1,024 (multilingual, typed) per the Laya README, so Laya is not plotted. Suite long_grid_plus, English and Chinese documents: preregistered 2026-09-24 and amended before any model was scored (+96 items per 24K depth decile, thresholds unchanged); 320 items per bin, 1,280 at 24K. Controlled accuracy = the real item is correct AND its question-only and state-swap controls pass; both controls are at chance in every length bin. 25K claim rule: |24K − 2K–4K reference| ≤ 5 points and every 24K evidence-depth decile within 10 points of it.</sub>
+
+### JevBench: ahead of Laya and every Qwen3.5-0.8B-based system
+
+![JevBench v1.4.1 public accuracy: v3 vs Laya and the Qwen3.5-0.8B-based systems](figures/jevbench.png)
+
+**On the 231 public JevBench v1.4.1 items, v3 scores 64.1% zero-shot**: 5.6 points above Laya, and ahead of every
+Qwen3.5-0.8B-based system on the board, including a dedicated 0.8B decision fine-tune (+4.8 points) and
+SimpleJev on the same base (+9.5 points). Every answer is a valid option (231 of 231), because v3 can only score
+the options it is given.
+
+<sub>JevBench v1.4.1, public items only (231). v3: self-run zero-shot with the vendored official harness (commit 24b9b5c), 148 / 231 correct, 95% CI 57.7–70.0% (Wilson); training-pool contamination scan: 0 hits; not an official leaderboard entry. Other rows: public accuracy as published in the board's [v1.4.1 results file](https://github.com/fstandhartinger/jevbench). Shown: Laya plus every Qwen3.5-0.8B-based system on the board; other board systems are not shown. Laya's and M. Ghafiri's scores lie inside v3's 95% CI, so those two leads are point estimates, not significant at n = 231.</sub>
+
+### Zero-shot topics: +12 points over English Laya
+
+![Zero-shot tweet_topic and fin_topic accuracy: v3 vs English Laya, with Jev on tweet_topic](figures/zeroshot.png)
+
+**On two topic sets it never trained on, v3 leads English Laya by +12.3 points on tweet_topic** (75.5% vs 63.2%)
+**and +12.5 points on the 20-way fin_topic** (46.7% vs 34.2%). On tweet_topic it lands **within 4 points of Jev**
+(75.5% vs 79.3%). Macro-F1 leads over English Laya are +13.8 points (59.9% vs 46.1%) and +8.9 points (45.2% vs 36.2%).
+With its shipped temperature, its probabilities are also better calibrated than Jev's on both sets: ECE 0.027 vs
+0.063 on tweet_topic and 0.046 vs 0.166 on fin_topic.
+
+<sub>Zero-shot for every system: neither set is in v3's training pool; accuracy over every row of the pinned test files (n = 1,693 and 4,117). Jev (1.13, API) and English Laya: numbers published by the [elcronos jev-vs-open-decision-models study](https://github.com/elcronos/jev-vs-open-decision-models) with its own prompt (results/cross_dataset_summary.json @ a1901bc), not re-run by us. v3: scored by us on the identical rows, label sets and instruction, in v3's own input format; tweet_topic accuracy 95% CI 73.4–77.5%. ECE: 15 equal-width bins as in the study; v3 with its shipped global temperature (0.880, fitted on v3's own calibration split, never on these sets), Jev's ECE as published (raw API probabilities).</sub>
+
+### Speed: many questions, one read
+
+![Warm p50 latency: v3 vs a Laya-architecture engine, 1K to 8K-token states](figures/latency.png)
+
+**Ask many questions about one long state and v3 pulls away.** v3 reads the state once and scores every question
+in a single call (GGUF runtime, `many_mode="batched"`; the default exact mode shares whole 1,024-token chunks of the
+state and gives results identical to one call per question). With 5 to 10 questions per state, that makes it **1.4× to 1.9× faster** than a Laya-architecture
+engine (our round-1 MacLaya-4K) on 1K-token states and **2.6× to 4.6× faster** on 4K-token states (4K tokens with
+10 questions: 1,381 ms vs 6,364 ms). It also answers
+questions about 8K-token states in 2.3 to 2.6 s, which the 4K-budget engine cannot run at all.
+
+<sub>Identical-architecture timing: untrained Qwen3.5-0.8B export (latency does not depend on the weights). v3 = llama.cpp GGUF F16, one call per state with all questions scored together (`many_mode="batched"` in the GGUF runtime). Comparison engine = round-1 MacLaya-4K, our own fine-tune of the Laya multilingual architecture (4,096-token budget, FP32 on Apple MPS), one call per question; it is not an official Laya checkpoint. Compared at 5 and 10 questions per state. Apple M1 Max 64 GB, warm end-to-end p50, idle run 2026-09-23, prefix reuse off.</sub>
+
+### Quantization: 4-bit, 0.53 GB, same calls
+
+![Top-1 agreement with full precision and file size for every format of v1, v2 and v3](figures/quantization.png)
+
+**Quantize it to 4-bit and it still makes the same call.** Every shipped v3 format (GGUF F16, Q8_0 and Q4_K_M;
+MLX bf16 and 8-bit) matches the PyTorch FP32 model on **240 of 240 parity rows**, plus 6 of 6 prompts of about 16K
+and 25.6K tokens. The 0.53 GB Q4_K_M file is about **2.4× smaller** than the 2B v2's Q4_K_M (1.27 GB).
+
+<sub>v3: top-1 agreement with the PyTorch FP32 reference on 240 parity rows (a mixed fixture drawn from the training pool, 22 categories, English and Chinese), plus 6 extra rows at about 16K and 25.6K tokens (3 each), where every format also agrees 6/6. v3 sizes are the exported weight files (GB = 10^9 bytes). 2B v1/v2 numbers and sizes are as reported on their public Hugging Face GGUF cards: 500 held-out decisions each, against bf16 for v1 and CUDA merged BF16 for v2. The fixtures (training-pool rows for v3, held-out rows for v1/v2) and references differ, so the rows are not a paired comparison and no agreement gap is claimed.</sub>
+
+## Third generation: what changed
+
+![Design comparison of Jev-Style v1, v2 and v3](figures/design_table.png)
+
+<sub>v1/v2: as reported on their public Hugging Face cards (v1 GGUF card; v2 and v2-GGUF cards; v1's typed-decisions number is reported on the v2 card). v3: release manifest, training config and evaluation files; Q4_K_M size = exported file (GB = 10^9 bytes), parity rows drawn from the training pool. Typed decisions: same 2,000 decisions from 400 states; v1/v2 scored by the v2 card's harness, v3 by ours. In-domain for v3; v1 was not trained on typed decisions; v2's pool included typed workflow decisions. "51 evaluated" = MASSIVE locales, 14 trained + 37 held out; the fine-tuning pool covers 19 languages. 25,600 tokens = the runtime's whole-input limit, 25× v2's 1,024-token prompt.</sub>
+
+<details>
+<summary><strong>The same comparison as a text table</strong></summary>
+
+| | Jev-Style 2B v1 | Jev-Style 2B v2 | **Jev-Style 0.8B v3** |
+|---|---|---|---|
+| Parameters | 2B (Qwen3.5-2B-Base) | 2B (continued from v1) | **0.8B** (752M text-model parameters) |
+| Training | LoRA rank 16 (all linear layers) | LoRA rank 32 (33.6M trainable parameters) | **Full fine-tune** (every weight trained) |
+| Readout | Option-letter token (one letter per option) | Option-letter token (' A' ... ' Z') | **Verdict slot per option** (every option scored, one pass) |
+| Options per decision | Up to 26 (20 via top_logprobs) | 2–26 (letter-capped) | **No letter cap** (tested with 77 options) |
+| Context | Not stated (quickstart: server default) | 1,024-token prompt (quickstart runs -c 2048) | **25,600 tokens** (preregistered 25K claim passed) |
+| Languages | English (five English task families) | English (English state required) | **51 evaluated** (MASSIVE locales; 19 languages in fine-tuning) |
+| Questions per state read | 1 (one question per prompt) | 1 (one question per prompt) | **Many** (all questions in one call) |
+| Q4_K_M file | 1.3 GB (as reported on the v1 GGUF card) | 1.27 GB (as reported on the v2 GGUF card) | **0.53 GB** (matches FP32 on 240 / 240 parity rows) |
+| Typed decisions, teacher agreement | 53.35% (2,000 decisions / 400 states) | 73.45% (same 2,000 decisions) | **79.15%** (same 2,000; 1,583 correct) |
+
+</details>
+
+Three ceilings of v1 and v2 are gone in v3:
+
+- **No letter cap.** v1 and v2 read one option-letter token, so a question could have at most 26 options. v3
+  scores a verdict slot per option, so the options are whatever you pass. Banking77 was run with all 77 intents
+  in one pass.
+- **25× the prompt budget.** v2's interface is a 1,024-token prompt. v3 takes 25,600 tokens, and its long-context
+  claim was preregistered and passed.
+- **Read once, ask many.** v1 and v2 put one question in each prompt. v3 renders the state once and answers any
+  number of questions about it. With 10 questions on a 4K-token state this is 4.6× faster (GGUF, `many_mode="batched"`) than a
+  Laya-architecture engine (our round-1 MacLaya-4K) that calls once per question.
+
+## What "Jev-style" means
+
+[Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) (TypeSafe AI, 2026) introduced *System One*
+decision models. Instead of generating text, the model takes a state and a typed question and returns a probability
+for each allowed answer in a single pass. Jev-Style models follow that pattern with open weights:
+
+- **choice**: pick one of N named options, with a probability for each;
+- **noul** (yes/no): the probability that a statement about the state is true;
+- **score**: a distribution over 2 to 10 ordered levels.
+
+The model cannot answer outside the options it is given, and it never decodes text.
+
+> **Independent project.** Jev-Style is not affiliated with, endorsed by or connected to TypeSafe AI or Jev, and no
+> Jev weights, code or outputs are used. It is also not affiliated with the Laya authors or the Qwen team. Jev and
+> Laya numbers on this card come from the sources named under each result.
+
+## Quick start (transformers)
+
+```bash
+pip install -U huggingface_hub
+hf download chaoliangUNSW/Jev-Style-0.8B-Decision-v3 --local-dir jev-v3
+cd jev-v3
+pip install -r requirements.txt      # torch, transformers>=5.0 (Qwen3.5 support), tokenizers, numpy
+```
+
+The repository ships `jev_style_decision.py`, a self-contained runtime. It handles input rendering, the verdict
+readout, the fitted temperatures and budget checks.
+
+```python
+from jev_style_decision import JevStyleDecision
+
+m = JevStyleDecision(".")        # float32 on CUDA, Apple MPS or CPU (device="cpu" to force)
+r = m.decide(
+    {"ticket": "I was charged twice for my subscription this month.", "customer_tier": "pro"},
+    "Which team should handle this ticket?",
+    options={"billing": "payments, invoices, refunds",
+             "technical": "bugs and outages",
+             "sales": "new purchases"},
+    category="theme_routing",
+)
+print(r["answer"], r["probabilities"])
+# billing  (probabilities ≈ billing 0.978, sales 0.017, technical 0.004 on CPU, float32)
+```
+
+Other question types, and several questions about one state:
+
+```python
+state = "Order #1182: paid, packed, handed to the courier on Monday. Tracking shows 'delivered' on Wednesday."
+m.decide(state, "Has the order been delivered?", qtype="noul")                       # {"false": p, "true": p}
+m.decide(state, "How urgent is a follow-up?", qtype="score",
+         options=["not urgent", "somewhat urgent", "urgent", "critical"])            # levels "0".."3"
+m.decide_many(state, [
+    {"t": "noul", "ins": "Was the order paid?", "crit": None},
+    {"t": "choice", "ins": "Which step is the order at?",
+     "crit": {"packing": None, "in transit": None, "delivered": None}},
+])
+```
+
+From the command line:
+
+```bash
+python jev_style_decision.py --state "The film was excellent." \
+  --question "What is the sentiment of this review?" \
+  --options '["negative", "positive"]' --category general_sentiment
+# -> "answer": "positive", probability 0.989
+```
+
+`decide` returns a dict with `answer`, `probabilities`, the raw `scores`, the `temperature` used,
+`top_probability`, `entropy_concentration`, `input_tokens`, `head_tokens`, `model` and `backend`. Batch mode reads JSON lines (`--jsonl file|-`), and
+`--verify` checks every file against the sha256 manifest before loading.
+
+**GGUF (llama.cpp):** [Jev-Style-0.8B-Decision-v3-GGUF](https://huggingface.co/chaoliangUNSW/Jev-Style-0.8B-Decision-v3-GGUF).
+It includes `jev_score.cpp`, a small libllama scorer that reads logits only at the verdict slots and scores many
+questions on one decoded state, plus `jev_style_decision_gguf.py` with the same API.
+
+**MLX (Apple silicon):** [bf16](https://huggingface.co/chaoliangUNSW/Jev-Style-0.8B-Decision-v3-MLX-bf16) and
+[8-bit](https://huggingface.co/chaoliangUNSW/Jev-Style-0.8B-Decision-v3-MLX-8bit), each with
+`jev_style_decision_mlx.py` and the same API.
+
+## Input format and readout
+
+Each segment is tokenised on its own and the pieces are concatenated. Text inside the state or the options is
+tokenised with special tokens disabled, so a string such as `<|im_end|>` in user data stays plain text.
+
+```text
+State:
+<state: plain text, or any JSON value serialised with ensure_ascii=False>
+
+Question [<choice|noul|score>]: <question>
+Options:
+- <option 1>
+- <option 2>
+Judge each option:
+<option 1> ->
+<option 2> ->
+```
+
+- **Verdict slot.** The hidden state at the ` ->` token that ends option k's line is option k's verdict slot. Its
+  score is `logit(" yes") − logit(" no")` at that position, computed as `h_k · (w_yes − w_no)` from the final
+  normalised hidden state and the tied embedding rows, in float32. No parameters are added, so the weights stay a
+  standard Qwen3.5 text model.
+- **Probabilities.** `softmax(scores / T)`. `T` is looked up by calibration family × question type × option-count
+  bucket in `readout_config.json` (20 fitted groups), and the global value 0.880 is the fallback. Pass
+  `category=` (for example `theme_routing`, `general_topic`, `intent`, `typed_official`) to pick the family. With
+  no `category`, the runtime uses the global temperature (0.880), and `temperature=1.0` gives the raw scores.
+- **Option text.** Choice options render as `name` or `name: description`. Score levels render as
+  `level i: description`. Yes/no questions render as `false: …` / `true: …`, with default descriptions when none
+  are given.
+
+## Usage notes
+
+- **Input budget.** The whole rendered input, meaning state, question, options and readout, may be up to 25,600
+  tokens. The head (question, options and readout) may be up to 2,048 tokens. Over-budget inputs raise
+  `InputBudgetError`, and nothing is ever truncated silently.
+- **Decisions only.** The model scores the options you give it and returns probabilities. It does not generate
+  text, and it takes no actions on its own.
+- **Options.** `choice` takes any number of named options within the 2,048-token head (77 is the largest set we
+  evaluated; the GGUF runtime accepts up to 256 options per question). `score` takes 2 to 10 ordered levels, lowest first. `noul` needs no options.
+- **Several questions about one state.** Use `decide_many`. In the GGUF runtime it sends all questions to the
+  bundled scorer in one request. By default the results are identical to one `decide` call per question; to keep
+  them identical, the state is shared only in whole 1,024-token blocks, so the time saved starts at 1,024-token
+  states and grows with the state length. `JevStyleDecisionGGUF(..., many_mode="batched")` reads the whole state
+  once and scores all questions together, as in the latency chart; its probabilities differed from `decide` by at
+  most 0.002 in our tests, and a near-tied top answer can change.
+- **Precision.** The evaluation numbers on this card were computed with the PyTorch weights. The GGUF and MLX
+  builds were checked for top-1 parity with the PyTorch FP32 reference (see Quantization).
+- **Weights.** This is a text-only `Qwen3_5ForCausalLM`: 752,393,024 parameters, 24 layers (18 Gated DeltaNet +
+  6 full attention), hidden size 1,024, tied embeddings. The vision tower and the multi-token-prediction head
+  were removed.
+
+## Training
+
+- **Base:** [Qwen/Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) (revision `2fc06364`).
+- **Run:** full fine-tune in bf16 on one NVIDIA H100 80GB. It took 994 optimizer steps over 131.4M tokens, with
+  a learning rate of 2e-5 and about 96 minutes of training steps (5,780 s). Checkpoint step 981 was selected by
+  the preregistered development score.
+- **Mixture:** a 321,756-row training pool in 19 languages, with inputs up to 25,600 tokens for long documents and agent
+  histories:
+  - typed decisions: the LocalLLaMA/typed-decisions train split plus 27,300 synthetic typed items;
+  - general classification and QA: MNLI and translated NLI, AG News, GoEmotions, DAIR Emotion, SQuAD v2, SST-5
+    and BoolQ;
+  - intents: MASSIVE intent and scenario in 14 locales, CLINC150 and HWU64;
+  - application themes: spam, phishing, jailbreak and prompt injection, toxicity, ticket triage and model routing;
+  - Mac agent step-gate and goal-done checks;
+  - long-context retrieval, tables and QA.
+- **Calibration:** 20 group temperatures plus a global one, fitted on 15,655 held-out calibration rows (never
+  test rows).
+
+## Training data and licences
+
+- **Base model:** Qwen/Qwen3.5-0.8B by the Qwen team (Alibaba Cloud), Apache-2.0. The Apache License 2.0 text is
+  in `LICENSE`, and `NOTICE` lists the modifications. This model is released under Apache-2.0.
+- **Datasets with restrictive or unclear terms** (kept in training by the author's decision):
+  - DAIR Emotion (14,757 training rows): its dataset card says it should be used for educational and research
+    purposes only.
+  - AG News (30,000 training rows): its licence is listed as unknown, and its card describes it as provided by the
+    academic community for research and non-commercial use.
+  - SST-5, MNLI, an HWU64 mirror, QuALITY, Enron spam and a phishing-email dataset also carry their own terms.
+    Check each source before commercial use.
+- **Outputs of other models** (kept by the author's decision): an OpenAI GPT model wrote the theme data for
+  routing, triage and jailbreak, the teacher-translated NLI data and the Chinese filler text for long documents.
+  OpenAI GPT and Anthropic Claude models designed the synthetic typed-decision workflows, and Anthropic Claude
+  models labelled them. The providers' terms of use may restrict how models trained on such outputs may be used,
+  so check them for your use case.
+- **Evaluation-only data** (0 training rows): Banking77, toxic-chat, XNLI, ContractNLI, MS MARCO, tweet_topic,
+  fin_topic, the JevBench items and the support-ticket set.
+
+<details>
+<summary><strong>Evaluation records and vector charts</strong></summary>
+
+- Every chart is also provided as SVG: [headline_typed](figures/headline_typed.svg),
+  [beyond_laya](figures/beyond_laya.svg), [multilingual](figures/multilingual.svg),
+  [calibration](figures/calibration.svg), [long_context](figures/long_context.svg),
+  [jevbench](figures/jevbench.svg), [zeroshot](figures/zeroshot.svg), [latency](figures/latency.svg),
+  [quantization](figures/quantization.svg), [design_table](figures/design_table.svg).
+- Plotted values, sources and protocol labels for each chart:
+  [headline_typed](figures/headline_typed.data.json), [beyond_laya](figures/beyond_laya.json),
+  [multilingual](figures/multilingual.data.json), [calibration](figures/calibration.data.json),
+  [long_context](figures/long_context.json), [jevbench](figures/jevbench.data.json),
+  [zeroshot](figures/zeroshot.json), [latency](figures/latency.data.json),
+  [quantization](figures/quantization.data.json), [design_table](figures/design_table.data.json).
+- Every v3 and re-run Laya number comes from prediction files that were each scored once. Paired differences use
+  a case-cluster bootstrap within suites (2,000 resamples). A win is only claimed when the 95% CI excludes zero,
+  except where a chart or note says otherwise (JevBench leads over Laya and M. Ghafiri, and per-language MASSIVE
+  gaps against the best of three Laya checkpoints, are point estimates).
+- Public sources: [Laya](https://huggingface.co/convaiinnovations/laya) (official checkpoints and README;
+  [BENCHMARKS.md](https://github.com/NandhaKishorM/laya/blob/main/BENCHMARKS.md)),
+  [LocalLLaMA/typed-decisions](https://huggingface.co/datasets/LocalLLaMA/typed-decisions) (dataset card with Jev's
+  numbers), [JevBench](https://github.com/fstandhartinger/jevbench) (v1.4.1 results file),
+  [elcronos/jev-vs-open-decision-models](https://github.com/elcronos/jev-vs-open-decision-models) (zero-shot topic
+  study), and the [2B v1](https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF) and
+  [2B v2](https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2) cards.
+
+</details>
+
+## Citation
+
+```bibtex
+@misc{jevstyle2026v3,
+  title        = {Jev-Style-0.8B-Decision-v3: a long-context, multilingual 0.8B decision model},
+  author       = {chaoliangUNSW},
+  year         = {2026},
+  howpublished = {\url{https://huggingface.co/chaoliangUNSW/Jev-Style-0.8B-Decision-v3}},
+  note         = {Fine-tuned from Qwen/Qwen3.5-0.8B}
+}
+```
+
+## Contact
+
+I welcome internship, employment, and research collaboration opportunities. Please contact me at [**yanchaoliang369@gmail.com**](mailto:yanchaoliang369@gmail.com).
+
+欢迎提供实习、工作及科研合作机会，请邮件联系：[yanchaoliang369@gmail.com](mailto:yanchaoliang369@gmail.com)。

