cloudyu commited on
Commit
3b960eb
·
verified ·
1 Parent(s): c3ee780

Model card: repository renamed to autotrust/JEV-9B; all ids, paths and commands updated

Browse files
Files changed (1) hide show
  1. README.md +33 -33
README.md CHANGED
@@ -29,7 +29,7 @@ metrics:
29
  - brier
30
  - ece
31
  model-index:
32
- - name: autotrust/JEV (student of TypeSafe Jev 1.13)
33
  results:
34
  - task:
35
  type: text-classification
@@ -62,17 +62,17 @@ model-index:
62
  value: 0.954
63
  ---
64
 
65
- # autotrust/JEV — an open-weights student of TypeSafe Jev 1.13 (Qwen3.5-9B, calibrated *System One* decisions)
66
 
67
  > **Two different models, two different organisations — please read the names carefully.**
68
  > **TypeSafe Jev 1.13** is the hosted, closed model made by TypeSafe AI; it is the *teacher* whose public
69
- > outputs this model was trained on. **autotrust/JEV** (this repository) is an independent open-weights
70
  > *student* built by autotrust from the Apache-2.0 corpus
71
  > [`SargeDev/jev-distill-corpus-v3`](https://huggingface.co/datasets/SargeDev/jev-distill-corpus-v3).
72
- > autotrust/JEV is not affiliated with, endorsed by, or a product of TypeSafe AI. Throughout this card the
73
- > teacher is always written *TypeSafe Jev 1.13* and this model is always written *autotrust/JEV*.
74
 
75
- **One forward pass in, a calibrated probability distribution out.** autotrust/JEV answers typed questions
76
  (`noul` yes/no · `choice` over 2–16 options · `score` on a 0–5 scale) about any text or JSON state.
77
  No JSON parsing, no prompt engineering — it was distilled from the full output distributions of
78
  **TypeSafe Jev 1.13** onto a pretrained **Qwen3.5-9B** backbone. The bundle ships **two heads on one
@@ -81,14 +81,14 @@ generation). The backbone weights on disk are **bit-identical to Qwen3.5-9B**; t
81
  separate 154 MB LoRA that is only applied for decisions — so generation quality is exactly the base model's.
82
 
83
  **Model family** — same recipe, same API, same two-head packaging:
84
- **autotrust/JEV** (this repository, Qwen3.5-9B, 18 GB) · [autotrust/JEV-27B](https://huggingface.co/autotrust/JEV-27B)
85
  (Qwen3.8-27B, 54 GB: KL 0.019, choice 90.3 %, out-of-distribution KL 0.104, HumanEval 78.0 %).
86
 
87
  ## Headline results
88
 
89
- Held-out test set: 29,955 questions, 53 domains, all three question types. Numbers are for **autotrust/JEV**.
90
 
91
- | What is measured | autotrust/JEV | How to read it |
92
  |---|---|---|
93
  | **How closely it reproduces the teacher's probabilities** — mean KL divergence, lower is better, 0 = identical | **0.021** | The student's probability distributions are nearly indistinguishable from TypeSafe Jev 1.13's |
94
  | **Yes/no questions** (`noul`) — AUROC, higher is better | **0.996** | Ranks true vs. false almost perfectly |
@@ -121,7 +121,7 @@ financial tweets, under three kinds of pressure: more options, near-miss options
121
  are human gold labels, none of this data is in our training set, and TypeSafe Jev 1.13's results are published with
122
  it. We re-ran the same items, up to our 16-option limit.
123
 
