AgentJev 0.6B

Exact model and requirements

Artifact:

aimeigaoshou (native release), malevrigns (official source) / AgentJev v1 complete trained Qwen3 backbone and candidate decision head · checkpoint aimeigaoshou/agent-jev v1

Publisher revision:

7d433994fbde17a3f0993c2f2b02fe8ca1370db1

Artifact digest:

sha256:8166e46dc6019ae13f0d8fc97d603cdbcc2d20eb1882c7350c8deb9fc6bae215

Download size:

1.11 GiB

Memory requirement (unknown):

Unknown — No measured peak for both networks, CPU tensors, native decoding, dependencies and serialization.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; path microbatch one, serial HTTP request handling. Complete trained backbone and head on CPU; no separate language decoder, base weights, GPU or offload.

KV cache:

Complete independent causal paths use native backbone use_cache=False; no autoregressive output decoding or shared-prefix compute is requested.

Workload:

2,048 token context; Unknown; Not applicable; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Exact v1 local Qwen3 tokenizer, native path framing without added special tokens · 7d433994fbde17a3f0993c2f2b02fe8ca1370db1

Sampling / reasoning:

No autoregressive sampler. Exact pinned boolean, choice and score temperature scalars are loaded and checked against the complete native calibration file; these are relative head distributions, not arbitrary-domain calibration or action-success probabilities. Each candidate is a complete causal [STATE], [QUESTION], [CANDIDATE] path. The bare v1 strict safetensors route has no training config and rejects paths beyond 2048 rather than truncating. Native config positional capacity 32768 is distinct from this service bound. Boolean returns true mass and threshold; choice returns relative argmax, top probability and margin; score returns ordered expected level plus argmax. Transport identifiers are not semantic inputs; no tool or action executes.

Runtime prerequisites:

Ubuntu24.04 x86_64; See prerequisites above.

Official artifact · License

The native source contains Apache 2.0 and the exact v1 model card declares Apache 2.0; the snapshot lists no separate license file. Retain applicable copyright, license, notices and modification notices; Apache trademark and patent conditions apply. The identified Qwen3 upstream currently publishes Apache 2.0 too, but its training revision and Base versus non-Base labeling are not proven by these sources. Source access is not legal clearance or permission for unrelated data. No decision accuracy or fitness guarantee.

• temperatures.json · 0 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1
• config.json · 0 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1
• tokenizer.json · 0.01 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1
• tokenizer_config.json · 0 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1
• vocab.json · 0 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1
• merges.txt · 0 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1

• Native model file is 1,196,881,242 bytes; all seven required release assets total 1,208,372,367 bytes. These are artifact sizes, not RAM requirements.
• Publisher verifier counts 598,418,689 parameter elements; their FP32 storage lower bound is 2,393,674,756 bytes before loading copies, activations, masks, tokenizer, libraries and OS. Peak and fit remain unknown.
• Four CPU threads, 32 GiB RAM and 40 GiB free disk are planning allocations, not measured minima.

Install the runtime

1. 1. Install the exact native CPU dependency path

   Use the official hashed CPU wheel and complete pinned source package without stripping dependencies. The imported native losses, data, model, encoders and service path are source-audited. Eager Qwen3 attention is selected through the released Transformers API; no CUDA driver or flash-attention package is required. Record the resolved environment; target execution remains unverified.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv git curl ca-certificates openssl
   mkdir -p "$HOME/agentjev-v1-hosting/model"
   cd "$HOME/agentjev-v1-hosting"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install 'https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91'
   git clone https://github.com/malevrigns/agent-jev.git runtime
   git -C runtime checkout --detach 3b0042a6171124c5d2942748ae01f453ed2cce61
   test "$(git -C runtime rev-parse HEAD)" = '3b0042a6171124c5d2942748ae01f453ed2cce61'
   .venv/bin/python -m pip install 'transformers==4.57.6' 'numpy==1.26.4' 'PyYAML==6.0.2' 'tqdm==4.67.1' 'safetensors==0.6.2' -e ./runtime
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Pinned source checkout, direct package versions and pip check succeed; resolved environment saved. Stop on any dependency failure.