chat_template.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is true %}
+        {{- '<think>\n' }}
+    {%- else %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

config.json

@@ -0,0 +1,75 @@
+{
+  "architectures": [
+    "Qwen3_5ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attn_output_gate": true,
+  "bos_token_id": null,
+  "dtype": "bfloat16",
+  "eos_token_id": 248044,
+  "full_attention_interval": 4,
+  "head_dim": 256,
+  "hidden_act": "silu",
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 3584,
+  "layer_types": [
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention",
+    "linear_attention",
+    "linear_attention",
+    "linear_attention",
+    "full_attention"
+  ],
+  "linear_conv_kernel_dim": 4,
+  "linear_key_head_dim": 128,
+  "linear_num_key_heads": 16,
+  "linear_num_value_heads": 16,
+  "linear_value_head_dim": 128,
+  "mamba_ssm_dtype": "float32",
+  "max_position_embeddings": 262144,
+  "mlp_only_layers": [],
+  "model_type": "qwen3_5_text",
+  "mtp_num_hidden_layers": 1,
+  "mtp_use_dedicated_embeddings": false,
+  "num_attention_heads": 8,
+  "num_hidden_layers": 24,
+  "num_key_value_heads": 2,
+  "pad_token_id": null,
+  "partial_rotary_factor": 0.25,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "mrope_interleaved": true,
+    "mrope_section": [
+      11,
+      11,
+      10
+    ],
+    "partial_rotary_factor": 0.25,
+    "rope_theta": 10000000,
+    "rope_type": "default"
+  },
+  "tie_word_embeddings": true,
+  "transformers_version": "5.17.0",
+  "use_cache": true,
+  "vocab_size": 248320
+}

figures/beyond_laya.json

@@ -0,0 +1,94 @@
+{
+  "chart": "beyond_laya",
+  "unit": "percent (value x 100)",
+  "source": "runs/macjev/report_r2/scoreboard/scoreboard.json via hf_staging/v3_card/chart_data.json",
+  "delta_ci_method": "case-cluster bootstrap within suites, 2000 resamples (scoreboard n_boot)",
+  "rows": [
+    {
+      "task": "Banking77",
+      "metric": "accuracy, 77 intents",
+      "v3": 0.6825,
+      "best_laya": 0.4925,
+      "best_laya_checkpoint": "typed",
+      "delta_pts": 19.0,
+      "delta_ci95_pts": [
+        13.99,
+        24.0
+      ],
+      "n": 400,
+      "v3_entry": "banking77.v3",
+      "laya_entry": "banking77.laya_best",
+      "scoreboard_field": "metrics['banking77']",
+      "protocol_label": "v3: held-out (banking77 never trained; CLINC150/HWU64 intents are) / Laya: held-out. Laya numbers: official checkpoints re-run by us on identical rows, shipped temperatures."
+    },
+    {
+      "task": "Jailbreak",
+      "metric": "balanced accuracy",
+      "v3": 0.9036238842064085,
+      "best_laya": 0.8311462000782389,
+      "best_laya_checkpoint": "multilingual",
+      "delta_pts": 7.25,
+      "delta_ci95_pts": [
+        2.99,
+        11.43
+      ],
+      "n": 400,
+      "v3_entry": "theme.guardrails_jailbreak.balanced_accuracy.v3",
+      "laya_entry": "theme.guardrails_jailbreak.balanced_accuracy.laya_best",
+      "scoreboard_field": "metrics['theme.guardrails_jailbreak.balanced_accuracy']",
+      "protocol_label": "v3: held-out source (other permissive jailbreak sets + teacher) / Laya: held-out. Laya numbers: official checkpoints re-run by us on identical rows, shipped temperatures."
+    },
+    {
+      "task": "Toxicity",
+      "metric": "macro-F1",
+      "v3": 0.7089254055198327,
+      "best_laya": 0.41360068097985436,
+      "best_laya_checkpoint": "multilingual",
+      "delta_pts": 29.53,
+      "delta_ci95_pts": [
+        24.65,
+        34.37
+      ],
+      "n": 400,
+      "v3_entry": "theme.moderation_toxicity.macro_f1.v3",
+      "laya_entry": "theme.moderation_toxicity.macro_f1.laya_best",
+      "scoreboard_field": "metrics['theme.moderation_toxicity.macro_f1']",
+      "protocol_label": "v3: held-out source (civil_comments + teacher; toxic-chat eval-only) / Laya: held-out. Laya numbers: official checkpoints re-run by us on identical rows, shipped temperatures."
+    },
+    {
+      "task": "Model routing",
+      "metric": "accuracy",
+      "v3": 0.9624060150375939,
+      "best_laya": 0.6591478696741855,
+      "best_laya_checkpoint": "typed",
+      "delta_pts": 30.33,
+      "delta_ci95_pts": [
+        26.06,
+        35.09
+      ],
+      "n": 399,
+      "v3_entry": "theme.model_routing_domain.v3",
+      "laya_entry": "theme.model_routing_domain.laya_best",
+      "scoreboard_field": "metrics['theme.model_routing_domain']",
+      "protocol_label": "v3: near-domain (teacher-written; gsm8k/mbpp/AG rows eval-only) / Laya: held-out. Laya numbers: official checkpoints re-run by us on identical rows, shipped temperatures."
+    },
+    {
+      "task": "MASSIVE intent",
+      "metric": "37 held-out locales",
+      "v3": 0.6551351351351351,
+      "best_laya": 0.36108108108108106,
+      "best_laya_checkpoint": "multilingual",
+      "delta_pts": 29.41,
+      "delta_ci95_pts": [
+        27.54,
+        31.3
+      ],
+      "n": 3700,
+      "v3_entry": "massive51.unseen37.macro_accuracy.v3",
+      "laya_entry": "massive51.unseen37.macro_accuracy.laya_best",
+      "scoreboard_field": "metrics['massive51.unseen37.macro_accuracy']",
+      "protocol_label": "v3: held-out (37 languages never trained) / Laya: held-out. Laya numbers: official checkpoints re-run by us on identical rows, shipped temperatures."
+    }
+  ],
+  "footnote": "Laya numbers: official checkpoints (English, typed-decisions, multilingual) re-run by us on identical rows with their shipped temperatures and\ndefault token budgets; the best of the three is shown per task. v3 trained on same-kind tasks from other datasets, never on these eval rows:\nintent (CLINC150/HWU64; Banking77 never trained), jailbreak (other permissive sets + teacher), toxicity (civil_comments + teacher; toxic-chat\neval-only), routing (teacher-written; gsm8k/mbpp/AG rows eval-only), MASSIVE in 14 other locales (no MASSIVE rows in these 37).\nn = 400 / 400 / 400 / 399 / 3,700 (37 x 100). Every gap's paired 95% bootstrap CI excludes zero. Plotted values: figures/beyond_laya.json."
+}

figures/calibration.data.json

@@ -0,0 +1,99 @@
+{
+ "chart": "calibration",
+ "metric": "49-suite macro NLL / Brier / ECE, as deployed (lower is better)",
+ "protocol_label": "v3: mixed (49 suites) / Laya: mixed (49 suites). Laya numbers: official checkpoints re-run by us on identical rows, shipped temperatures.",
+ "rows": [
+  {
+   "panel": "NLL",
+   "metric": "49-suite macro NLL, as deployed",
+   "n_rows": 17416,
+   "v3": 0.4928839178581892,
+   "best_laya": 2.212565130608199,
+   "best_laya_checkpoint": "multilingual",
+   "all_laya_checkpoints": {
+    "english": 9.669232280968933,
+    "typed": 7.344998151307391,
+    "multilingual": 2.212565130608199
+   },
+   "ratio_best_laya_over_v3": 4.48901871301224,
+   "diff_v3_minus_best": -1.7196812127500098,
+   "diff_ci95": [
+    -1.7632826815264786,
+    -1.6751892907984907
+   ],
+   "diff_ci_method": "case-cluster bootstrap within suites, 2000 resamples (scoreboard n_boot)",
+   "source_path": "runs/macjev/report_r2/scoreboard/scoreboard.json",
+   "fields": [
+    "metrics['t4.macro_nll'].ours_value",
+    "metrics['t4.macro_nll'].laya_best"
+   ],
+   "entries": [
+    "t4.macro_nll.v3",
+    "t4.macro_nll.laya_best"
+   ],
+   "claim": "t4.macro_nll_vs_best_laya"
+  },
+  {
+   "panel": "Brier score",
+   "metric": "49-suite macro Brier, as deployed",
+   "n_rows": 17416,
+   "v3": 0.23930092306086748,
+   "best_laya": 0.7120664110846202,
+   "best_laya_checkpoint": "multilingual",
+   "all_laya_checkpoints": {
+    "english": 1.0124090901595464,
+    "typed": 0.9474950671291875,
+    "multilingual": 0.7120664110846202
+   },
+   "ratio_best_laya_over_v3": 2.975610799894417,
+   "diff_v3_minus_best": -0.47276548802375273,
+   "diff_ci95": [
+    -0.4843787573639837,
+    -0.46074418926883626
+   ],
+   "diff_ci_method": "case-cluster bootstrap within suites, 2000 resamples (scoreboard n_boot)",
+   "source_path": "runs/macjev/report_r2/scoreboard/scoreboard.json",
+   "fields": [
+    "metrics['t4.macro_brier'].ours_value",
+    "metrics['t4.macro_brier'].laya_best"
+   ],
+   "entries": [
+    "t4.macro_brier.v3",
+    "t4.macro_brier.laya_best"
+   ],
+   "claim": "t4.macro_brier_vs_best_laya"
+  },
+  {
+   "panel": "ECE",
+   "metric": "49-suite macro ECE, as deployed",
+   "n_rows": 17416,
+   "v3": 0.05416919804670487,
+   "best_laya": 0.2994167384382702,
+   "best_laya_checkpoint": "multilingual",
+   "all_laya_checkpoints": {
+    "english": 0.4647782759946152,
+    "typed": 0.40828649732151606,
+    "multilingual": 0.2994167384382702
+   },
+   "ratio_best_laya_over_v3": 5.527435318132494,
+   "diff_v3_minus_best": -0.24524754039156535,
+   "diff_ci95": [
+    -0.24448168599481393,
+    -0.22893121774401398
+   ],
+   "diff_ci_method": "case-cluster bootstrap within suites, 2000 resamples (scoreboard n_boot)",
+   "source_path": "runs/macjev/report_r2/scoreboard/scoreboard.json",
+   "fields": [
+    "metrics['t4.macro_ece_as_deployed'].ours_value",
+    "metrics['t4.macro_ece_as_deployed'].laya_best"
+   ],
+   "entries": [
+    "t4.macro_ece_as_deployed.v3",
+    "t4.macro_ece_as_deployed.laya_best"
+   ],
+   "claim": "t4.macro_ece_as_deployed_vs_best_laya"
+  }
+ ],
+ "footnote": "Macro average over 49 suites, 17,416 identical rows for both models. Protocol: mixed for both v3 and Laya (in-domain on some suites, held-out on others). As deployed: v3 with its supplied deployment temperatures; Laya numbers: official checkpoints re-run by us on identical rows with their shipped temperatures. Best Laya = best of the three official checkpoints (English, typed-decisions, multilingual) per metric; multilingual is best on all three. Paired case-cluster bootstrap 95% CIs (2,000 resamples) of every v3 minus Laya difference exclude zero. Plotted values: figures/calibration.data.json.",
+ "not_plotted": "optional typed-decisions reliability diagram omitted (see agent report)"
+}