124
- | | TypeSafe Jev 1.13 (published) | **autotrust/JEV** (9 B) | autotrust/JEV-27B |
125
  |---|---|---|---|
126
  | Accuracy with 2 / 4 / 8 / 16 options (800 items, 4 domains) | 0.890 / 0.801 / 0.782 / 0.769 | **0.868 / 0.774 / 0.735 / 0.694** | 0.876 / 0.784 / 0.767 / 0.740 |
127
  | 16 options — CLINC / DBpedia / GoEmotions / MTOP | 0.945 / 0.900 / 0.470 / 0.760 | **0.875 / 0.855 / 0.325 / 0.720** | 0.930 / 0.885 / 0.415 / 0.730 |
@@ -141,14 +141,14 @@ seeded differently, so compare the aggregates, not individual items.
141
  The expected answers were written by hand before the model was run. These are illustrations (≈ 110 decisions), not a
142
  benchmark.
143
 
144
- | task | autotrust/JEV (9 B) | autotrust/JEV-27B |
145
  |---|---|---|
146
  | Topic of 19 HN front-page stories (10 options) + "is it about AI?" | **38 / 38** | 38 / 38 |
147
  | 12 comments from a heated HN thread: "insults or attacks someone?" + "what is it mainly doing?" (6 options) | **22 / 24** | 23 / 24 |
148
  | 10 V2EX hot posts in **Chinese**: "contains a referral / invite code?" + "promotes a product or paid offer?" | **18 / 19** | 19 / 19 |
149
  | Community use cases: code-rule checks in the style of `adhere`, injection filtering, ticket routing, phishing, code-review diffs, urgency scores | **14 / 15** | 15 / 15 |
150
 
151
- | input | question | autotrust/JEV (9 B) |
152
  |---|---|---|
153
  | HN: "Two-tier encryption in the UK" | topic (10 options) | security & privacy · 0.87 |
154
  | HN: "Using LLMs to trace alchemical knowledge and decode 17th century letters" | about AI? | P(true) = 0.88 |
@@ -187,9 +187,9 @@ decision head. `adapter_vllm/` contains the backbone LoRA plus the 24-slot decis
187
  ### 1 — Start the server (OpenAI-compatible)
188
 
189
  ```bash
190
- hf download autotrust/JEV --local-dir JEV # ~18 GB
191
- vllm serve JEV --served-model-name autotrust/JEV \
192
- --enable-lora --max-lora-rank 32 --lora-modules jev-decision=JEV/adapter_vllm \
193
  --logprobs-mode processed_logprobs --max-model-len 4096
194
  ```
195
 
@@ -201,7 +201,7 @@ state (see notes below).
201
 
202
  ```bash
203
  curl -s localhost:8000/v1/chat/completions -H 'content-type: application/json' -d '{
204
- "model": "autotrust/JEV",
205
  "messages": [{"role": "user", "content": "In one sentence, what is safety stock?"}],
206
  "max_tokens": 60, "chat_template_kwargs": {"enable_thinking": false}}'
207
  ```
@@ -212,7 +212,7 @@ curl -s localhost:8000/v1/chat/completions -H 'content-type: application/json' -
212
  import json, math, requests
213
  from huggingface_hub import hf_hub_download
214
 
215
- REPO, URL = "autotrust/JEV", "http://localhost:8000"
216
  dh = json.load(open(hf_hub_download(REPO, "adapter_vllm/decision_head.json"))) # bias + verbalizer token ids
217
  T = json.load(open(hf_hub_download(REPO, "calibration.json")))["per_kind"] # per-kind temperatures
218
 
@@ -248,8 +248,8 @@ that vLLM applies cancels out, so the result is exactly the decision head's cali
248
  from vllm import LLM, SamplingParams
249
  from vllm.lora.request import LoRARequest
250
 
251
- llm = LLM("JEV", enable_lora=True, max_lora_rank=32, logprobs_mode="processed_logprobs", max_model_len=4096)
252
- decision = LoRARequest("jev-decision", 1, "JEV/adapter_vllm")
253
 
254
  gen = llm.generate(["..."], SamplingParams(temperature=0.0, max_tokens=256)) # base model, no LoRA
255
  dec = llm.generate([prompt], [SamplingParams(max_tokens=1, temperature=1.0, # decision head
@@ -345,7 +345,7 @@ from peft import PeftModel
345
  from safetensors.torch import load_file
346
  from transformers import AutoModelForCausalLM, AutoTokenizer
347
 