Download the model

1. 1. Download all seven exact v1 assets and notices

   All hashes and sizes come from the native release verifier and pinned public metadata. Download no alternate branch. The wrapper checks every size and digest and the publisher BF16 header composition before importing the native loader. HF API F32 summary disagrees with the verifier; a mismatch is a stop, not permission to relabel the release.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   curl --fail --location --output model/model.safetensors 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/model.safetensors'
   curl --fail --location --output model/temperatures.json 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/temperatures.json'
   curl --fail --location --output model/config.json 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/config.json'
   curl --fail --location --output model/tokenizer.json 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/tokenizer.json'
   curl --fail --location --output model/tokenizer_config.json 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/tokenizer_config.json'
   curl --fail --location --output model/vocab.json 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/vocab.json'
   curl --fail --location --output model/merges.txt 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/merges.txt'
   sha256sum --check <<'SHA256'
   8166e46dc6019ae13f0d8fc97d603cdbcc2d20eb1882c7350c8deb9fc6bae215  model/model.safetensors
   5e5032896d77c72feb86b4725abe418488b8a1c7476014c706bd8215eb43c86c  model/temperatures.json
   504a6b58c4271583724e66584b6b7698aea18450209df6b2f7582df0e89cee59  model/config.json
   c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539  model/tokenizer.json
   3c04ed3ca964ea2f6b2b5faf0dc4d31aec1cb1e8b4bcf63f402d295046b422b5  model/tokenizer_config.json
   ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910  model/vocab.json
   8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5  model/merges.txt
   SHA256
   curl --fail --location --output AGENTJEV-LICENSE 'https://raw.githubusercontent.com/malevrigns/agent-jev/3b0042a6171124c5d2942748ae01f453ed2cce61/LICENSE'
   printf '%s  %s\n' 'f4d9b0aad8ff9e650f5890fa3eeb7c1311479cab901f8a9b416a2d9acaf54266' 'AGENTJEV-LICENSE' | sha256sum --check
   curl --fail --location --output QWEN-UPSTREAM-LICENSE 'https://huggingface.co/Qwen/Qwen3-0.6B-Base/resolve/da87bfb608c14b7cf20ba1ce41287e8de496c0cd/LICENSE'
   printf '%s  %s\n' '832dd9e00a68dd83b3c3fb9f5588dad7dcf337a0db50f7d9483f310cd292e92e' 'QWEN-UPSTREAM-LICENSE' | sha256sum --check

   Expected result: Every asset and both retained notices pass SHA checks. Native weights have not been fetched by the author.