figures/design_table.data.json

@@ -0,0 +1,175 @@
+{
+ "chart": "design_table",
+ "columns": [
+  "Jev-Style 2B v1",
+  "Jev-Style 2B v2",
+  "Jev-Style 0.8B v3"
+ ],
+ "rows": [
+  {
+   "row": "Parameters",
+   "Jev-Style 2B v1": {
+    "main": "2B",
+    "sub": "Qwen3.5-2B-Base"
+   },
+   "Jev-Style 2B v2": {
+    "main": "2B",
+    "sub": "continued from v1"
+   },
+   "Jev-Style 0.8B v3": {
+    "main": "0.8B",
+    "sub": "752M text-model params"
+   }
+  },
+  {
+   "row": "Training",
+   "Jev-Style 2B v1": {
+    "main": "LoRA rank 16",
+    "sub": "all linear layers"
+   },
+   "Jev-Style 2B v2": {
+    "main": "LoRA rank 32",
+    "sub": "33.6M trainable params"
+   },
+   "Jev-Style 0.8B v3": {
+    "main": "Full fine-tune",
+    "sub": "every weight trained"
+   }
+  },
+  {
+   "row": "Readout",
+   "Jev-Style 2B v1": {
+    "main": "Option-letter token",
+    "sub": "one letter per option"
+   },
+   "Jev-Style 2B v2": {
+    "main": "Option-letter token",
+    "sub": "' A' ... ' Z'"
+   },
+   "Jev-Style 0.8B v3": {
+    "main": "Verdict slot per option",
+    "sub": "every option scored, one pass"
+   }
+  },
+  {
+   "row": "Options per decision",
+   "Jev-Style 2B v1": {
+    "main": "Up to 26",
+    "sub": "20 via top_logprobs"
+   },
+   "Jev-Style 2B v2": {
+    "main": "2-26",
+    "sub": "letter-capped"
+   },
+   "Jev-Style 0.8B v3": {
+    "main": "No letter cap",
+    "sub": "tested with 77 options"
+   }
+  },
+  {
+   "row": "Context",
+   "Jev-Style 2B v1": {
+    "main": "Not stated",
+    "sub": "quickstart: server default"
+   },
+   "Jev-Style 2B v2": {
+    "main": "1,024-token prompt",
+    "sub": "quickstart runs -c 2048"
+   },
+   "Jev-Style 0.8B v3": {
+    "main": "25,600 tokens",
+    "sub": "preregistered 25K claim passed"
+   }
+  },
+  {
+   "row": "Languages",
+   "Jev-Style 2B v1": {
+    "main": "English",
+    "sub": "five English task families"
+   },
+   "Jev-Style 2B v2": {
+    "main": "English",
+    "sub": "English state required"
+   },
+   "Jev-Style 0.8B v3": {
+    "main": "51 evaluated",
+    "sub": "MASSIVE locales; 19 in fine-tuning"
+   }
+  },
+  {
+   "row": "Questions per state read",
+   "Jev-Style 2B v1": {
+    "main": "1",
+    "sub": "one question per prompt"
+   },
+   "Jev-Style 2B v2": {
+    "main": "1",
+    "sub": "one question per prompt"
+   },
+   "Jev-Style 0.8B v3": {
+    "main": "Many",
+    "sub": "all questions in one call"
+   }
+  },
+  {
+   "row": "Q4_K_M file",
+   "Jev-Style 2B v1": {
+    "main": "1.3 GB",
+    "sub": "as reported on the v1 GGUF card"
+   },
+   "Jev-Style 2B v2": {
+    "main": "1.27 GB",
+    "sub": "as reported on the v2 GGUF card"
+   },
+   "Jev-Style 0.8B v3": {
+    "main": "0.53 GB",
+    "sub": "matches FP32 on 240/240 parity rows"
+   }
+  },
+  {
+   "row": "Typed decisions, teacher agreement",
+   "Jev-Style 2B v1": {
+    "main": "53.35%",
+    "sub": "2,000 decisions / 400 states"
+   },
+   "Jev-Style 2B v2": {
+    "main": "73.45%",
+    "sub": "same 2,000 decisions"
+   },
+   "Jev-Style 0.8B v3": {
+    "main": "79.15%",
+    "sub": "same 2,000 \u00b7 1,583 correct"
+   }
+  }
+ ],
+ "numbers": {
+  "v1_q4_k_m_agreement": 0.944,
+  "v2_q4_k_m_agreement": 0.914,
+  "v3_q4_k_m_agreement": {
+   "agree": 240,
+   "n": 240,
+   "value": 1.0
+  },
+  "typed_teacher_agreement": {
+   "v1": 0.5335,
+   "v2": 0.7345,
+   "v3": 0.7915,
+   "v3_correct": 1583,
+   "n": 2000,
+   "states": 400
+  },
+  "params": {
+   "v1": "2B",
+   "v2": "2B",
+   "v3_text_model": 752393024
+  },
+  "size_ratio_v3_over_2b": 0.4,
+  "context_ratio_v3_over_v2_prompt": 25.0
+ },
+ "sources": {
+  "v1": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF/raw/main/README.md (file table 'Same decision as bf16' Q4_K_M 94.4%; 'LoRA rank 16 on all linear layers'; 'Up to 26 options (20 when ... top_logprobs)'; 'Trained on five English task families'; quickstart llama-server without -c; prompt has one [Question])",
+  "v2": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/raw/main/README.md (Q4_K_M 91.4% choice agreement vs CUDA BF16 on frozen 500-decision subset; '2-26 unique options, within a 1,024-token prompt'; 'English state'; quickstart -c 2048; rank-32 LoRA, 33,638,400 trainable; continued from v1; ' A' through ' Z' readout; typed-decisions 53.35% v1 / 73.45% v2, 2,000 decisions from 400 states)",
+  "v3": "chart_data.json: design.generations (export_manifest.json parameters_text_model=752393024; config.resolved.json readout=verdict, no LoRA keys; jevbench/results.json scorer.max_len=25600; apps.jsonl banking77_full 400 rows x 77 options; scoreboard long_grid_plus claim_25k_ok=true), quant.v3 gguf-q4_k_m 240/240 (export_r2/main/parity/report.json), typed.accuracy.v3 1583/2000 (typed_test.jsonl, 400 group_ids)"
+ },
+ "footnote": "v1/v2: as reported on their public Hugging Face cards (v1 GGUF card; v2 and v2-GGUF cards; v1's typed-decisions number is reported on the v2 card). v3: release manifest, training config and eval files; Q4_K_M size = exported file (GB = 10^9 bytes), parity rows drawn from the training pool. Typed decisions: same 2,000 decisions from 400 states; v1/v2 scored by the v2 card's harness, v3 by ours. In-domain for v3; v1 not trained on typed decisions; v2's pool included typed workflow decisions. '51 evaluated' = MASSIVE locales (14 trained, 37 held out); fine-tuning covers 19 languages. 25,600 tokens = the runtime's whole-input limit; 25x = vs v2's 1,024-token prompt."
+}

figures/headline_typed.data.json

@@ -0,0 +1,90 @@
+{
+ "figure": "headline_typed",
+ "panels": {
+  "accuracy_pct": [
+   {
+    "label": "Jev-Style 2B v1",
+    "entry": "typed.teacher_agreement.v1",
+    "raw": 0.5335,
+    "plotted": 53.4,
+    "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2/raw/main/README.md",
+    "field": "card text: 'The separate typed-decisions group contains 2,000 teacher-reference decisions from 400 states. Teacher agreement is 53.35% for v1, 37.55% for English Laya and 73.45% for v2'"
+   },
+   {
+    "label": "Jev",
+    "entry": "typed.accuracy.jev",
+    "raw": 0.727,
+    "plotted": 72.7,
+    "source": "docs/round2_audit/track_jev_results.json",
+    "field": "string 'Jev typed-decisions (zero-shot)': '0.727 acc; KL 1.442; Brier 0.148; ECE 0.144; soft-acc 0.580' (from LocalLLaMA/typed-decisions dataset card / Laya HF card)"
+   },
+   {
+    "label": "Jev-Style 2B v2",
+    "entry": "typed.teacher_agreement.v2",
+    "raw": 0.7345,
+    "plotted": 73.5,
+    "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2/raw/main/README.md",
+    "field": "card text: 'The separate typed-decisions group contains 2,000 teacher-reference decisions from 400 states. Teacher agreement is 53.35% for v1, 37.55% for English Laya and 73.45% for v2'"
+   },
+   {
+    "label": "Laya (typed ckpt)",
+    "entry": "typed.accuracy.laya_typed",
+    "raw": 0.766,
+    "plotted": 76.6,
+    "source": "runs/macjev/report_r2/scoreboard/scoreboard.json",
+    "field": "metrics['typed.accuracy'].laya.typed.value"
+   },
+   {
+    "label": "Jev-Style 0.8B v3",
+    "entry": "typed.accuracy.v3",
+    "raw": 0.7915,
+    "plotted": 79.2,
+    "source": "runs/macjev/report_r2/scoreboard/scoreboard.json",
+    "field": "metrics['typed.accuracy'].ours_value"
+   }
+  ],
+  "brier_vs_soft": [
+   {
+    "label": "Jev",
+    "entry": "typed.Brier_vs_soft_labels.jev",
+    "raw": 0.148,
+    "plotted": 0.148,
+    "source": "docs/round2_audit/track_jev_results.json",
+    "field": "string 'Jev typed-decisions (zero-shot)': '0.727 acc; KL 1.442; Brier 0.148; ECE 0.144; soft-acc 0.580' (from LocalLLaMA/typed-decisions dataset card / Laya HF card)"
+   },
+   {
+    "label": "Laya (typed ckpt)",
+    "entry": "typed.brier_vs_soft.laya_typed",
+    "raw": 0.06146485330135357,
+    "plotted": 0.061,
+    "source": "runs/macjev/report_r2/scoreboard/scoreboard.json",
+    "field": "metrics['typed.brier_vs_soft'].laya.typed.value"
+   },
+   {
+    "label": "Jev-Style 0.8B v3",
+    "entry": "typed.brier_vs_soft.v3",
+    "raw": 0.04583493309263009,
+    "plotted": 0.046,
+    "source": "runs/macjev/report_r2/scoreboard/scoreboard.json",
+    "field": "metrics['typed.brier_vs_soft'].ours_value"
+   }
+  ]
+ },
+ "annotations": {
+  "delta_vs_jev_pts": 6.4,
+  "delta_vs_v2_pts": 5.7,
+  "delta_vs_laya_typed_pts": 2.6,
+  "brier_ratio_jev_over_v3": 3.229,
+  "brier_pct_lower_than_laya_typed": 25.4,
+  "v3_acc_ci95_wilson": [
+   0.7731,
+   0.8087
+  ],
+  "v3_minus_laya_typed_paired_ci95": [
+   0.010499999999999954,
+   0.04249999999999998
+  ]
+ },
+ "protocol": "in-domain for v3 and Laya typed; zero-shot for Jev (dataset card); 2B v1/v2 as reported on the v2 card",
+ "footnote": "Typed-decisions test set (LocalLLaMA/typed-decisions), 2,000 decisions from 400 states. In-domain for v3 and Laya typed (both trained on its\ntrain split); zero-shot for Jev (dataset-card numbers, Jev API, all 2,000 decisions). Laya: official typed-decisions checkpoint re-run by us on\nidentical rows with its shipped temperature. 2B v1/v2: teacher agreement as reported on the v2 card (same 2,000 decisions, that card's harness;\nv1 was not trained on typed decisions, v2's pool included typed workflow decisions). v3 95% CI 77.3\u201380.9% (Wilson);\nv3 minus Laya typed, paired bootstrap 95% CI +1.0 to +4.2 pts. Plotted values: figures/headline_typed.data.json."
+}

figures/jevbench.data.json

@@ -0,0 +1,60 @@
+{
+ "chart": "jevbench",
+ "metric": "JevBench v1.4.1 public accuracy (231 items)",
+ "rows": [
+  {
+   "label": "Jev-Style 0.8B v3",
+   "key": "v3",
+   "accuracy": 0.6406926406926406,
+   "correct": 148,
+   "n": 231,
+   "colour": "#3456F0",
+   "source": "runs/macjev/received/ext_evals/main/jevbench/results.json :: public_accuracy"
+  },
+  {
+   "label": "Qwen3.5-0.8B Decision Model (M. Ghafiri)",
+   "key": "mghafiri-qwen3.5-0.8b-decision-model",
+   "accuracy": 0.5930735930735931,
+   "correct": 137,
+   "n": 231,
+   "colour": "#C4C4CB",
+   "source": "data/external/jevbench_v1.4.1/results/v1.4.1/jevbench-v1.4.1-results.json :: systems[key=mghafiri-qwen3.5-0.8b-decision-model].public_accuracy",
+   "v3_lead_pts": 4.76
+  },
+  {
+   "label": "Laya (ModernBERT-large, 421M)",
+   "key": "laya",
+   "accuracy": 0.5844155844155844,
+   "correct": 135,
+   "n": 231,
+   "colour": "#D0D0D6",
+   "source": "data/external/jevbench_v1.4.1/results/v1.4.1/jevbench-v1.4.1-results.json :: systems[key=laya].public_accuracy",
+   "v3_lead_pts": 5.63
+  },
+  {
+   "label": "SimpleJev (Qwen3.5-0.8B)",
+   "key": "simplejev-qwen3.5-0.8b",
+   "accuracy": 0.5454545454545454,
+   "correct": 126,
+   "n": 231,
+   "colour": "#C4C4CB",
+   "source": "data/external/jevbench_v1.4.1/results/v1.4.1/jevbench-v1.4.1-results.json :: systems[key=simplejev-qwen3.5-0.8b].public_accuracy",
+   "v3_lead_pts": 9.52
+  }
+ ],
+ "footnote": "JevBench v1.4.1, public items only (231). v3: self-run zero-shot with the vendored official harness (commit 24b9b5c); training-pool contamination scan: 0 hits; not an official leaderboard entry. Other rows: public accuracy as published in the board's v1.4.1 results file (github.com/fstandhartinger/jevbench). Shown: Laya plus every Qwen3.5-0.8B-based system on the board. Laya and M. Ghafiri lie inside v3's 95% CI (57.7\u201370.0%, Wilson): those two leads are point estimates, not significant at n = 231.",
+ "v3_run": {
+  "source_path": "runs/macjev/received/ext_evals/main/jevbench/results.json",
+  "field": "public_accuracy (148/231; independent check equal)",
+  "protocol_label": "v3 self-run with the vendored official harness on the 231 public items, zero-shot (contamination scan of the training pool: 0 hits, runs/macjev/external/jevbench/contamination.json); other rows as published in jevbench v1.4.1 results.json (full runs by the board); not an official leaderboard entry",
+  "ci95": [
+   0.577,
+   0.6998
+  ],
+  "ci_method": "Wilson 95%"
+ },
+ "board": {
+  "source_path": "data/external/jevbench_v1.4.1/results/v1.4.1/jevbench-v1.4.1-results.json",
+  "protocol_label": "as published by the board (commit 24b9b5c1609a, tag v1.4.1)"
+ }
+}