348
- repo = "autotrust/JEV"
349
  tok = AutoTokenizer.from_pretrained(repo)
350
  base = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="cuda") # == Qwen3.5-9B text model
351
 
@@ -386,7 +386,7 @@ inside `with model.disable_adapter():`.
386
  ## Why start from a pretrained LLM
387
 
388
  A System One judge is "just a classifier", so one could train a small encoder from scratch or fine-tune a
389
- BERT-class model. autotrust/JEV starts from a pretrained 9 B LLM instead, and that choice — not the 3 GPU-hours of
390
  distillation — is where most of the result above comes from:
391
 
392
  1. **Pretraining does most of the work; distillation only sharpens.** The 24-slot head is *initialised from
@@ -402,7 +402,7 @@ distillation — is where most of the result above comes from:
402
  after the first 64 k rows validation KL was already below 0.10.
403
 
404
  3. **Transfer to unseen task families.** On the Open-Jev *OOD* split (game-state and workflow-policy
405
- decisions never seen in training) autotrust/JEV reaches KL 0.234 and top-1 0.918 — a pretrained LLM reads the
406
  *content* of a new task instead of matching surface patterns of the 53 training domains.
407
 
408
  4. **Reads real, long, structured states.** States in the corpus are prose, JSON game states, policy
@@ -425,7 +425,7 @@ distillation — is where most of the result above comes from:
425
 
426
  ### Additional metrics (`test_set_30k`, temperature applied)
427
 
428
- | metric | autotrust/JEV |
429
  |---|---|
430
  | `noul` Brier score against the teacher's probability (lower is better) | 0.0015 |
431
  | `score` ranked probability score (lower is better) | 0.0085 |
@@ -440,7 +440,7 @@ distillation — is where most of the result above comes from:
440
  | weights | pass@1 | note |
441
  |---|---|---|
442
  | Qwen3.5-9B (base) | 70.7 % (116/164) | same loader and protocol as below |
443
- | **autotrust/JEV — generation path (backbone + `lm_head`, adapter off)** | **70.7 % (116/164)** | all 164 completions byte-identical to the base model |
444
  | LoRA folded into the backbone + base `lm_head` (*not shipped*) | 61.6 % (101/164) | why the adapter is kept separate |
445
 
446
  ### Per source × primitive (`test_set_30k`, temperature applied)
@@ -466,7 +466,7 @@ Validation KL on a fixed 4 k-row subset; `test_set_30k` metrics after calibratio
466
 
467
  | stage | rows seen | val KL | t30k KL | choice top-1 | score MAE | noul AUROC | ECE |
468
  |---|---|---|---|---|---|---|---|
469
- | **autotrust/JEV v0.8.0 — released weights (4,750 steps ≈ 0.93 epoch, LR annealed to ≈ 0.07×)** | 608 k | **0.019** | **0.0210** | **0.898** | **0.103** | **0.996** | **0.0007** |
470
  | v0.7.0 (step 2000 + 500-step LR cool-down) | 320 k | 0.026 | 0.0276 | 0.884 | 0.119 | 0.994 | 0.0014 |
471
  | step 2000 | 256 k | 0.0325 | 0.037 | 0.865 | 0.143 | 0.992 | 0.004 |
472
  | step 1500 | 192 k | 0.038 | 0.040 | 0.861 | 0.151 | 0.991 | 0.0025 |
@@ -495,9 +495,9 @@ another 24 % and added 1.4 points of choice agreement. Checkpoints taken at high
495
 