Configure

1. 1. Write the manifest, original fixture and persistent key

   Create a key only when absent; an empty existing key must fail rather than silently rotate. Protect files and review the original synthetic state and criteria.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   cat > release-manifest.json <<'MANIFEST'
   {
     "repository": "aimeigaoshou/agent-jev",
     "revision": "7d433994fbde17a3f0993c2f2b02fe8ca1370db1",
     "files": [
       {
         "path": "model.safetensors",
         "bytes": 1196881242,
         "sha256": "8166e46dc6019ae13f0d8fc97d603cdbcc2d20eb1882c7350c8deb9fc6bae215"
       },
       {
         "path": "temperatures.json",
         "bytes": 389,
         "sha256": "5e5032896d77c72feb86b4725abe418488b8a1c7476014c706bd8215eb43c86c"
       },
       {
         "path": "config.json",
         "bytes": 727,
         "sha256": "504a6b58c4271583724e66584b6b7698aea18450209df6b2f7582df0e89cee59"
       },
       {
         "path": "tokenizer.json",
         "bytes": 7031645,
         "sha256": "c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539"
       },
       {
         "path": "tokenizer_config.json",
         "bytes": 9678,
         "sha256": "3c04ed3ca964ea2f6b2b5faf0dc4d31aec1cb1e8b4bcf63f402d295046b422b5"
       },
       {
         "path": "vocab.json",
         "bytes": 2776833,
         "sha256": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"
       },
       {
         "path": "merges.txt",
         "bytes": 1671853,
         "sha256": "8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5"
       }
     ]
   }
   MANIFEST
   cat > fixture.json <<'FIXTURE'
   {
     "id": "original-library-schedule",
     "state": "A fictional community library opens weekdays 9am–6pm, Saturday 10am–4pm, and is closed Sunday. These are original synthetic facts for a low-stakes demonstration.",
     "questions": [
       {
         "id": "sunday",
         "type": "boolean",
         "question": "Is the library open on Sunday according to this state?",
         "criteria": {
           "true": "The supplied state says it is open Sunday.",
           "false": "The supplied state says it is closed Sunday."
         }
       },
       {
         "id": "saturday-hours",
         "type": "choice",
         "question": "How many hours is the library open on Saturday?",
         "options": {
           "four": "Four hours.",
           "six": "Six hours.",
           "nine": "Nine hours."
         }
       },
       {
         "id": "weekday-duration",
         "type": "score",
         "question": "Rate the weekday opening duration by the ordered definitions.",
         "levels": [
           "At most four hours open per weekday.",
           "More than four but at most eight hours open per weekday.",
           "More than eight hours open per weekday."
         ]
       }
     ]
   }
   FIXTURE
   if [ ! -e .api-key ]; then (umask 077; openssl rand -hex 32 > .api-key); fi
   test -s .api-key
   chmod 600 .api-key

   Expected result: Manifest and fictional fixture saved; private nonempty key retained.