figures/latency.data.json

@@ -0,0 +1,51 @@
+{
+ "chart": "latency",
+ "from_chart_data_entry": "latency.idle",
+ "source_path": "runs/macjev/latency/m1max_untrained_base_2026-09-23_review_idle/latency.json",
+ "field": "results['llamacpp-f16'|'laya-multilingual-mps-fp32'][cell].p50_ms",
+ "protocol_label": "M1 Max 64 GB, warm end-to-end p50 ms, idle run 2026-09-23 (load avg 3-8); untrained identical-architecture Qwen3.5-0.8B export (latency does not depend on weights); v3 = llama.cpp GGUF F16, one call per state; comparison engine = round-1 MacLaya-4K (our Laya-multilingual fine-tune, FP32/MPS, 4,096-token budget), one decide() per question; prefix reuse off",
+ "plotted": [
+  {
+   "cell": "1024x5",
+   "v3_p50_ms": 394,
+   "maclaya4k_p50_ms": 543,
+   "speedup_label": "1.4x"
+  },
+  {
+   "cell": "1024x10",
+   "v3_p50_ms": 525,
+   "maclaya4k_p50_ms": 1019,
+   "speedup_label": "1.9x"
+  },
+  {
+   "cell": "4096x5",
+   "v3_p50_ms": 1212,
+   "maclaya4k_p50_ms": 3204,
+   "speedup_label": "2.6x"
+  },
+  {
+   "cell": "4096x10",
+   "v3_p50_ms": 1381,
+   "maclaya4k_p50_ms": 6364,
+   "speedup_label": "4.6x"
+  },
+  {
+   "cell": "8192x1",
+   "v3_p50_ms": 2295,
+   "maclaya4k_p50_ms": null,
+   "speedup_label": "v3 only"
+  },
+  {
+   "cell": "8192x5",
+   "v3_p50_ms": 2424,
+   "maclaya4k_p50_ms": null,
+   "speedup_label": "v3 only"
+  },
+  {
+   "cell": "8192x10",
+   "v3_p50_ms": 2613,
+   "maclaya4k_p50_ms": null,
+   "speedup_label": "v3 only"
+  }
+ ]
+}

figures/long_context.json

@@ -0,0 +1,216 @@
+{
+ "figure": "long_context",
+ "plotted": [
+  {
+   "length_bin": 1024,
+   "items": 320,
+   "controlled_pct": 57.81,
+   "ci95_pct": [
+    52.34,
+    63.1
+   ]
+  },
+  {
+   "length_bin": 2048,
+   "items": 320,
+   "controlled_pct": 54.69,
+   "ci95_pct": [
+    49.21,
+    60.05
+   ]
+  },
+  {
+   "length_bin": 4096,
+   "items": 320,
+   "controlled_pct": 58.13,
+   "ci95_pct": [
+    52.65,
+    63.4
+   ]
+  },
+  {
+   "length_bin": 8192,
+   "items": 320,
+   "controlled_pct": 54.37,
+   "ci95_pct": [
+    48.9,
+    59.75
+   ]
+  },
+  {
+   "length_bin": 12288,
+   "items": 320,
+   "controlled_pct": 54.37,
+   "ci95_pct": [
+    48.9,
+    59.75
+   ]
+  },
+  {
+   "length_bin": 16384,
+   "items": 320,
+   "controlled_pct": 54.69,
+   "ci95_pct": [
+    49.21,
+    60.05
+   ]
+  },
+  {
+   "length_bin": 24576,
+   "items": 1280,
+   "controlled_pct": 55.31,
+   "ci95_pct": [
+    52.58,
+    58.02
+   ]
+  }
+ ],
+ "reference_controlled_2k_4k_pct": 56.41,
+ "gap_24k_vs_ref_pts": -1.09,
+ "spread_all_bins_pts": 3.75,
+ "claim_25k_ok": true,
+ "a_gap_ok": true,
+ "b_deciles_ok": true,
+ "c_controls_ok": true,
+ "verified_length_bin": 4096,
+ "caution": "verified_length_bin = 4096: the stricter every-bin (8K-16K) decile criterion fails in some deciles; the card may say 'preregistered 25K claim passed' but not 'verified at every length'",
+ "controlled_24k_by_decile": {
+  "0": 0.5234375,
+  "1": 0.515625,
+  "2": 0.546875,
+  "3": 0.5390625,
+  "4": 0.5546875,
+  "5": 0.5546875,
+  "6": 0.5703125,
+  "7": 0.5625,
+  "8": 0.5390625,
+  "9": 0.625
+ },
+ "controls_pooled_by_length": {
+  "1024": {
+   "question_only": {
+    "n": 320,
+    "accuracy": 0.290625,
+    "chance": 0.28003348214285717,
+    "p_above_chance": 0.3511374281963401,
+    "at_chance": true
+   },
+   "state_swap": {
+    "n": 320,
+    "accuracy": 0.315625,
+    "chance": 0.28003348214285717,
+    "p_above_chance": 0.07484199516389486,
+    "at_chance": true
+   }
+  },
+  "2048": {
+   "question_only": {
+    "n": 320,
+    "accuracy": 0.2875,
+    "chance": 0.28029017857142857,
+    "p_above_chance": 0.40558180434986946,
+    "at_chance": true
+   },
+   "state_swap": {
+    "n": 320,
+    "accuracy": 0.303125,
+    "chance": 0.28029017857142857,
+    "p_above_chance": 0.18406466252955953,
+    "at_chance": true
+   }
+  },
+  "4096": {
+   "question_only": {
+    "n": 320,
+    "accuracy": 0.253125,
+    "chance": 0.28010044642857146,
+    "p_above_chance": 0.8864307901965104,
+    "at_chance": true
+   },
+   "state_swap": {
+    "n": 320,
+    "accuracy": 0.29375,
+    "chance": 0.28010044642857146,
+    "p_above_chance": 0.304486548332512,
+    "at_chance": true
+   }
+  },
+  "8192": {
+   "question_only": {
+    "n": 320,
+    "accuracy": 0.28125,
+    "chance": 0.28131324404761904,
+    "p_above_chance": 0.5273741085290076,
+    "at_chance": true
+   },
+   "state_swap": {
+    "n": 320,
+    "accuracy": 0.284375,
+    "chance": 0.28131324404761904,
+    "p_above_chance": 0.4747527191420883,
+    "at_chance": true
+   }
+  },
+  "12288": {
+   "question_only": {
+    "n": 320,
+    "accuracy": 0.278125,
+    "chance": 0.2804017857142857,
+    "p_above_chance": 0.5644923218252806,
+    "at_chance": true
+   },
+   "state_swap": {
+    "n": 320,
+    "accuracy": 0.284375,
+    "chance": 0.2804017857142857,
+    "p_above_chance": 0.4593971620159274,
+    "at_chance": true
+   }
+  },
+  "16384": {
+   "question_only": {
+    "n": 320,
+    "accuracy": 0.265625,
+    "chance": 0.28533854166666667,
+    "p_above_chance": 0.8139517721332229,
+    "at_chance": true
+   },
+   "state_swap": {
+    "n": 320,
+    "accuracy": 0.28125,
+    "chance": 0.28533854166666667,
+    "p_above_chance": 0.5936977453471026,
+    "at_chance": true
+   }
+  },
+  "24576": {
+   "question_only": {
+    "n": 1280,
+    "accuracy": 0.2890625,
+    "chance": 0.2837481398809524,
+    "p_above_chance": 0.33937799270227736,
+    "at_chance": true
+   },
+   "state_swap": {
+    "n": 1280,
+    "accuracy": 0.28515625,
+    "chance": 0.2837481398809524,
+    "p_above_chance": 0.4658977510199284,
+    "at_chance": true
+   }
+  }
+ },
+ "laya_budgets": {
+  "laya_english": 512,
+  "laya_multilingual": 1024,
+  "laya_typed_decisions": 1024,
+  "source": "third_party/laya/README.md lines 366-368 (defaults); models/laya-official/.../multilingual/rl_agent_config.json max_len 1024"
+ },
+ "v3_context_tokens": 25600,
+ "sources": [
+  "runs/macjev/received/own_evals/main/test.own/long_grid_plus.jsonl + runs/macjev/received/own_evals/main/temperatures.json via macjev.eval.aggregate.suite_metrics('long_grid_plus') (recomputed); matches runs/macjev/report_r2/scoreboard/scoreboard.json long_grid_plus",
+  "chart_data.json#long_grid_plus.bins"
+ ],
+ "protocol_label": "v3 only (Laya cannot run >4K; not applicable). Controlled = real row correct AND question_only + state_swap controls pass; Wilson CI; 2K-4K reference; preregistered 2026-09-24, amended pre-data (+96 items per 24K decile, thresholds unchanged)",
+ "footnote": "v3 only. Laya's default input budget is 512 tokens (English) / 1,024 (multilingual, typed) per the Laya README;\nlonger rows exceed Laya's default budgets, so Laya is not plotted. Suite long_grid_plus: preregistered 2026-09-24, amended pre-data\n(+96 items per 24K depth decile, thresholds unchanged); 320 items per bin, 1,280 at 24K. Controlled accuracy = real-state\nanswer correct AND its question-only and state-swap controls pass; both controls are at chance in every length bin.\n25K claim rule: |24K \u2212 2K\u20134K reference| \u2264 5 pts and every 24K evidence-depth decile within 10 pts of it.\nPlotted values: figures/long_context.json (bin-level controlled accuracy and CIs)."
+}

figures/multilingual.data.json

@@ -0,0 +1,486 @@
+{
+ "chart": "multilingual",
+ "metric": "per-language MASSIVE intent accuracy, argmax over 20 scored options, 100 rows/language",
+ "sources": {
+  "v3": "runs/macjev/received/20260924-0313/evals/main/test/massive51.jsonl",
+  "laya_multilingual": "runs/macjev/laya_baselines/multilingual/massive51.jsonl",
+  "chart_data_entry": "chart_data.json#massive51.per_language (re-verified against raw files)"
+ },
+ "protocol": "14 trained + 37 locales held out of MASSIVE training for v3; held-out for all Laya checkpoints. Laya numbers: official checkpoints re-run by us on identical rows, shipped temperatures. Wilson 95% CI per language (n=100)",
+ "chance": 0.05,
+ "three_x_chance": 0.15000000000000002,
+ "summary": {
+  "languages": 51,
+  "v3_beats_laya_multilingual": 51,
+  "v3_beats_best_laya_checkpoint": 51,
+  "v3_beats_laya_multilingual_never_trained": 37,
+  "v3_above_3x_chance": 51,
+  "laya_multilingual_above_3x_chance": 48,
+  "macro_v3": 0.7174509803921569,
+  "macro_laya_multilingual": 0.4007843137254902,
+  "macro_v3_never_trained37": 0.6551351351351351,
+  "macro_laya_multilingual_never_trained37": 0.36108108108108106,
+  "min_margin_vs_laya_multilingual": 0.11
+ },
+ "rows_plotted_sorted": [
+  {
+   "lang": "zh-CN",
+   "name": "Chinese (CN)",
+   "trained": true,
+   "v3": 0.95,
+   "laya_multilingual": 0.65,
+   "best_laya": 0.65,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "ja",
+   "name": "Japanese",
+   "trained": true,
+   "v3": 0.95,
+   "laya_multilingual": 0.64,
+   "best_laya": 0.64,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "es",
+   "name": "Spanish",
+   "trained": true,
+   "v3": 0.95,
+   "laya_multilingual": 0.58,
+   "best_laya": 0.58,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "ru",
+   "name": "Russian",
+   "trained": true,
+   "v3": 0.94,
+   "laya_multilingual": 0.57,
+   "best_laya": 0.57,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "en",
+   "name": "English",
+   "trained": true,
+   "v3": 0.93,
+   "laya_multilingual": 0.71,
+   "best_laya": 0.82,
+   "best_laya_ckpt": "english"
+  },
+  {
+   "lang": "zh-TW",
+   "name": "Chinese (TW)",
+   "trained": false,
+   "v3": 0.93,
+   "laya_multilingual": 0.61,
+   "best_laya": 0.61,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "de",
+   "name": "German",
+   "trained": true,
+   "v3": 0.93,
+   "laya_multilingual": 0.5,
+   "best_laya": 0.5,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "fr",
+   "name": "French",
+   "trained": true,
+   "v3": 0.9,
+   "laya_multilingual": 0.6,
+   "best_laya": 0.6,
+   "best_laya_ckpt": "typed"
+  },
+  {
+   "lang": "it",
+   "name": "Italian",
+   "trained": false,
+   "v3": 0.89,
+   "laya_multilingual": 0.52,
+   "best_laya": 0.52,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "pl",
+   "name": "Polish",
+   "trained": false,
+   "v3": 0.89,
+   "laya_multilingual": 0.5,
+   "best_laya": 0.5,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "ko",
+   "name": "Korean",
+   "trained": true,
+   "v3": 0.89,
+   "laya_multilingual": 0.47,
+   "best_laya": 0.47,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "pt",
+   "name": "Portuguese",
+   "trained": true,
+   "v3": 0.88,
+   "laya_multilingual": 0.5,
+   "best_laya": 0.5,
+   "best_laya_ckpt": "typed"
+  },
+  {
+   "lang": "hi",
+   "name": "Hindi",
+   "trained": true,
+   "v3": 0.87,
+   "laya_multilingual": 0.46,
+   "best_laya": 0.46,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "sv",
+   "name": "Swedish",
+   "trained": false,
+   "v3": 0.85,
+   "laya_multilingual": 0.49,
+   "best_laya": 0.49,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "ar",
+   "name": "Arabic",
+   "trained": true,
+   "v3": 0.85,
+   "laya_multilingual": 0.46,
+   "best_laya": 0.46,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "vi",
+   "name": "Vietnamese",
+   "trained": false,
+   "v3": 0.85,
+   "laya_multilingual": 0.34,
+   "best_laya": 0.34,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "fa",
+   "name": "Persian",
+   "trained": false,
+   "v3": 0.84,
+   "laya_multilingual": 0.51,
+   "best_laya": 0.51,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "id",
+   "name": "Indonesian",
+   "trained": false,
+   "v3": 0.84,
+   "laya_multilingual": 0.51,
+   "best_laya": 0.51,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "tr",
+   "name": "Turkish",
+   "trained": true,
+   "v3": 0.83,
+   "laya_multilingual": 0.4,
+   "best_laya": 0.4,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "nl",
+   "name": "Dutch",
+   "trained": false,
+   "v3": 0.82,
+   "laya_multilingual": 0.47,
+   "best_laya": 0.47,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "nb",
+   "name": "Norwegian",
+   "trained": false,
+   "v3": 0.81,
+   "laya_multilingual": 0.56,
+   "best_laya": 0.56,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "da",
+   "name": "Danish",
+   "trained": false,
+   "v3": 0.81,
+   "laya_multilingual": 0.52,
+   "best_laya": 0.52,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "ro",
+   "name": "Romanian",
+   "trained": false,
+   "v3": 0.77,
+   "laya_multilingual": 0.37,
+   "best_laya": 0.37,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "af",
+   "name": "Afrikaans",
+   "trained": false,
+   "v3": 0.76,
+   "laya_multilingual": 0.35,
+   "best_laya": 0.35,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "ta",
+   "name": "Tamil",
+   "trained": true,
+   "v3": 0.74,
+   "laya_multilingual": 0.31,
+   "best_laya": 0.31,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "sw",
+   "name": "Swahili",
+   "trained": true,
+   "v3": 0.74,
+   "laya_multilingual": 0.23,
+   "best_laya": 0.23,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "ms",
+   "name": "Malay",
+   "trained": false,
+   "v3": 0.73,
+   "laya_multilingual": 0.41,
+   "best_laya": 0.41,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "el",
+   "name": "Greek",
+   "trained": false,
+   "v3": 0.72,
+   "laya_multilingual": 0.44,
+   "best_laya": 0.44,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "he",
+   "name": "Hebrew",
+   "trained": false,
+   "v3": 0.72,
+   "laya_multilingual": 0.37,
+   "best_laya": 0.37,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "ur",
+   "name": "Urdu",
+   "trained": false,
+   "v3": 0.7,
+   "laya_multilingual": 0.42,
+   "best_laya": 0.42,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "az",
+   "name": "Azerbaijani",
+   "trained": false,
+   "v3": 0.66,
+   "laya_multilingual": 0.36,
+   "best_laya": 0.36,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "bn",
+   "name": "Bengali",
+   "trained": false,
+   "v3": 0.65,
+   "laya_multilingual": 0.45,
+   "best_laya": 0.45,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "sl",
+   "name": "Slovenian",
+   "trained": false,
+   "v3": 0.65,
+   "laya_multilingual": 0.37,
+   "best_laya": 0.37,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "th",
+   "name": "Thai",
+   "trained": false,
+   "v3": 0.64,
+   "laya_multilingual": 0.48,
+   "best_laya": 0.48,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "hu",
+   "name": "Hungarian",
+   "trained": false,
+   "v3": 0.63,
+   "laya_multilingual": 0.36,
+   "best_laya": 0.36,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "te",
+   "name": "Telugu",
+   "trained": false,
+   "v3": 0.62,
+   "laya_multilingual": 0.22,
+   "best_laya": 0.22,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "kn",
+   "name": "Kannada",
+   "trained": false,
+   "v3": 0.61,
+   "laya_multilingual": 0.3,
+   "best_laya": 0.3,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "fi",
+   "name": "Finnish",
+   "trained": false,
+   "v3": 0.6,
+   "laya_multilingual": 0.34,
+   "best_laya": 0.34,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "jv",
+   "name": "Javanese",
+   "trained": false,
+   "v3": 0.6,
+   "laya_multilingual": 0.3,
+   "best_laya": 0.3,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "hy",
+   "name": "Armenian",
+   "trained": false,
+   "v3": 0.59,
+   "laya_multilingual": 0.25,
+   "best_laya": 0.25,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "ka",
+   "name": "Georgian",
+   "trained": false,
+   "v3": 0.56,
+   "laya_multilingual": 0.15,
+   "best_laya": 0.15,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "tl",
+   "name": "Tagalog",
+   "trained": false,
+   "v3": 0.54,
+   "laya_multilingual": 0.35,
+   "best_laya": 0.35,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "km",
+   "name": "Khmer",
+   "trained": false,
+   "v3": 0.54,
+   "laya_multilingual": 0.2,
+   "best_laya": 0.2,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "sq",
+   "name": "Albanian",
+   "trained": false,
+   "v3": 0.52,
+   "laya_multilingual": 0.3,
+   "best_laya": 0.3,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "ml",
+   "name": "Malayalam",
+   "trained": false,
+   "v3": 0.52,
+   "laya_multilingual": 0.28,
+   "best_laya": 0.28,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "lv",
+   "name": "Latvian",
+   "trained": false,
+   "v3": 0.51,
+   "laya_multilingual": 0.31,
+   "best_laya": 0.31,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "is",
+   "name": "Icelandic",
+   "trained": false,
+   "v3": 0.5,
+   "laya_multilingual": 0.35,
+   "best_laya": 0.35,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "my",
+   "name": "Burmese",
+   "trained": false,
+   "v3": 0.46,
+   "laya_multilingual": 0.16,
+   "best_laya": 0.16,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "cy",
+   "name": "Welsh",
+   "trained": false,
+   "v3": 0.37,
+   "laya_multilingual": 0.13,
+   "best_laya": 0.17,
+   "best_laya_ckpt": "typed"
+  },
+  {
+   "lang": "mn",
+   "name": "Mongolian",
+   "trained": false,
+   "v3": 0.28,
+   "laya_multilingual": 0.16,
+   "best_laya": 0.16,
+   "best_laya_ckpt": "multilingual"
+  },
+  {
+   "lang": "am",
+   "name": "Amharic",
+   "trained": false,
+   "v3": 0.26,
+   "laya_multilingual": 0.15,
+   "best_laya": 0.16,
+   "best_laya_ckpt": "typed"
+  }
+ ]
+}