496
  ## Limitations
497
 
498
- * **autotrust/JEV mirrors TypeSafe Jev 1.13, including its mistakes.** This is a distillation, not an independent judge: where
499
- the teacher was wrong or uncalibrated, so is autotrust/JEV. Published evaluations of the teacher show it is unreliable
500
- for multi-hop reasoning, arithmetic, dates, counting, and adversarial inputs; autotrust/JEV inherits all of that. It
501
  cannot generate text or explanations *as a decision model*; the generation head is simply the unmodified base
502
  model and knows nothing about the decisions it is packaged with.
503
  * **It inherits the teacher's blind spots, and at 9 B adds some of its own.** On fresh inputs it shoves a poker spot
@@ -507,7 +507,7 @@ another 24 % and added 1.4 points of choice agreement. Checkpoints taken at high
507
  * **Choice agreement is capped by teacher ambiguity.** The teacher's `choice` distributions are soft (median
508
  top-1 probability 0.70). On the 14 % of rows where the teacher's top-2 options are within 0.1 of each other,
509
  argmax agreement is near chance for *any* faithful mimic (0.46 where the gap is < 0.05). On teacher-decisive
510
- rows agreement is 0.954, and autotrust/JEV's argmax captures 97.7 % of the teacher probability mass a perfect mimic
511
  could (0.693 vs 0.709).
512
  * **Fixed option sets.** `noul` and `score` accept only their canonical options; `choice` accepts 2–16
513
  options. Inputs longer than 1,024 tokens are truncated (state only, head 60 % / tail 40 %) at serving unless
@@ -515,7 +515,7 @@ another 24 % and added 1.4 points of choice agreement. Checkpoints taken at high
515
  * **English-centric.** The corpus is English; multilingual behaviour is inherited from the backbone but was not
516
  measured.
517
  * **Placeholder labels in the corpus.** The `yuri_v1` memory-relevance stream is 100 % exact-uniform
518
- `[0.5, 0.5]` — those rows teach nothing about relevance. autotrust/JEV outputs ≈ 0.5 on them by design; do not use it
519
  for memory-relevance scoring without further training.
520
  * **Not for high-stakes decisions.** Use confidence gating: act automatically only above a threshold you
521
  validated on your own data, and route the rest to a stronger model or a human.
@@ -541,14 +541,14 @@ reports/ evaluation reports: test-set evaluation,
541
 
542
  Weights: **Apache-2.0** (base model `Qwen/Qwen3.5-9B` is Apache-2.0; training corpus
543
  `SargeDev/jev-distill-corpus-v3` is Apache-2.0, its `openjev_v2` stream additionally CC0). The System One
544
- framing and the `noul` / `choice` / `score` primitives originate with TypeSafe AI's Jev; autotrust/JEV is an
545
  independent student model trained on public data and shares no weights, code or affiliation with TypeSafe AI.
546
 
547
  ```bibtex
548
  @misc{autotrust_jev_2026,
549
- title = {autotrust/JEV: an open-weights student of TypeSafe Jev 1.13 for calibrated System One decisions (Qwen3.5-9B)},
550
  author = {autotrust},
551
  year = {2026},
552
- url = {https://huggingface.co/autotrust/JEV}
553
  }
554
  ```
 
29
  - brier
30
  - ece
31
  model-index:
32
+ - name: autotrust/JEV-9B (student of TypeSafe Jev 1.13)
33
  results:
34
  - task:
35
  type: text-classification
 
62
  value: 0.954
63
  ---
64
 
65
+ # autotrust/JEV-9B — an open-weights student of TypeSafe Jev 1.13 (Qwen3.5-9B, calibrated *System One* decisions)
66
 
67
  > **Two different models, two different organisations — please read the names carefully.**
68
  > **TypeSafe Jev 1.13** is the hosted, closed model made by TypeSafe AI; it is the *teacher* whose public
69
+ > outputs this model was trained on. **autotrust/JEV-9B** (this repository) is an independent open-weights
70
  > *student* built by autotrust from the Apache-2.0 corpus
71
  > [`SargeDev/jev-distill-corpus-v3`](https://huggingface.co/datasets/SargeDev/jev-distill-corpus-v3).
72
+ > autotrust/JEV-9B is not affiliated with, endorsed by, or a product of TypeSafe AI. Throughout this card the
73
+ > teacher is always written *TypeSafe Jev 1.13* and this model is always written *autotrust/JEV-9B*.
74
 