2. 2. Write request and result boundaries

   Reader-authored service code is independent of publisher HTTP serving code. It rejects unknown fields, duplicate JSON keys, unsupported content framing and malformed probability algebra. No publisher model code executes while writing this file.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   cat > reader-decision-server.py <<'PYWRAPPER'
   """Reader-authored serial private AgentJev decision service; no native imports at import time."""
   import hashlib,hmac,json,math,pathlib,re,socket,struct
   from http.server import HTTPServer,BaseHTTPRequestHandler
   PORT=8999
   EXPECTED_WEIGHT='8166e46dc6019ae13f0d8fc97d603cdbcc2d20eb1882c7350c8deb9fc6bae215'
   def digest(path):
    h=hashlib.sha256()
    with open(path,'rb')as f:
     for b in iter(lambda:f.read(4*1024*1024),b''):h.update(b)
    return h.hexdigest()
   def verify_files(root,manifest):
    for row in manifest:
     f=root/row['path']
     if not f.is_file()or f.stat().st_size!=row['bytes']or digest(f)!=row['sha256']:raise ValueError('Pinned release asset mismatch')
   def header_check(path):
    with open(path,'rb')as f:
     raw=f.read(8)
     if len(raw)!=8:raise ValueError('Missing safetensors header')
     n=struct.unpack('<Q',raw)[0]
     if not 1<=n<=1048576:raise ValueError('Header bound')
     h=json.loads(f.read(n))
    tensors=[v for k,v in h.items()if k!='__metadata__']
    if len(tensors)!=343 or any(v.get('dtype')!='BF16'for v in tensors):raise ValueError('Expected exact publisher v1 BF16 header; metadata discrepancy is not waived')
    total=0
    for v in tensors:
     shape=v.get('shape')
     if not isinstance(shape,list)or any(type(x)is not int or x<0 for x in shape):raise ValueError('Invalid shape')
     total+=math.prod(shape)
    if total!=598418689:raise ValueError('Full parameter composition mismatch')
   def parse_request(raw):
    def unique(pairs):
     d={}
     for k,v in pairs:
      if k in d:raise ValueError('Duplicate field')
      d[k]=v
     return d
    p=json.loads(raw.decode('utf-8'),object_pairs_hook=unique,parse_constant=lambda _:(_ for _ in()).throw(ValueError('Nonfinite JSON')))
    if not isinstance(p,dict)or set(p)!={'id','state','questions'}:raise ValueError('Exact state request schema')
    def text(v,limit):
     if not isinstance(v,str)or not v.strip()or len(v.encode())>limit or any(x in v for x in ['[STATE]','[QUESTION]','[CANDIDATE]']):raise ValueError('Bounded plain text required')
    text(p['id'],80);text(p['state'],4096)
    if not isinstance(p['questions'],list)or not 1<=len(p['questions'])<=3:raise ValueError('One to three questions required')
    ids=set()
    for q in p['questions']:
     if not isinstance(q,dict):raise ValueError('Question object required')
     t=q.get('type')
     if not isinstance(t,str):raise ValueError('Typed primitive string required')
     extra={'boolean':'criteria','choice':'options','score':'levels'}.get(t)
     if extra is None or set(q)!={'id','type','question',extra}:raise ValueError('Exact typed question schema')
     text(q['id'],80);text(q['question'],1024)
     if q['id']in ids:raise ValueError('Duplicate question id')
     ids.add(q['id']);v=q[extra]
     if t=='boolean':
      if not isinstance(v,dict)or set(v)!={'true','false'}:raise ValueError('Both boolean criteria required')
      vals=list(v.values())
     elif t=='choice':
      if not isinstance(v,dict)or not 2<=len(v)<=8:raise ValueError('Two to eight choice options')
      for k in v:text(k,80)
      vals=list(v.values())
     else:
      if not isinstance(v,list)or not 2<=len(v)<=5:raise ValueError('Two to five ordered score levels')
      vals=v
     for v in vals:text(v,1024)
     if len(set(vals))!=len(vals):raise ValueError('Distinct descriptions required')
    return p
   def request_length(headers):
    vals=headers.get_all('Content-Length',[])
    if headers.get('Transfer-Encoding')is not None or len(vals)!=1 or not re.fullmatch('[0-9]{1,5}',vals[0]):raise ValueError('One bounded decimal length')
    n=int(vals[0])
    if not 1<=n<=16384:raise ValueError('Body outside limit')
    return n
   def validate_result(request,result):
    if result.get('api_version')!='agentjev.decision.v1':raise ValueError('Native API version mismatch')
    usage=result.get('usage',{})
    if any(type(usage.get(k))is not int or usage[k]!=0 for k in ['generated_tokens','truncated_inputs']):raise ValueError('No decoding or truncation allowed')
    rows=result.get('results');questions=request['questions']
    if not isinstance(rows,list)or len(rows)!=1 or rows[0].get('id')!=request['id']:raise ValueError('Native state identity mismatch')
    answers=rows[0].get('answers')
    if not isinstance(answers,list)or len(answers)!=len(questions):raise ValueError('Missing typed answers')
    paths=0
    for q,a in zip(questions,answers):
     t=q['type'];keys=['true','false']if t=='boolean'else list(q['options'])if t=='choice'else[str(i)for i in range(len(q['levels']))];paths+=len(keys)
     dist=a.get('distribution')
     if a.get('id')!=q['id']or a.get('type')!=t or not isinstance(dist,dict)or list(dist)!=keys:raise ValueError('Typed answer identity mismatch')
     probs=list(dist.values())
     if any(type(x)not in(int,float)or not math.isfinite(x)or not 0<=x<=1 for x in probs)or abs(sum(probs)-1)>1e-4:raise ValueError('Invalid finite distribution')
     best=max(range(len(probs)),key=probs.__getitem__)
     def close(a,b):return type(a)in(int,float)and math.isfinite(a)and abs(a-b)<=1e-6
     if t=='boolean':
      if type(a.get('value'))is not bool or a['value']!=(probs[0]>=.5)or not close(a.get('probability'),probs[0]):raise ValueError('Boolean semantics mismatch')
     elif t=='choice':
      ranked=sorted(probs,reverse=True)
      if a.get('value')!=keys[best]or a.get('description')!=q['options'][keys[best]]or not close(a.get('top_probability'),probs[best])or not close(a.get('margin'),ranked[0]-ranked[1]):raise ValueError('Relative choice semantics mismatch')
     elif type(a.get('level'))is not int or a['level']!=best or a.get('legend')!=q['levels']or not close(a.get('score'),sum(i*x for i,x in enumerate(probs))):raise ValueError('Ordered expected-score semantics mismatch')
    if usage.get('questions')!=len(questions)or usage.get('candidate_paths')!=paths or type(usage.get('input_path_tokens'))is not int or not paths<=usage['input_path_tokens']<=paths*2048:raise ValueError('Native path usage mismatch')
    return result
   PYWRAPPER