figures/quantization.data.json

@@ -0,0 +1,151 @@
+{
+ "chart": "quantization",
+ "metric": "top-1 agreement with the full-precision reference (%), plus file size (GB = bytes/1e9)",
+ "rows": [
+  {
+   "tier": "16-bit",
+   "model": "v3",
+   "label": "Jev-Style 0.8B v3 \u00b7 GGUF F16",
+   "agree": 100.0,
+   "n": 240,
+   "agree_count": 240,
+   "size_gb": 1.516744128,
+   "size_note": "local export file (bytes / 1e9)",
+   "reference": "torch FP32",
+   "source": "runs/macjev/export_r2/main/parity/report.json; sizes: chart_data.json export.sizes"
+  },
+  {
+   "tier": "16-bit",
+   "model": "v3",
+   "label": "Jev-Style 0.8B v3 \u00b7 MLX bf16",
+   "agree": 100.0,
+   "n": 240,
+   "agree_count": 240,
+   "size_gb": 1.504827355,
+   "size_note": "local export file (bytes / 1e9)",
+   "reference": "torch FP32",
+   "source": "runs/macjev/export_r2/main/parity/report.json; sizes: chart_data.json export.sizes"
+  },
+  {
+   "tier": "16-bit",
+   "model": "v2",
+   "label": "Jev-Style 2B v2 \u00b7 GGUF BF16",
+   "agree": 99.6,
+   "n": 500,
+   "size_gb": 3.78,
+   "size_note": "as reported on card",
+   "reference": "CUDA merged BF16",
+   "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/raw/main/README.md"
+  },
+  {
+   "tier": "16-bit",
+   "model": "v2",
+   "label": "Jev-Style 2B v2 \u00b7 MLX BF16",
+   "agree": 99.6,
+   "n": 500,
+   "size_gb": 3.76,
+   "size_note": "as reported on card",
+   "reference": "CUDA merged BF16",
+   "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/raw/main/README.md (table: 'Native MLX BF16 | 3.76 GB | 99.6% choice agreement')"
+  },
+  {
+   "tier": "16-bit",
+   "model": "v1",
+   "label": "Jev-Style 2B v1 \u00b7 GGUF BF16",
+   "agree": 99.8,
+   "n": 500,
+   "size_gb": 3.9,
+   "size_note": "as reported on card",
+   "reference": "bf16",
+   "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF/raw/main/README.md"
+  },
+  {
+   "tier": "8-bit",
+   "model": "v3",
+   "label": "Jev-Style 0.8B v3 \u00b7 GGUF Q8_0",
+   "agree": 100.0,
+   "n": 240,
+   "agree_count": 240,
+   "size_gb": 0.811843008,
+   "size_note": "local export file (bytes / 1e9)",
+   "reference": "torch FP32",
+   "source": "runs/macjev/export_r2/main/parity/report.json; sizes: chart_data.json export.sizes"
+  },
+  {
+   "tier": "8-bit",
+   "model": "v3",
+   "label": "Jev-Style 0.8B v3 \u00b7 MLX 8-bit",
+   "agree": 100.0,
+   "n": 240,
+   "agree_count": 240,
+   "size_gb": 0.799973748,
+   "size_note": "local export file (bytes / 1e9)",
+   "reference": "torch FP32",
+   "source": "runs/macjev/export_r2/main/parity/report.json; sizes: chart_data.json export.sizes"
+  },
+  {
+   "tier": "8-bit",
+   "model": "v2",
+   "label": "Jev-Style 2B v2 \u00b7 GGUF Q8_0",
+   "agree": 99.2,
+   "n": 500,
+   "size_gb": 2.01,
+   "size_note": "as reported on card",
+   "reference": "CUDA merged BF16",
+   "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/raw/main/README.md"
+  },
+  {
+   "tier": "8-bit",
+   "model": "v1",
+   "label": "Jev-Style 2B v1 \u00b7 GGUF Q8_0",
+   "agree": 99.4,
+   "n": 500,
+   "size_gb": 2.1,
+   "size_note": "as reported on card",
+   "reference": "bf16",
+   "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF/raw/main/README.md"
+  },
+  {
+   "tier": "4-bit",
+   "model": "v3",
+   "label": "Jev-Style 0.8B v3 \u00b7 GGUF Q4_K_M",
+   "agree": 100.0,
+   "n": 240,
+   "agree_count": 240,
+   "size_gb": 0.529296832,
+   "size_note": "local export file (bytes / 1e9)",
+   "reference": "torch FP32",
+   "source": "runs/macjev/export_r2/main/parity/report.json; sizes: chart_data.json export.sizes"
+  },
+  {
+   "tier": "4-bit",
+   "model": "v2",
+   "label": "Jev-Style 2B v2 \u00b7 GGUF Q4_K_M",
+   "agree": 91.4,
+   "n": 500,
+   "size_gb": 1.27,
+   "size_note": "as reported on card",
+   "reference": "CUDA merged BF16",
+   "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/raw/main/README.md"
+  },
+  {
+   "tier": "4-bit",
+   "model": "v1",
+   "label": "Jev-Style 2B v1 \u00b7 GGUF Q4_K_M",
+   "agree": 94.39999999999999,
+   "n": 500,
+   "size_gb": 1.3,
+   "size_note": "as reported on card",
+   "reference": "bf16",
+   "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF/raw/main/README.md"
+  }
+ ],
+ "deltas": {
+  "v2 Q4_K_M size / v3 Q4_K_M size": 2.4
+ },
+ "no_agreement_delta_claimed": "v3 parity rows are training-pool rows; v1/v2 used 500 held-out decisions, so no agreement gap is claimed",
+ "long_points": "v3 formats also 100% top-1 at 16,384 and 25,600 tokens (3 rows each), parity report long_points",
+ "not_plotted": "mlx-4bit (98.75%, 237/240) is not a published format and is omitted",
+ "footnote": "v3: top-1 agreement with the PyTorch FP32 reference on a 240-row parity fixture drawn from the training pool (22 categories, en+zh), plus 6 extra rows at about 16K and 25.6K tokens (6/6 agree). v3 sizes = exported weight files (GB = 10^9 bytes). 2B v1/v2: as reported on their public HF GGUF cards (500 held-out decisions each; vs bf16 for v1, vs CUDA merged BF16 for v2; card sizes). Different fixtures (training-pool rows for v3, held-out rows for v1/v2) and references: rows are not a paired comparison. x-axis starts at 88%.",
+ "protocol_label": "v3 G5 parity: 240-row mixed fixture (drawn from training-pool rows, 22 categories, en+zh) vs torch FP32; plus 16K and 25.6K long points (3 rows each). 2B numbers are from their cards on different fixtures (500 decisions vs bf16)"
+}

figures/zeroshot.json

@@ -0,0 +1,58 @@
+{
+ "chart": "zeroshot",
+ "metric": "accuracy (argmax over the set's labels; all rows, none unsupported)",
+ "plotted": {
+  "tweet_topic": {
+   "v3": 75.49,
+   "laya_en": 63.2,
+   "delta_pts_vs_laya_en": 12.29
+  },
+  "fin_topic": {
+   "v3": 46.71,
+   "laya_en": 34.2,
+   "delta_pts_vs_laya_en": 12.51
+  }
+ },
+ "jev_marker": {
+  "set": "tweet_topic",
+  "jev": 79.33,
+  "gap_pts": 3.84
+ },
+ "v3_recomputed": {
+  "tweet_topic": {
+   "correct": 1278,
+   "n": 1693,
+   "ci95": [
+    0.7341996455995274,
+    0.7749556999409333
+   ]
+  },
+  "fin_topic": {
+   "correct": 1923,
+   "n": 4117,
+   "ci95": [
+    0.45202817585620597,
+    0.48239008987126547
+   ]
+  }
+ },
+ "entries": [
+  "zeroshot.tweet_topic.accuracy.v3",
+  "zeroshot.tweet_topic.accuracy.laya_en",
+  "zeroshot.tweet_topic.accuracy.jev",
+  "zeroshot.fin_topic.accuracy.v3",
+  "zeroshot.fin_topic.accuracy.laya_en",
+  "zeroshot.fin_topic.accuracy.jev"
+ ],
+ "claims": [
+  "tweet_topic_accuracy_vs_laya_en",
+  "fin_topic_accuracy_vs_laya_en",
+  "tweet_topic_accuracy_vs_jev"
+ ],
+ "sources": {
+  "v3": "runs/macjev/received/ext_evals/main/zeroshot_topics/metrics.json comparison.clean[*].ours.accuracy; recomputed from ext.zeroshot.<set>.jsonl",
+  "jev_laya_en": "src/macjev/eval/external/zeroshot_topics.py PUBLISHED = elcronos results/cross_dataset_summary.json @ a1901bc3d520e73936de8d4326545c0cdcf742fb"
+ },
+ "not_plotted_on_purpose": "Jev on fin_topic (not a win, not within 5 pts); macro-F1 vs Jev",
+ "footnote": "Zero-shot for every system: neither set is in v3's training pool; accuracy over every row of the pinned test files. Jev (1.13, API) and English Laya: numbers published by the elcronos jev-vs-open-decision-models study with its own prompt (results/cross_dataset_summary.json @ a1901bc), not re-run by us. v3: scored by us on the identical rows, label sets and instruction, in v3's own input format."
+}

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": 248044,
+  "transformers_version": "5.17.0",
+  "use_cache": true
+}