75
+ **One forward pass in, a calibrated probability distribution out.** autotrust/JEV-9B answers typed questions
76
  (`noul` yes/no · `choice` over 2–16 options · `score` on a 0–5 scale) about any text or JSON state.
77
  No JSON parsing, no prompt engineering — it was distilled from the full output distributions of
78
  **TypeSafe Jev 1.13** onto a pretrained **Qwen3.5-9B** backbone. The bundle ships **two heads on one
 
81
  separate 154 MB LoRA that is only applied for decisions — so generation quality is exactly the base model's.
82
 
83
  **Model family** — same recipe, same API, same two-head packaging:
84
+ **autotrust/JEV-9B** (this repository, Qwen3.5-9B, 18 GB) · [autotrust/JEV-27B](https://huggingface.co/autotrust/JEV-27B)
85
  (Qwen3.8-27B, 54 GB: KL 0.019, choice 90.3 %, out-of-distribution KL 0.104, HumanEval 78.0 %).
86
 
87
  ## Headline results
88
 
89
+ Held-out test set: 29,955 questions, 53 domains, all three question types. Numbers are for **autotrust/JEV-9B**.
90
 
91
+ | What is measured | autotrust/JEV-9B | How to read it |
92
  |---|---|---|
93
  | **How closely it reproduces the teacher's probabilities** — mean KL divergence, lower is better, 0 = identical | **0.021** | The student's probability distributions are nearly indistinguishable from TypeSafe Jev 1.13's |
94
  | **Yes/no questions** (`noul`) — AUROC, higher is better | **0.996** | Ranks true vs. false almost perfectly |
 
121
  are human gold labels, none of this data is in our training set, and TypeSafe Jev 1.13's results are published with
122
  it. We re-ran the same items, up to our 16-option limit.
123
 
124
+ | | TypeSafe Jev 1.13 (published) | **autotrust/JEV-9B** | autotrust/JEV-27B |
125
  |---|---|---|---|
126
  | Accuracy with 2 / 4 / 8 / 16 options (800 items, 4 domains) | 0.890 / 0.801 / 0.782 / 0.769 | **0.868 / 0.774 / 0.735 / 0.694** | 0.876 / 0.784 / 0.767 / 0.740 |
127
  | 16 options — CLINC / DBpedia / GoEmotions / MTOP | 0.945 / 0.900 / 0.470 / 0.760 | **0.875 / 0.855 / 0.325 / 0.720** | 0.930 / 0.885 / 0.415 / 0.730 |
 
141
  The expected answers were written by hand before the model was run. These are illustrations (≈ 110 decisions), not a
142
  benchmark.
143
 
144
+ | task | autotrust/JEV-9B | autotrust/JEV-27B |
145
  |---|---|---|
146
  | Topic of 19 HN front-page stories (10 options) + "is it about AI?" | **38 / 38** | 38 / 38 |
147
  | 12 comments from a heated HN thread: "insults or attacks someone?" + "what is it mainly doing?" (6 options) | **22 / 24** | 23 / 24 |
148
  | 10 V2EX hot posts in **Chinese**: "contains a referral / invite code?" + "promotes a product or paid offer?" | **18 / 19** | 19 / 19 |
149
  | Community use cases: code-rule checks in the style of `adhere`, injection filtering, ticket routing, phishing, code-review diffs, urgency scores | **14 / 15** | 15 / 15 |
150
 
151
+ | input | question | autotrust/JEV-9B |
152
  |---|---|---|
153
  | HN: "Two-tier encryption in the UK" | topic (10 options) | security & privacy · 0.87 |
154
  | HN: "Using LLMs to trace alchemical knowledge and decode 17th century letters" | about AI? | P(true) = 0.88 |
 