   Expected result: First wrapper portion saved.

3. 3. Write strict native loading and serial loopback HTTP

   Startup verifies every native file and header before loading the complete trained state with strict=True. It checks CPU FP32, all three calibration scalars and native identity/path mode. It has a 30 second socket-read timeout, no native inference deadline and no public-service hardening.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   cat >> reader-decision-server.py <<'PYWRAPPER'
   class Engine:
    def __init__(self,root):
     import os
     manifest=json.loads((root/'release-manifest.json').read_text())['files'];verify_files(root/'model',manifest);header_check(root/'model/model.safetensors')
     os.environ.update(HF_HUB_OFFLINE='1',TRANSFORMERS_OFFLINE='1')
     import torch
     from jev_service.engine import DecisionEngine
     from jev_service.contract import prepare,encode_paths
     torch.set_default_dtype(torch.float32);torch.set_num_threads(4);torch.set_num_interop_threads(1)
     self.engine=DecisionEngine(root/'model/model.safetensors',str(root/'model'),device='cpu',max_tokens=2048,path_batch=1,encoder='path',temperatures=root/'model/temperatures.json')
     self.engine.model.path_encoder.backbone.set_attn_implementation('eager')
     if self.engine.model.path_encoder.backbone.config._attn_implementation!='eager':raise ValueError('Explicit eager attention required')
     if any(p.is_floating_point()and p.dtype!=torch.float32 for p in self.engine.model.parameters()):raise ValueError('CPU FP32 parameters required')
     info=self.engine.info();expected=json.loads((root/'model/temperatures.json').read_text());temps={k:v['temperature']for k,v in expected.items()}
     if set(temps)!={'boolean','choice','score'}or temps!={'boolean':1.0717734625362931,'choice':1.0352649238413776,'score':1.0717734625362931}:raise ValueError('Complete exact calibration required')
     if info['api_version']!='agentjev.decision.v1'or info['checkpoint_sha256']!=EXPECTED_WEIGHT or info['encoder']!='path'or info['max_path_tokens']!=2048 or info['temperatures']!=temps or info['types']!=['boolean','choice','score']or info['output_token_decoding']is not False or info['shared_prefix_compute']is not False:raise ValueError('Native release/calibration identity mismatch')
     self.prepare=prepare;self.encode_paths=encode_paths;self.info=info
    def evaluate(self,request):
     paths,_,_=self.encode_paths(self.prepare(request),self.engine.tokenizer,2048)
     if not paths or any(not p or len(p)>2048 for p in paths):raise ValueError('Complete per-path bound')
     return validate_result(request,self.engine.evaluate(request))
   def serve(engine,key):
    class Handler(BaseHTTPRequestHandler):
     protocol_version='HTTP/1.0'
     def setup(self):super().setup();self.connection.settimeout(30)
     def log_message(self,*a):pass
     def reply(self,status,obj):
      b=json.dumps(obj,ensure_ascii=False,allow_nan=False).encode();self.send_response(status);self.send_header('Content-Type','application/json');self.send_header('Content-Length',str(len(b)));self.end_headers();self.wfile.write(b)
     def auth(self):
      v=self.headers.get_all('Authorization',[])
      try:return len(v)==1 and hmac.compare_digest(v[0].encode('ascii'),('Bearer '+key).encode('ascii'))
      except UnicodeError:return False
     def do_GET(self):
      if not self.auth():return self.reply(401,{'error':'Unauthorized'})
      if self.path!='/health':return self.reply(404,{'error':'Not found'})
      return self.reply(200,{'ready':True,'native':engine.info,'reader_path_microbatch':1,'reader_context_per_path':2048})
     def do_POST(self):
      if not self.auth():return self.reply(401,{'error':'Unauthorized'})
      if self.path!='/api/evaluate':return self.reply(404,{'error':'Not found'})
      try:
       if self.headers.get('Content-Type')!='application/json':raise ValueError('JSON required')
       n=request_length(self.headers);b=self.rfile.read(n)
       if len(b)!=n:raise ValueError('Incomplete body')
       req=parse_request(b)
      except(ValueError,UnicodeError,socket.timeout):return self.reply(400,{'error':'Invalid bounded request'})
      try:out=engine.evaluate(req)
      except Exception:return self.reply(422,{'error':'Native path or finite-distribution check failed; no action executed'})
      return self.reply(200,out)
    server=HTTPServer(('127.0.0.1',PORT),Handler)
    try:server.serve_forever()
    finally:server.server_close()
   if __name__=='__main__':
    root=pathlib.Path(__file__).resolve().parent;key=(root/'.api-key').read_text().strip()
    if not re.fullmatch('[a-f0-9]{64}',key):raise ValueError('Restore private service key')
    serve(Engine(root),key)
   PYWRAPPER
   printf '%s  %s\n' 'a997d981592ccc21575be799483bf41f65ce43990175e579936cc643d5d59cce' 'reader-decision-server.py' | sha256sum --check
   .venv/bin/python -m py_compile reader-decision-server.py