jev_style_decision.py

@@ -0,0 +1,517 @@
+"""Jev-Style-0.8B-Decision-v3: typed decisions with transformers / PyTorch (CUDA, MPS, CPU).
+
+Self-contained runtime for chaoliangUNSW/Jev-Style-0.8B-Decision-v3 (Apache-2.0). No dependency on any training code:
+rendering, verdict readout and calibration are implemented below and reproduce the reference
+implementation used for evaluation (see release_config.json -> "runtime_parity").
+
+Calibration temperature: with no category (the default) probabilities use the global temperature of
+readout_config.json (temperatures.global = 0.880); pass category=... (CLI --category, JSONL "category")
+for the fitted group temperature of that category's family x question type x option-count bucket, or
+temperature=... to override (1.0 = uncalibrated scores).
+"""
+# ----------------------------------------------------------------------------------------------
+# Shared core (identical in jev_style_decision.py, jev_style_decision_gguf.py and
+# jev_style_decision_mlx.py): input rendering, verdict readout, calibrated probabilities.
+#
+# Input layout ("macjev-render-v1"; token segments are encoded separately and concatenated):
+#
+#   State:\n<state>\n\n
+#   Question [<type>]: <question>\nOptions:\n
+#   - <option 1>\n ... - <option K>\n
+#   Judge each option:\n
+#   <option 1> ->\n ... <option K> ->\n
+#
+# Score of option k = logit(" yes") - logit(" no") at the k-th " ->" token (computed from the final
+# hidden state and the tied embedding rows, float32). Probabilities = softmax(scores / T), where T is
+# the calibration temperature shipped in readout_config.json:
+#   * no category given (the default): T = temperatures.global (the file's global temperature);
+#   * category="..." given: T = the fitted group temperature of (family of that category x question
+#     type x option-count bucket), or temperatures.global when that group was not fitted;
+#   * temperature=... given: that value (1.0 = uncalibrated scores).
+# T is clamped to temperatures.clamp. Text inside the state or options is tokenised with special
+# tokens disabled, so e.g. "<|im_end|>" in user text can never act as a control token.
+#
+# Budgets: whole input <= 25,600 tokens; question + options + readout ("head") <= 2,048 tokens.
+# Larger inputs raise InputBudgetError. Nothing is ever truncated.
+# ----------------------------------------------------------------------------------------------
+import argparse
+import hashlib
+import json
+import math
+import sys
+from pathlib import Path
+
+import numpy as np
+
+MODEL_NAME = "Jev-Style-0.8B-Decision-v3"
+TEMPLATE_VERSION = "macjev-render-v1"
+READOUT_FORMAT = "macjev-readout-v1"
+CONTEXT_LIMIT = 25_600          # state + question + options + readout
+HARD_HEAD_MAX = 2048            # question + options + readout
+QTYPES = ("choice", "score", "noul")
+HERE = Path(__file__).resolve().parent
+
+# calibration families (category prefix -> family), same table the temperatures were fitted with
+FAMILY_BY_CATEGORY_PREFIX = (("typed_official", "typed"), ("typed_synthetic", "typed_synth"), ("general_", "general"),
+                             ("intent", "intent"), ("nli", "nli"), ("theme_", "theme"), ("mac_", "mac"),
+                             ("long_", "long"))
+
+
+class InputBudgetError(ValueError):
+    """The rendered input exceeds a token budget. Nothing was truncated."""
+
+
+class QuestionError(ValueError):
+    """The question/options are malformed."""
+
+
+# -- questions ------------------------------------------------------------------------------------
+def option_names(question):
+    """Canonical option identifiers, in the order the probabilities are returned."""
+    if not isinstance(question, dict):
+        raise QuestionError("question must be a dict {'t', 'ins', 'crit'}")
+    t, crit = question.get("t"), question.get("crit")
+    if not isinstance(question.get("ins"), str) or not question["ins"].strip():
+        raise QuestionError("question text ('ins') must be a non-empty string")
+    if t == "choice":
+        if not isinstance(crit, dict) or not crit:
+            raise QuestionError("choice needs a non-empty dict {option name: description or None}")
+        return [str(k) for k in crit]
+    if t == "score":
+        if not isinstance(crit, list) or not 2 <= len(crit) <= 10:
+            raise QuestionError("score needs a list of 2..10 level descriptions")
+        return [str(i) for i in range(len(crit))]
+    if t == "noul":
+        if crit is not None and not isinstance(crit, dict):
+            raise QuestionError("noul criteria must be None or {'false': ..., 'true': ...}")
+        return ["false", "true"]
+    raise QuestionError(f"unknown question type {t!r} (expected one of {QTYPES})")
+
+
+def make_question(question, options=None, qtype=None):
+    """Build a typed question.
+
+    * ``question`` already a dict {"t", "ins", "crit"}: validated and returned.
+    * ``qtype="choice"`` (default when ``options`` is given): ``options`` = {name: description or None}
+      or a list of names.
+    * ``qtype="score"``: ``options`` = list of 2..10 level descriptions (level 0 first).
+    * ``qtype="noul"`` (default when no options): a true/false statement; ``options`` may be
+      {"false": "...", "true": "..."} to describe the two outcomes.
+    """
+    if isinstance(question, dict):
+        q = dict(question)
+    else:
+        if qtype is None:
+            qtype = "choice" if options is not None else "noul"
+        if qtype == "choice":
+            if isinstance(options, (list, tuple)):
+                if len(set(map(str, options))) != len(options):
+                    raise QuestionError("duplicate option names")
+                crit = {str(o): None for o in options}
+            else:
+                crit = options
+        elif qtype == "score":
+            crit = list(options) if options is not None else None
+        else:
+            crit = options
+        q = {"t": qtype, "ins": question, "crit": crit}
+    option_names(q)
+    return q
+
+
+def serialize_state(state):
+    """Strings pass through unchanged; any other JSON value is serialised (ensure_ascii=False)."""
+    if isinstance(state, str):
+        return state
+    return json.dumps(state, ensure_ascii=False)
+
+
+def _criterion(value):
+    if isinstance(value, str):
+        return value
+    return json.dumps(value, ensure_ascii=False, separators=(", ", ": "), default=str)
+
+
+def render_options(question):
+    t, crit = question["t"], question.get("crit")
+    if t == "choice":
+        return [k if v is None or v == "" else f"{k}: {_criterion(v)}" for k, v in crit.items()]
+    if t == "score":
+        return [f"level {i}: {_criterion(c)}" for i, c in enumerate(crit)]
+    crit = crit or {}
+    false_c, true_c = crit.get("false"), crit.get("true")
+    return ["false: " + (_criterion(false_c) if false_c not in (None, "") else "no, the statement does not hold"),
+            "true: " + (_criterion(true_c) if true_c not in (None, "") else "yes, the statement holds")]
+
+
+# -- tokenizer + renderer -----------------------------------------------------------------------
+class TextEncoder:
+    """HF ``tokenizers`` tokenizer.json; no BOS/EOS, special tokens in text are split (never control tokens)."""
+
+    def __init__(self, tokenizer_json):
+        from tokenizers import Tokenizer
+        self.tk = Tokenizer.from_file(str(tokenizer_json))
+        self.tk.encode_special_tokens = True
+
+    def __call__(self, text):
+        return self.tk.encode(text, add_special_tokens=False).ids
+
+
+class Rendered:
+    __slots__ = ("ids", "prefix_len", "slots", "names", "head_tokens")
+
+    def __init__(self, ids, prefix_len, slots, names, head_tokens):
+        self.ids, self.prefix_len, self.slots, self.names, self.head_tokens = ids, prefix_len, slots, names, head_tokens
+
+
+class Renderer:
+    def __init__(self, encode, readout_cfg, max_len=CONTEXT_LIMIT, head_max=HARD_HEAD_MAX):
+        if readout_cfg.get("format") != READOUT_FORMAT or readout_cfg.get("template") != TEMPLATE_VERSION:
+            raise ValueError("readout_config.json is not a macjev-readout-v1 / macjev-render-v1 config")
+        if readout_cfg.get("readout") != "verdict":
+            raise ValueError("this runtime implements the verdict readout only")
+        if not 0 < int(max_len) <= CONTEXT_LIMIT:
+            raise ValueError(f"max_len must be in 1..{CONTEXT_LIMIT}")
+        if not 0 < int(head_max) <= HARD_HEAD_MAX:
+            raise ValueError(f"head_max must be in 1..{HARD_HEAD_MAX}")
+        self.enc, self.max_len, self.head_max = encode, int(max_len), int(head_max)
+        st = readout_cfg["slot_tokens"]
+        self.yes, self.no, arrow = int(st["yes"]["id"]), int(st["no"]["id"]), int(st["verdict_slot"]["id"])
+        for text, want in ((" yes", self.yes), (" no", self.no), (" ->", arrow)):
+            got = self.enc(text)
+            if got != [want]:
+                raise ValueError(f"tokenizer mismatch: {text!r} -> {got}, readout_config expects [{want}]")
+        self.arrow = [arrow]
+        self.newline = self.enc("\n")
+        self.dash = self.enc("- ")
+        self.judge = self.enc("Judge each option:\n")
+
+    def prefix_ids(self, state):
+        return self.enc("State:\n") + self.enc(serialize_state(state)) + self.enc("\n\n")
+
+    def render(self, state, question, head_max=None, max_len=None):
+        head_max = self.head_max if head_max is None else int(head_max)
+        max_len = self.max_len if max_len is None else min(int(max_len), self.max_len)
+        if head_max > HARD_HEAD_MAX:
+            raise InputBudgetError(f"head_max may not exceed {HARD_HEAD_MAX}")
+        names = option_names(question)
+        opts = [self.enc(o) for o in render_options(question)]
+        suffix = self.enc(f"Question [{question['t']}]: {question['ins']}\nOptions:\n")
+        for o in opts:
+            suffix += self.dash + o + self.newline
+        suffix += self.judge
+        rel = []
+        for o in opts:
+            suffix += o + self.arrow
+            rel.append(len(suffix) - 1)
+            suffix += self.newline
+        if len(suffix) > head_max:
+            raise InputBudgetError(f"question + options + readout need {len(suffix)} tokens; the head budget is "
+                                   f"{head_max} (hard cap {HARD_HEAD_MAX}). Nothing was truncated: shorten the "
+                                   f"question/options or split the options over several questions.")
+        prefix = self.prefix_ids(state)
+        ids = prefix + suffix
+        if len(ids) > max_len:
+            raise InputBudgetError(f"input needs {len(ids)} tokens (state {len(prefix)} + head {len(suffix)}); the "
+                                   f"limit is {max_len} (model maximum {CONTEXT_LIMIT}). Nothing was truncated: "
+                                   f"shorten the state.")
+        return Rendered(ids, len(prefix), [len(prefix) + s for s in rel], names, len(suffix))
+
+
+# -- calibration ----------------------------------------------------------------------------------
+def family(category):
+    for prefix, fam in FAMILY_BY_CATEGORY_PREFIX:
+        if category.startswith(prefix):
+            return fam
+    return "other"
+
+
+def option_bucket(k):
+    return "2" if k <= 2 else "3-5" if k <= 5 else "6-10" if k <= 10 else "11-20" if k <= 20 else "21+"
+
+
+def lookup_temperature(temps, category, qtype, n_options):
+    """Calibration temperature. ``category`` None/"" -> the global temperature; otherwise the fitted
+    group (family(category) x qtype x option bucket), falling back to the global temperature."""
+    g = None
+    if category:
+        g = (temps.get("groups") or {}).get(f"{family(category)}|{qtype}|{option_bucket(n_options)}")
+    t = g["T"] if g else temps.get("global", 1.0)
+    lo, hi = temps.get("clamp", [0.3, 5.0])
+    return float(min(hi, max(lo, t)))
+
+
+def concentration(p):
+    k = len(p)
+    if k < 2:
+        return 1.0
+    ent = -(p * np.log(np.clip(p, 1e-12, 1.0))).sum()
+    return float(np.clip(1.0 - ent / math.log(k), 0.0, 1.0))
+
+
+def _sha256(path):
+    h = hashlib.sha256()
+    with open(path, "rb") as f:
+        for b in iter(lambda: f.read(1 << 22), b""):
+            h.update(b)
+    return h.hexdigest()
+
+
+def verify_manifest(model_dir, only=None):
+    """Re-hash the files listed in manifest.json (all, or those whose path starts with one of ``only``).
+    Documentation (README.md, figures/, assets/) is recorded in the manifest but not checked here, so
+    a card edit never makes the runtime refuse to load."""
+    model_dir = Path(model_dir)
+    man = json.loads((model_dir / "manifest.json").read_text())
+    bad, missing, checked = [], [], 0
+    for name, rec in man["files"].items():
+        if name == "README.md" or name.startswith(("assets/", "figures/")):
+            continue
+        if only and not any(name == o or name.startswith(o.rstrip("/") + "/") for o in only):
+            continue
+        p = model_dir / name
+        if not p.exists():
+            missing.append(name)
+        elif _sha256(p) != rec["sha256"]:
+            bad.append(name)
+        checked += 1
+    return {"ok": not bad and not missing, "checked": checked, "bad": bad, "missing": missing}
+
+
+class DecisionBase:
+    """Backend-independent part. Subclasses implement ``_scores(rendered) -> list[float]`` and may
+    override ``_scores_many(list of rendered) -> list of list[float]`` (several questions, one state).
+
+    Calibration: ``category`` (constructor default or per call) selects the fitted group temperature
+    of that category's family; with no category anywhere, the global temperature of
+    readout_config.json (temperatures.global) is used."""
+    backend = "base"
+
+    def _setup(self, model_dir, tokenizer_json, category=None, head_max=HARD_HEAD_MAX, max_len=CONTEXT_LIMIT):
+        self.model_dir = Path(model_dir)
+        self.readout_config = json.loads((self.model_dir / "readout_config.json").read_text())
+        self.temperatures = self.readout_config["temperatures"]
+        self.default_category = category or None       # None -> temperatures.global
+        self.encode = TextEncoder(tokenizer_json)
+        self.renderer = Renderer(self.encode, self.readout_config, max_len=max_len, head_max=head_max)
+
+    def temperature(self, question, category=None):
+        """T for ``question``: group temperature of ``category`` (or the constructor's default category);
+        the global temperature when neither is given."""
+        return lookup_temperature(self.temperatures, category or self.default_category, question["t"],
+                                  len(option_names(question)))
+
+    def _scores_many(self, rendered):
+        return [self._scores(r) for r in rendered]
+
+    def _result(self, r, q, scores, category=None, temperature=None):
+        scores = [float(x) for x in scores]
+        t = float(temperature) if temperature is not None else self.temperature(q, category)
+        z = np.asarray(scores, float) / t
+        if not np.all(np.isfinite(z)):
+            raise FloatingPointError("non-finite decision scores")
+        p = np.exp(z - z.max())
+        p /= p.sum()
+        i = int(p.argmax())
+        return {"answer": r.names[i], "probabilities": dict(zip(r.names, p.tolist())),
+                "scores": dict(zip(r.names, scores)), "temperature": t, "top_probability": float(p[i]),
+                "entropy_concentration": concentration(p), "input_tokens": len(r.ids), "head_tokens": r.head_tokens,
+                "model": MODEL_NAME, "backend": self.backend}
+
+    def decide(self, state, question, options=None, qtype=None, category=None, temperature=None, head_max=None):
+        """Score one question about ``state``.
+
+        Returns {"answer", "probabilities" {option: p}, "scores" {option: logit(yes)-logit(no)},
+        "temperature", "top_probability", "entropy_concentration", "input_tokens", "head_tokens"}.
+        Temperature: with no ``category`` (here or in the constructor) the global temperature of
+        readout_config.json is used; ``category`` picks the fitted group temperature of its family
+        (e.g. "mac_gate", "general_topic", "theme_routing", "intent", "typed_official");
+        ``temperature`` overrides both (1.0 = uncalibrated scores).
+        Raises InputBudgetError (never truncates) or QuestionError.
+        """
+        q = make_question(question, options, qtype)
+        r = self.renderer.render(state, q, head_max=head_max)
+        return self._result(r, q, self._scores(r), category, temperature)
+
+    def decide_many(self, state, questions, category=None, temperature=None, head_max=None):
+        """Several questions about the same state (each a dict {"t","ins","crit"}); results in order.
+        Same outputs as calling decide() per question. The llama.cpp runtime sends all questions in one
+        request and shares the state in whole 1,024-token ubatches (see JevStyleDecisionGGUF, also for
+        its faster, not bit-identical many_mode="batched"). All questions are rendered and
+        budget-checked before any scoring."""
+        qs = [make_question(q) for q in questions]
+        rs = [self.renderer.render(state, q, head_max=head_max) for q in qs]
+        if not rs:
+            return []
+        return [self._result(r, q, sc, category, temperature) for r, q, sc in zip(rs, qs, self._scores_many(rs))]
+
+
+def base_arg_parser(description):
+    ap = argparse.ArgumentParser(description=description)
+    ap.add_argument("--model-dir", default=str(HERE), help="folder with the weights and readout_config.json")
+    ap.add_argument("--state", help="state as plain text")
+    ap.add_argument("--state-json", help="state as a JSON value")
+    ap.add_argument("--question", help="question text (or a JSON question {'t','ins','crit'})")
+    ap.add_argument("--options", help="JSON: {name: description} or [names] (choice); [levels] (score)")
+    ap.add_argument("--qtype", choices=QTYPES)
+    ap.add_argument("--category", help="calibration family key, e.g. mac_gate, general_topic, theme_routing, intent "
+                                       "(default: none -> the global temperature of readout_config.json)")
+    ap.add_argument("--temperature", type=float, help="override the calibrated temperature")
+    ap.add_argument("--head-max", type=int, default=HARD_HEAD_MAX)
+    ap.add_argument("--max-len", type=int, default=CONTEXT_LIMIT)
+    ap.add_argument("--jsonl", help="batch mode: input JSON lines {id?, state, question, options?, qtype?, "
+                                    "category?}; one JSON result per line on stdout")
+    ap.add_argument("--verify", action="store_true", help="check sha256 of the files in manifest.json first")
+    return ap
+
+
+def run_cli(args, engine):
+    def one(rec):
+        q = rec["question"]
+        return engine.decide(rec.get("state", ""), q, options=rec.get("options"), qtype=rec.get("qtype"),
+                             category=rec.get("category"), temperature=rec.get("temperature", args.temperature))
+    if args.jsonl:
+        src = sys.stdin if args.jsonl == "-" else open(args.jsonl, encoding="utf-8")
+        for n, line in enumerate(src):
+            if not line.strip():
+                continue
+            rec = json.loads(line)
+            rid = rec.get("id", n)
+            try:
+                out = {"id": rid, **one(rec)}
+            except (InputBudgetError, QuestionError) as e:
+                out = {"id": rid, "error": f"{type(e).__name__}: {e}"}
+            print(json.dumps(out, ensure_ascii=False), flush=True)
+        return 0
+    if args.question is None:
+        raise SystemExit("--question (or --jsonl) is required")
+    state = json.loads(args.state_json) if args.state_json is not None else (args.state or "")
+    question = args.question
+    if question.lstrip().startswith("{"):
+        question = json.loads(question)
+    rec = {"state": state, "question": question, "options": json.loads(args.options) if args.options else None,
+           "qtype": args.qtype, "category": args.category}
+    print(json.dumps(one(rec), ensure_ascii=False, indent=2))
+    return 0
+# ---------------------------------------------------------------------------- end of shared core
+
+
+# ------------------------------------------------------------------------------ PyTorch backend
+ATTN_CHUNK = 1024
+_CHUNKED_NAME = "jev_chunked_sdpa"
+_CHUNKED_REGISTERED = False
+
+
+def _chunked_sdpa_forward(module, query, key, value, attention_mask, dropout=0.0, scaling=None, is_causal=None,
+                          **kwargs):
+    """Query-chunked SDPA (MPS / CPU): the same computation as transformers' sdpa, but the
+    [heads x queries x keys] score matrix is built ATTN_CHUNK queries at a time, so a 25,600-token
+    input does not need ~21 GB for one attention call. Inputs of <= ATTN_CHUNK tokens take the
+    unchanged sdpa path."""
+    import torch
+    from transformers.integrations.sdpa_attention import sdpa_attention_forward
+    q_len, kv_len = query.shape[2], key.shape[2]
+    chunk = int(getattr(module, "jev_attn_chunk", ATTN_CHUNK) or ATTN_CHUNK)
+    if q_len <= chunk:
+        return sdpa_attention_forward(module, query, key, value, attention_mask, dropout=dropout, scaling=scaling,
+                                      is_causal=is_causal, **kwargs)
+    causal = is_causal if is_causal is not None else getattr(module, "is_causal", True)
+    past = kv_len - q_len
+    outs = []
+    for s in range(0, q_len, chunk):
+        e = min(q_len, s + chunk)
+        if attention_mask is not None:
+            m = attention_mask if attention_mask.shape[-2] == 1 else attention_mask[:, :, s:e, :]
+        elif causal:
+            qpos = torch.arange(s + past, e + past, device=query.device)
+            kpos = torch.arange(kv_len, device=query.device)
+            m = (kpos[None, :] <= qpos[:, None])[None, None]
+        else:
+            m = None
+        out, _ = sdpa_attention_forward(module, query[:, :, s:e], key, value, m, dropout=dropout, scaling=scaling,
+                                        is_causal=False, **kwargs)
+        outs.append(out)
+    return torch.cat(outs, dim=1), None
+
+
+def _enable_chunked_attention(model, chunk=ATTN_CHUNK):
+    global _CHUNKED_REGISTERED
+    if getattr(model.config, "_attn_implementation", None) not in ("sdpa", _CHUNKED_NAME):
+        return False
+    try:
+        from transformers import AttentionInterface
+        from transformers.masking_utils import ALL_MASK_ATTENTION_FUNCTIONS, AttentionMaskInterface
+    except ImportError:
+        return False
+    if not _CHUNKED_REGISTERED:
+        AttentionInterface.register(_CHUNKED_NAME, _chunked_sdpa_forward)
+        AttentionMaskInterface.register(_CHUNKED_NAME, ALL_MASK_ATTENTION_FUNCTIONS["sdpa"])
+        _CHUNKED_REGISTERED = True
+    model.set_attn_implementation(_CHUNKED_NAME)
+    for m in model.modules():
+        if hasattr(m, "is_causal"):
+            m.jev_attn_chunk = int(chunk)
+    return True
+
+
+class JevStyleDecision(DecisionBase):
+    """Transformers / PyTorch runtime (CUDA, Apple MPS or CPU).
+
+    >>> m = JevStyleDecision(".")                       # float32 on the best available device
+    >>> m.decide({"messages": ["Refund still missing after 3 weeks"]},
+    ...          "Which team should handle this ticket?",
+    ...          options={"billing": "payments, refunds", "tech": "bugs, crashes", "sales": "pricing, plans"},
+    ...          category="theme_routing")["probabilities"]
+    """
+    backend = "torch"
+
+    def __init__(self, model_dir=HERE, device=None, dtype="float32", category=None, head_max=HARD_HEAD_MAX,
+                 max_len=CONTEXT_LIMIT, attn_chunk=ATTN_CHUNK, verify=False):
+        import torch
+        self.torch = torch
+        model_dir = Path(model_dir)
+        if verify:
+            res = verify_manifest(model_dir)
+            if not res["ok"]:
+                raise RuntimeError(f"integrity check failed: {res}")
+        self._setup(model_dir, model_dir / "tokenizer.json", category, head_max, max_len)
+        if device is None:
+            device = ("cuda" if torch.cuda.is_available() else
+                      "mps" if getattr(torch.backends, "mps", None) and torch.backends.mps.is_available() else "cpu")
+        dt = getattr(torch, dtype) if isinstance(dtype, str) else dtype
+        try:
+            from transformers.models.qwen3_5.modeling_qwen3_5 import Qwen3_5ForCausalLM as cls
+        except ImportError as e:
+            raise ImportError("this model needs a transformers version with Qwen3.5 support "
+                              "(transformers.models.qwen3_5)") from e
+        try:
+            model = cls.from_pretrained(str(model_dir), dtype=dt)
+        except TypeError:                                       # transformers 4.x keyword
+            model = cls.from_pretrained(str(model_dir), torch_dtype=dt)
+        self.model = model.to(device).eval()
+        self.device, self.dtype = device, dt
+        self.chunked_attention = _enable_chunked_attention(self.model, attn_chunk) if device != "cuda" else False
+        w = self.model.get_output_embeddings().weight
+        self.direction = (w[self.renderer.yes].float() - w[self.renderer.no].float()).detach()
+
+    def _scores(self, r):
+        torch = self.torch
+        with torch.no_grad():
+            ids = torch.tensor([r.ids], device=self.device)
+            h = self.model.model(input_ids=ids, use_cache=False).last_hidden_state      # final normed hidden states
+            hs = h[0, torch.tensor(r.slots, device=self.device)].float()
+            return (hs @ self.direction).cpu().tolist()
+
+
+def main(argv=None):
+    ap = base_arg_parser(f"{MODEL_NAME}: typed decisions with transformers / PyTorch")
+    ap.add_argument("--device", choices=["cuda", "mps", "cpu"])
+    ap.add_argument("--dtype", default="float32", choices=["float32", "bfloat16", "float16"])
+    args = ap.parse_args(argv)
+    engine = JevStyleDecision(args.model_dir, device=args.device, dtype=args.dtype, category=args.category,
+                              head_max=args.head_max, max_len=args.max_len, verify=args.verify)
+    return run_cli(args, engine)
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