187
  ### 1 — Start the server (OpenAI-compatible)
188
 
189
  ```bash
190
+ hf download autotrust/JEV-9B --local-dir JEV-9B # ~18 GB
191
+ vllm serve JEV-9B --served-model-name autotrust/JEV-9B \
192
+ --enable-lora --max-lora-rank 32 --lora-modules jev-decision=JEV-9B/adapter_vllm \
193
  --logprobs-mode processed_logprobs --max-model-len 4096
194
  ```
195
 
 
201
 
202
  ```bash
203
  curl -s localhost:8000/v1/chat/completions -H 'content-type: application/json' -d '{
204
+ "model": "autotrust/JEV-9B",
205
  "messages": [{"role": "user", "content": "In one sentence, what is safety stock?"}],
206
  "max_tokens": 60, "chat_template_kwargs": {"enable_thinking": false}}'
207
  ```
 
212
  import json, math, requests
213
  from huggingface_hub import hf_hub_download
214
 
215
+ REPO, URL = "autotrust/JEV-9B", "http://localhost:8000"
216
  dh = json.load(open(hf_hub_download(REPO, "adapter_vllm/decision_head.json"))) # bias + verbalizer token ids
217
  T = json.load(open(hf_hub_download(REPO, "calibration.json")))["per_kind"] # per-kind temperatures
218
 
 
248
  from vllm import LLM, SamplingParams
249
  from vllm.lora.request import LoRARequest
250
 
251
+ llm = LLM("JEV-9B", enable_lora=True, max_lora_rank=32, logprobs_mode="processed_logprobs", max_model_len=4096)
252
+ decision = LoRARequest("jev-decision", 1, "JEV-9B/adapter_vllm")
253
 
254
  gen = llm.generate(["..."], SamplingParams(temperature=0.0, max_tokens=256)) # base model, no LoRA
255
  dec = llm.generate([prompt], [SamplingParams(max_tokens=1, temperature=1.0, # decision head
 
345
  from safetensors.torch import load_file
346
  from transformers import AutoModelForCausalLM, AutoTokenizer
347
 
348
+ repo = "autotrust/JEV-9B"
349
  tok = AutoTokenizer.from_pretrained(repo)
350
  base = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="cuda") # == Qwen3.5-9B text model
351
 
 
386
  ## Why start from a pretrained LLM
387
 
388
  A System One judge is "just a classifier", so one could train a small encoder from scratch or fine-tune a
389
+ BERT-class model. autotrust/JEV-9B starts from a pretrained 9 B LLM instead, and that choice — not the 3 GPU-hours of
390
  distillation — is where most of the result above comes from:
391
 
392
  1. **Pretraining does most of the work; distillation only sharpens.** The 24-slot head is *initialised from
 
402
  after the first 64 k rows validation KL was already below 0.10.
403
 
404
  3. **Transfer to unseen task families.** On the Open-Jev *OOD* split (game-state and workflow-policy
405
+ decisions never seen in training) autotrust/JEV-9B reaches KL 0.234 and top-1 0.918 — a pretrained LLM reads the
406
  *content* of a new task instead of matching surface patterns of the 53 training domains.
407
 
408
  4. **Reads real, long, structured states.** States in the corpus are prose, JSON game states, policy
 
425
 
426
  ### Additional metrics (`test_set_30k`, temperature applied)
427
 
428
+ | metric | autotrust/JEV-9B |
429
  |---|---|
430
  | `noul` Brier score against the teacher's probability (lower is better) | 0.0015 |
431
  | `score` ranked probability score (lower is better) | 0.0085 |
 
440
  | weights | pass@1 | note |
441
  |---|---|---|
442
  | Qwen3.5-9B (base) | 70.7 % (116/164) | same loader and protocol as below |
443
+ | **autotrust/JEV-9B — generation path (backbone + `lm_head`, adapter off)** | **70.7 % (116/164)** | all 164 completions byte-identical to the base model |
444
  | LoRA folded into the backbone + base `lm_head` (*not shipped*) | 61.6 % (101/164) | why the adapter is kept separate |
445
 
446
  ### Per source × primitive (`test_set_30k`, temperature applied)
 