manifest.json

@@ -0,0 +1,181 @@
+{
+  "format": "jev-style-manifest-v1",
+  "repo": "chaoliangUNSW/Jev-Style-0.8B-Decision-v3",
+  "created_unix": 1790261965.8405728,
+  "files": {
+    "LICENSE": {
+      "sha256": "bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a",
+      "bytes": 11544
+    },
+    "NOTICE": {
+      "sha256": "3c1bee42827901d754cf64e437cf8837ea94241348f5f4be9e6c4e14fbfc5716",
+      "bytes": 1966
+    },
+    "README.md": {
+      "sha256": "ccf12d32e307f97aa6818c098fc874aec0f3666c79af3834fee799f98eb6871b",
+      "bytes": 31905
+    },
+    "chat_template.jinja": {
+      "sha256": "273d8e0e683b885071fb17e08d71e5f2a5ddfb5309756181681de4f5a1822d80",
+      "bytes": 7755
+    },
+    "config.json": {
+      "sha256": "3f56b6210db3c52e0bafb50da005eb52b6d340ab4ad980091879b397743675fc",
+      "bytes": 1790
+    },
+    "figures/beyond_laya.json": {
+      "sha256": "a6dfd582fe3aca32bcd67cf01b85190d92084a0790f7744dd03180df5923f946",
+      "bytes": 4357
+    },
+    "figures/beyond_laya.png": {
+      "sha256": "0dbd24837759acf05ed3a668016ecaea75a41a093b64588417d2fd750b95779b",
+      "bytes": 185526
+    },
+    "figures/beyond_laya.svg": {
+      "sha256": "4b64a4e71d3c9c05c9a785003bbfbd7269c5341bcc8abd560a2b156cfb771d6b",
+      "bytes": 22000
+    },
+    "figures/calibration.data.json": {
+      "sha256": "0396f5d48ba839c65718a7f1b449e7fa074503cbd1b9db48a2d40096c2d6142a",
+      "bytes": 3765
+    },
+    "figures/calibration.png": {
+      "sha256": "3176339f6e4f58c9efa583b101f5862141a326468cf177f2c50c3ce2d1adf06f",
+      "bytes": 148880
+    },
+    "figures/calibration.svg": {
+      "sha256": "7cdb177072406c736a03bcec696adacf037cf810f156dd8f94ab6cf24e1d76ec",
+      "bytes": 15047
+    },
+    "figures/design_table.data.json": {
+      "sha256": "84fd672a2c69d6881c199dc09ccbe7b1313b73bf7f56f54f574cf4eefe052003",
+      "bytes": 5313
+    },
+    "figures/design_table.png": {
+      "sha256": "5b907319a5a4f6ee8eea5800fbc732b95d15e941b1d0cd8d774d3705a1341744",
+      "bytes": 331936
+    },
+    "figures/design_table.svg": {
+      "sha256": "a0a23216766b235c950a071a614691b010dcea9f3a0c5bb8c6f8cc2a7ecae0bc",
+      "bytes": 26655
+    },
+    "figures/headline_typed.data.json": {
+      "sha256": "04380aa3823f80489ae37f893de0bda67ae2dae38fd07e37f88700f7c173f0ec",
+      "bytes": 3819
+    },
+    "figures/headline_typed.png": {
+      "sha256": "9e763e4cbd86a25299785b94a89945f2c3a7cab4b1c7d5000f9456c4d220a1f4",
+      "bytes": 175884
+    },
+    "figures/headline_typed.svg": {
+      "sha256": "a49ec30b28319d682d9465145bf6435ac4ec4ad11c46a28c552c9b851fb27f37",
+      "bytes": 20359
+    },
+    "figures/jevbench.data.json": {
+      "sha256": "f9173e65f5cd1fe9fadad8c93a8d00dbe5181e7314d49f18c6f547a0ed168543",
+      "bytes": 2645
+    },
+    "figures/jevbench.png": {
+      "sha256": "b75ab6864b4487d0a323c1019f7ce405fedcdb5a4d8dc910c86a29b3141b5388",
+      "bytes": 142720
+    },
+    "figures/jevbench.svg": {
+      "sha256": "2014d7664f268c2621e68226d4ec92eec6dd8dc7a04029024b5110a9341b812f",
+      "bytes": 12181
+    },
+    "figures/latency.data.json": {
+      "sha256": "c0eb510a69dfe1b1cd85beca6ba3c354be4984d588b4aae2672331606a119cbc",
+      "bytes": 1402
+    },
+    "figures/latency.png": {
+      "sha256": "d6d0d34b875997b8a733d4776dc34a1b564ddc64b48cea9d0def7f72a34af79e",
+      "bytes": 203081
+    },
+    "figures/latency.svg": {
+      "sha256": "c7bdf011eec488aae7197d653cb8486dfd8ccbd1a58f5ba222873de7410aca30",
+      "bytes": 24079
+    },
+    "figures/long_context.json": {
+      "sha256": "e6289e61da3bf4117ed4ed05820b99620e541c25cdeb5262e3a580fa9febff62",
+      "bytes": 5553
+    },
+    "figures/long_context.png": {
+      "sha256": "957fb443f67d1ed91ef61fa83b0496347299823799bc97692d25893060b1e26d",
+      "bytes": 224537
+    },
+    "figures/long_context.svg": {
+      "sha256": "ad7be559f77b11b56df84b4f286f9cea9c076d4caf47a2f8fb0f67ab46f7968b",
+      "bytes": 21403
+    },
+    "figures/multilingual.data.json": {
+      "sha256": "21018e6ed6ec35f30b2739a5fd0adfae10eee6aa6d24e0e877651317d7cc76e7",
+      "bytes": 9921
+    },
+    "figures/multilingual.png": {
+      "sha256": "544a3982019f7d0d8f788f4931ba88bde317956c87f7a7138fcd5e4ec22013d4",
+      "bytes": 300658
+    },
+    "figures/multilingual.svg": {
+      "sha256": "c4442d53ba20f1a3bfa3b5764db211a7a18db95c48ec834f78d454ca13efd7e4",
+      "bytes": 63997
+    },
+    "figures/quantization.data.json": {
+      "sha256": "a140816ef145a6dd096ef471e4afbff1cecd33adb663c6cb856e19845349fc2d",
+      "bytes": 5473
+    },
+    "figures/quantization.png": {
+      "sha256": "bc75c212caa11008ca929a685d3085040c2b4698afb58d04a0ad692b815be230",
+      "bytes": 253992
+    },
+    "figures/quantization.svg": {
+      "sha256": "70720bb46ec1e718dcbfac9a83d3b5d11d50b3a5cc3356915465d70d95f3d43b",
+      "bytes": 28756
+    },
+    "figures/zeroshot.json": {
+      "sha256": "003c087074b876efab789fadf94f51d134c604d5f2efce417557732de2272e9d",
+      "bytes": 1867
+    },
+    "figures/zeroshot.png": {
+      "sha256": "255e196d449c086130d0f06c493c64a7cc21730a38d21c1c2188641e4c08c002",
+      "bytes": 128831
+    },
+    "figures/zeroshot.svg": {
+      "sha256": "a148fa7cb8f4e0b6e529996441fadfa883595bfeb05dc93c4048609071655790",
+      "bytes": 14474
+    },
+    "generation_config.json": {
+      "sha256": "62153eb6c69f2e1f426beaa8002b7186437e949c7588167085df14e10e9c0a73",
+      "bytes": 116
+    },
+    "jev_style_decision.py": {
+      "sha256": "be38df723d83ef5fdffb8f7dfaee63b3c31e632b233fc3c7d5f6eb8c83fd994f",
+      "bytes": 25917
+    },
+    "model.safetensors": {
+      "sha256": "0f8c861605dcdfb356a63e056baa2106e81042e8981e8d0d26fe50b34599541e",
+      "bytes": 1504827608
+    },
+    "readout_config.json": {
+      "sha256": "01de9bcce7effbfd1ae0a3fa13e52d7fa9aa4dd1cd1e47977d6bbdcd5e63054a",
+      "bytes": 6236
+    },
+    "release_config.json": {
+      "sha256": "7ee881089c1ca5b019f98a26e60da4b0a454100d8cd557e5e352df3813aa31c0",
+      "bytes": 7088
+    },
+    "requirements.txt": {
+      "sha256": "855306c7cd8db9fcea2f865d61a0c53330c530c9dcf07e955a9250aaae2993f1",
+      "bytes": 201
+    },
+    "tokenizer.json": {
+      "sha256": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523",
+      "bytes": 19989325
+    },
+    "tokenizer_config.json": {
+      "sha256": "66e427c470fe580fe8c7b5725d857af23d8417e37fae62667ec698306a19987b",
+      "bytes": 1124
+    }
+  },
+  "readme_hashed": true,
+  "note": "manifest.json hashes every file of the repo except itself, README.md and figures/ included. The runtime --verify check skips the documentation (README.md, figures/, assets/) and checks every other file."
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f8c861605dcdfb356a63e056baa2106e81042e8981e8d0d26fe50b34599541e
+size 1504827608

readout_config.json

@@ -0,0 +1,247 @@
+{
+  "format": "macjev-readout-v1",
+  "model_name": "Jev-Style-0.8B-Decision-v3",
+  "readout": "verdict",
+  "template": "macjev-render-v1",
+  "slot_tokens": {
+    "yes": {
+      "text": " yes",
+      "id": 9542
+    },
+    "no": {
+      "text": " no",
+      "id": 874
+    },
+    "verdict_slot": {
+      "text": " ->",
+      "id": 1411
+    },
+    "letters": {
+      "A": 357,
+      "B": 417,
+      "C": 351,
+      "D": 414,
+      "E": 458,
+      "F": 426,
+      "G": 469,
+      "H": 462,
+      "I": 353,
+      "J": 604,
+      "K": 710,
+      "L": 436,
+      "M": 380,
+      "N": 443,
+      "O": 496,
+      "P": 387,
+      "Q": 1167,
+      "R": 423,
+      "S": 326,
+      "T": 345,
+      "U": 533,
+      "V": 629,
+      "W": 457,
+      "X": 1543,
+      "Y": 783,
+      "Z": 1799,
+      "a": 264,
+      "b": 292,
+      "c": 272,
+      "d": 293,
+      "e": 378,
+      "f": 281,
+      "g": 338,
+      "h": 304,
+      "i": 585,
+      "j": 492,
+      "k": 580,
+      "l": 324,
+      "m": 295,
+      "n": 307,
+      "o": 296,
+      "p": 280,
+      "q": 2715,
+      "r": 427,
+      "s": 274,
+      "t": 259,
+      "u": 560,
+      "v": 343,
+      "w": 288,
+      "x": 830,
+      "y": 374,
+      "z": 1110
+    }
+  },
+  "score": "per option k: logit[' yes'] - logit[' no'] at the k-th ' ->' slot, computed as h_slot . (w_yes - w_no) from the final normed hidden state and the tied embedding rows (float32)",
+  "probabilities": "softmax(scores / T); T = temperatures.groups['<family>|<qtype>|<option bucket>'].T (else temperatures.global), clamped to temperatures.clamp",
+  "families": {
+    "typed_official*": "typed",
+    "typed_synthetic*": "typed_synth",
+    "general_*": "general",
+    "intent*": "intent",
+    "nli*": "nli",
+    "theme_*": "theme",
+    "mac_*": "mac",
+    "long_*": "long",
+    "anything else": "other (global T)"
+  },
+  "option_buckets": [
+    "2",
+    "3-5",
+    "6-10",
+    "11-20",
+    "21+"
+  ],
+  "default_category": null,
+  "default_category_note": "no category given -> temperatures.global (the global temperature); pass category=... for the fitted group temperature of that category's family",
+  "tokenizer_sha256": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523",
+  "budgets": {
+    "max_len": 25600,
+    "head_max": 2048,
+    "hard_head_max": 2048,
+    "note": "whole input <= max_len tokens, question+options+readout <= head_max; larger inputs raise InputBudgetError, nothing is truncated"
+  },
+  "temperatures": {
+    "version": "macjev-temperatures-v1",
+    "global": 0.8800546821789332,
+    "groups": {
+      "general|choice|3-5": {
+        "T": 0.8563796906101172,
+        "T_raw": 0.8524962478467872,
+        "n": 600,
+        "weight": 0.8571428571428571
+      },
+      "general|choice|6-10": {
+        "T": 0.7883347591845132,
+        "T_raw": 0.7797058230459033,
+        "n": 1000,
+        "weight": 0.9090909090909091
+      },
+      "general|noul|2": {
+        "T": 0.9702474579038656,
+        "T_raw": 1.0023209290239254,
+        "n": 300,
+        "weight": 0.75
+      },
+      "general|score|3-5": {
+        "T": 0.9583320444900012,
+        "T_raw": 0.9748039508642105,
+        "n": 500,
+        "weight": 0.8333333333333334
+      },
+      "intent|choice|11-20": {
+        "T": 0.8536586990515279,
+        "T_raw": 0.8530447902066461,
+        "n": 4233,
+        "weight": 0.9769213016385876
+      },
+      "intent|choice|21+": {
+        "T": 0.7508751181814466,
+        "T_raw": 0.737974406022733,
+        "n": 916,
+        "weight": 0.9015748031496063
+      },
+      "long|choice|2": {
+        "T": 1.0101601686032944,
+        "T_raw": 2.5327604766936602,
+        "n": 15,
+        "weight": 0.13043478260869565
+      },
+      "long|choice|3-5": {
+        "T": 0.9789458925825579,
+        "T_raw": 0.9984431087771811,
+        "n": 540,
+        "weight": 0.84375
+      },
+      "long|choice|6-10": {
+        "T": 0.6918332634850016,
+        "T_raw": 0.6260903832047418,
+        "n": 241,
+        "weight": 0.7067448680351907
+      },
+      "long|noul|2": {
+        "T": 0.8183260460233314,
+        "T_raw": 0.7857458244621078,
+        "n": 179,
+        "weight": 0.6415770609318996
+      },
+      "long|score|3-5": {
+        "T": 0.8412851094005789,
+        "T_raw": 0.7952160414763221,
+        "n": 80,
+        "weight": 0.4444444444444444
+      },
+      "mac|choice|3-5": {
+        "T": 0.7640449061346866,
+        "T_raw": 0.753267027600698,
+        "n": 995,
+        "weight": 0.908675799086758
+      },
+      "mac|noul|2": {
+        "T": 0.922463080251143,
+        "T_raw": 0.9300179680535603,
+        "n": 577,
+        "weight": 0.8522895125553914
+      },
+      "mac|score|3-5": {
+        "T": 2.729450780811877,
+        "T_raw": 4.999707266277221,
+        "n": 187,
+        "weight": 0.6515679442508711
+      },
+      "nli|choice|3-5": {
+        "T": 1.003611674961589,
+        "T_raw": 1.0108425661039413,
+        "n": 1830,
+        "weight": 0.9481865284974094
+      },
+      "theme|choice|6-10": {
+        "T": 0.9848729122522978,
+        "T_raw": 0.9981552921230235,
+        "n": 840,
+        "weight": 0.8936170212765957
+      },
+      "theme|noul|2": {
+        "T": 0.896799495386032,
+        "T_raw": 0.89763584528285,
+        "n": 2022,
+        "weight": 0.9528746465598492
+      },
+      "typed|choice|3-5": {
+        "T": 0.9751541851571508,
+        "T_raw": 1.0336976619219436,
+        "n": 176,
+        "weight": 0.6376811594202898
+      },
+      "typed|noul|2": {
+        "T": 0.9892589014450378,
+        "T_raw": 1.054189849787923,
+        "n": 184,
+        "weight": 0.647887323943662
+      },
+      "typed|score|3-5": {
+        "T": 1.0056628114581752,
+        "T_raw": 1.0631515969968885,
+        "n": 240,
+        "weight": 0.7058823529411765
+      }
+    },
+    "clamp": [
+      0.3,
+      5.0
+    ],
+    "shrinkage_k": 100.0,
+    "key": "family|qtype|option_bucket",
+    "n_rows": 15655,
+    "fitted_on": [
+      "pool_cal.jsonl:dfb7e9beff96c6c5"
+    ],
+    "fit_quality": {
+      "nll_before": 0.37752055301970966,
+      "nll_after": 0.36671188108641545,
+      "ece_before": 0.03289307373502533,
+      "ece_after": 0.011376234101433989,
+      "n": 15655
+    }
+  },
+  "temperatures_sha256": "39ad8f6633934ff67770725993d7354ebebd6e7a08f73e50cab04df24715f2ce"
+}

release_config.json

@@ -0,0 +1,186 @@
+{
+  "format": "jev-style-release-v1",
+  "model_name": "Jev-Style-0.8B-Decision-v3",
+  "repo": "chaoliangUNSW/Jev-Style-0.8B-Decision-v3",
+  "generation": "v3 (third generation of the Jev-Style decision series)",
+  "lineage": {
+    "v1": "chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision",
+    "v1_public_gguf": "chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF",
+    "v2": "chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2",
+    "v3": "chaoliangUNSW/Jev-Style-0.8B-Decision-v3"
+  },
+  "base_model": "Qwen/Qwen3.5-0.8B",
+  "base_model_revision": "2fc06364715b967f1860aea9cf38778875588b17",
+  "base_model_relation": "finetune",
+  "architecture": "Qwen3_5ForCausalLM (text only, 24 layers: 18 Gated DeltaNet + 6 full attention, hidden 1024, tied embeddings, 752,393,024 parameters)",
+  "readout": "verdict",
+  "template": "macjev-render-v1",
+  "readout_config": "readout_config.json",
+  "budgets": {
+    "max_len": 25600,
+    "head_max": 2048
+  },
+  "source": {
+    "checkpoint_sha256": {
+      "best-0.safetensors": "b10d249adfa3e1475f0f633ab961130c06f16c898db23c846b035d24bdc5c945"
+    },
+    "text_only_model_safetensors_sha256": "0f8c861605dcdfb356a63e056baa2106e81042e8981e8d0d26fe50b34599541e",
+    "release_manifest_sha256": "4bc9f89795ccf2d749008de7739bc5294f3ec543404ca2db9cac6a3914cfb84a",
+    "llama_cpp_commit": "441df11f65ea0b6d0c72965aaf70c8241070ddcb",
+    "mlx": "0.32.2",
+    "mlx_lm": "0.31.3"
+  },
+  "calibration": {
+    "version": "macjev-temperatures-v1",
+    "global_T": 0.8800546821789332,
+    "groups": 20,
+    "n_rows": 15655,
+    "fit_quality": {
+      "nll_before": 0.37752055301970966,
+      "nll_after": 0.36671188108641545,
+      "ece_before": 0.03289307373502533,
+      "ece_after": 0.011376234101433989,
+      "n": 15655
+    },
+    "fitted_on": "calibration pool (dev/cal rows, never test rows)"
+  },
+  "g5_parity": {
+    "reference": "PyTorch float32 (same weights, same rendered inputs)",
+    "rows": 240,
+    "rows_note": "non-sealed training-distribution rows, 22 categories, en 215 / zh 25",
+    "report_sha256": "056dc3a3d1de5616f97dab3f8b5ec1f7e2c6c85258aa03f2a57b9fd65a71022b",
+    "g5_pass": true,
+    "verdict_readout": {
+      "gguf-f16": {
+        "top1_agreement": 1.0,
+        "dnll": 2.436279701012456e-05,
+        "n": 240,
+        "nan_rows": 0,
+        "gate": "top1>=0.99, |dNLL|<=0.02, no NaN, every reference row scored",
+        "pass": true
+      },
+      "gguf-q8_0": {
+        "top1_agreement": 1.0,
+        "dnll": -3.093173930673876e-05,
+        "n": 240,
+        "nan_rows": 0,
+        "gate": "top1>=0.98, |dNLL|<=0.02, no NaN, every reference row scored",
+        "pass": true
+      },
+      "gguf-q4_k_m": {
+        "top1_agreement": 1.0,
+        "dnll": 0.006187511546346225,
+        "n": 240,
+        "nan_rows": 0,
+        "gate": "report_only",
+        "pass": true
+      },
+      "mlx-bf16": {
+        "top1_agreement": 1.0,
+        "dnll": 0.00031018251516545803,
+        "n": 240,
+        "nan_rows": 0,
+        "gate": "top1>=0.99, |dNLL|<=0.02, no NaN, every reference row scored",
+        "pass": true
+      },
+      "mlx-bf16-f32act": {
+        "top1_agreement": 1.0,
+        "dnll": 0.00015659911501769708,
+        "n": 240,
+        "nan_rows": 0,
+        "gate": "top1>=0.99, |dNLL|<=0.02, no NaN, every reference row scored",
+        "pass": true
+      },
+      "mlx-8bit": {
+        "top1_agreement": 1.0,
+        "dnll": 0.00023178691602621093,
+        "n": 240,
+        "nan_rows": 0,
+        "gate": "top1>=0.98, |dNLL|<=0.02, no NaN, every reference row scored",
+        "pass": true
+      },
+      "mlx-4bit": {
+        "top1_agreement": 0.9875,
+        "dnll": 0.010326217052955111,
+        "n": 240,
+        "nan_rows": 0,
+        "gate": "report_only",
+        "pass": true
+      }
+    },
+    "gates": "top-1 >= 0.99 (16-bit) / >= 0.98 (8-bit), |dNLL| <= 0.02, no NaN; 4-bit report-only"
+  },
+  "decision_index": "not run for this release (optional follow-up)",
+  "related_repos": {
+    "main": "chaoliangUNSW/Jev-Style-0.8B-Decision-v3",
+    "gguf": "chaoliangUNSW/Jev-Style-0.8B-Decision-v3-GGUF",
+    "mlx-bf16": "chaoliangUNSW/Jev-Style-0.8B-Decision-v3-MLX-bf16",
+    "mlx-8bit": "chaoliangUNSW/Jev-Style-0.8B-Decision-v3-MLX-8bit"
+  },
+  "not_released": {
+    "mlx-4bit": "affine4-g64 reported only (G5 top-1 0.9875, report-only gate); not released"
+  },
+  "tested_with": {
+    "python": "3.12",
+    "torch": "2.14.0",
+    "transformers": "5.17.0",
+    "tokenizers": "0.23.2",
+    "numpy": "2.5.3",
+    "mlx": "0.32.2",
+    "mlx-lm": "0.31.3",
+    "llama.cpp": "441df11f65ea0b6d0c72965aaf70c8241070ddcb"
+  },
+  "weights": {
+    "model.safetensors": "bfloat16 (text-only export of the trained checkpoint)"
+  },
+  "runtime": {
+    "script": "jev_style_decision.py",
+    "class": "JevStyleDecision",
+    "default_dtype": "float32",
+    "devices": [
+      "cuda",
+      "mps",
+      "cpu"
+    ],
+    "long_inputs": "query-chunked SDPA on MPS/CPU (1024 queries per chunk)"
+  },
+  "runtime_parity": {
+    "protocol": "2026-09-24: each runtime script run in a clean subprocess (cwd = repo folder, empty PYTHONPATH, HF_HUB_OFFLINE=1, --verify) on 24 fixed parity-fixture rows (every 10th of the 240 G5 rows; 12 categories families, choice/score/noul, 92-8156 tokens) and compared with the reference (training-code) scorer of the same format on the same inputs, probabilities with the same calibration temperature; plus 2 synthetic long states (16,381 and 25,582 tokens). Rendered token ids identical to the reference renderer on 244/244 rows (240 fixture rows + 4 adversarial special-token/unicode rows). Re-run 2026-09-24 after the rename to Jev-Style-0.8B-Decision-v3, the no-category = global-temperature default and the GGUF general.name metadata edit: all six formats again identical to the reference scorer run on the original (pre-edit) files (max |prob diff| 0.0 same backend, 0 top-1 changes). Re-run 2026-09-25 after the GGUF decide_many fix (runtime code change in decide_many only, docstrings in all scripts): all six formats gave outputs identical (0.0) to the 2026-09-24 run on the 24 rows.",
+    "results": {
+      "torch-fp32-cpu": {
+        "rows": 24,
+        "errors": 0,
+        "vs_reference_same_backend": {
+          "max_abs_prob_diff": 0.0,
+          "max_abs_score_diff": 0.0,
+          "top1_changes": 0,
+          "top1_agreement": 1.0
+        },
+        "vs_reference_torch_fp32": {
+          "max_abs_prob_diff": 0.0,
+          "max_abs_score_diff": 0.0,
+          "top1_changes": 0,
+          "top1_agreement": 1.0
+        }
+      },
+      "torch-fp32-mps-long": {
+        "long16384:torch-mps-fp32": {
+          "tokens": 16381,
+          "answer_matches_gguf_mlx": true,
+          "max_abs_prob_diff_vs_gguf_f16": 0.0004853327842702093
+        },
+        "long25600:torch-mps-fp32": {
+          "tokens": 25582,
+          "answer_matches_gguf_mlx": true,
+          "max_abs_prob_diff_vs_gguf_f16": 0.0002457350394169111
+        }
+      },
+      "render_token_identity": {
+        "rows": 244,
+        "identical": 244,
+        "different": 0,
+        "bad": []
+      }
+    }
+  }
+}

requirements.txt

@@ -0,0 +1,5 @@
+# tested with: python 3.12, torch 2.14.0, transformers 5.17.0, tokenizers 0.23.2, numpy 2.5.3
+torch>=2.4
+transformers>=5.0  # needs Qwen3.5 support (transformers.models.qwen3_5)
+tokenizers>=0.21
+numpy

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+size 19989325

tokenizer_config.json

@@ -0,0 +1,32 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}