466
 
467
  | stage | rows seen | val KL | t30k KL | choice top-1 | score MAE | noul AUROC | ECE |
468
  |---|---|---|---|---|---|---|---|
469
+ | **autotrust/JEV-9B v0.8.0 — released weights (4,750 steps ≈ 0.93 epoch, LR annealed to ≈ 0.07×)** | 608 k | **0.019** | **0.0210** | **0.898** | **0.103** | **0.996** | **0.0007** |
470
  | v0.7.0 (step 2000 + 500-step LR cool-down) | 320 k | 0.026 | 0.0276 | 0.884 | 0.119 | 0.994 | 0.0014 |
471
  | step 2000 | 256 k | 0.0325 | 0.037 | 0.865 | 0.143 | 0.992 | 0.004 |
472
  | step 1500 | 192 k | 0.038 | 0.040 | 0.861 | 0.151 | 0.991 | 0.0025 |
 
495
 
496
  ## Limitations
497
 
498
+ * **autotrust/JEV-9B mirrors TypeSafe Jev 1.13, including its mistakes.** This is a distillation, not an independent judge: where
499
+ the teacher was wrong or uncalibrated, so is autotrust/JEV-9B. Published evaluations of the teacher show it is unreliable
500
+ for multi-hop reasoning, arithmetic, dates, counting, and adversarial inputs; autotrust/JEV-9B inherits all of that. It
501
  cannot generate text or explanations *as a decision model*; the generation head is simply the unmodified base
502
  model and knows nothing about the decisions it is packaged with.
503
  * **It inherits the teacher's blind spots, and at 9 B adds some of its own.** On fresh inputs it shoves a poker spot
 
507
  * **Choice agreement is capped by teacher ambiguity.** The teacher's `choice` distributions are soft (median
508
  top-1 probability 0.70). On the 14 % of rows where the teacher's top-2 options are within 0.1 of each other,
509
  argmax agreement is near chance for *any* faithful mimic (0.46 where the gap is < 0.05). On teacher-decisive
510
+ rows agreement is 0.954, and autotrust/JEV-9B's argmax captures 97.7 % of the teacher probability mass a perfect mimic
511
  could (0.693 vs 0.709).
512
  * **Fixed option sets.** `noul` and `score` accept only their canonical options; `choice` accepts 2–16
513
  options. Inputs longer than 1,024 tokens are truncated (state only, head 60 % / tail 40 %) at serving unless
 
515
  * **English-centric.** The corpus is English; multilingual behaviour is inherited from the backbone but was not
516
  measured.
517
  * **Placeholder labels in the corpus.** The `yuri_v1` memory-relevance stream is 100 % exact-uniform
518
+ `[0.5, 0.5]` — those rows teach nothing about relevance. autotrust/JEV-9B outputs ≈ 0.5 on them by design; do not use it
519
  for memory-relevance scoring without further training.
520
  * **Not for high-stakes decisions.** Use confidence gating: act automatically only above a threshold you
521
  validated on your own data, and route the rest to a stronger model or a human.
 
541
 
542
  Weights: **Apache-2.0** (base model `Qwen/Qwen3.5-9B` is Apache-2.0; training corpus
543
  `SargeDev/jev-distill-corpus-v3` is Apache-2.0, its `openjev_v2` stream additionally CC0). The System One
544
+ framing and the `noul` / `choice` / `score` primitives originate with TypeSafe AI's Jev; autotrust/JEV-9B is an
545
  independent student model trained on public data and shares no weights, code or affiliation with TypeSafe AI.
546
 
547
  ```bibtex
548
  @misc{autotrust_jev_2026,
549
+ title = {autotrust/JEV-9B: an open-weights student of TypeSafe Jev 1.13 for calibrated System One decisions (Qwen3.5-9B)},
550
  author = {autotrust},
551
  year = {2026},
552
+ url = {https://huggingface.co/autotrust/JEV-9B}
553
  }
554
  ```