   Expected result: Exact wrapper SHA and Python syntax pass; no inference has run.

Start the model

1. 1. Start the private CPU decision service

   Run in a foreground shell. Four intra-op threads and one inter-op thread are explicit CPU settings. Complete independent candidate paths use native path microbatch one and no shared-prefix compute. Health means initialized and checked, not successful model inference.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   export OMP_NUM_THREADS=4 MKL_NUM_THREADS=4 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1
   .venv/bin/python reader-decision-server.py

   Expected result: Exact native checkpoint/calibration identity checks pass and port 8999 binds only to loopback. Any size, hash, header, state-key, dtype or attention failure aborts startup.

Verify the setup

1. 1. Verify authentication, native algebra and grounded fixture

   Run in a second shell on the same host. Require unauthorized health to return 401, then check native identity and all three primitives. The checker independently revalidates complete distributions and manual fixture answers: Sunday closed, Saturday six hours, weekday duration above eight hours. Score is the probability-weighted ordinal expectation, not forced to its argmax. Wrong preferred alternatives fail; passing one fixture is not accuracy evidence.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   .venv/bin/python - <<'PYVERIFY'
   import json,urllib.request,urllib.error,pathlib
   root=pathlib.Path.cwd();key=(root/'.api-key').read_text().strip();base='http://127.0.0.1:8999'
   try:urllib.request.urlopen(base+'/health',timeout=10)
   except urllib.error.HTTPError as e:assert e.code==401
   else:raise AssertionError('Missing authentication rejection')
   h=json.load(urllib.request.urlopen(urllib.request.Request(base+'/health',headers={'Authorization':'Bearer '+key}),timeout=30))
   assert h['ready']is True and h['native']['types']==['boolean','choice','score']
   assert h['native']['output_token_decoding']is False and h['native']['shared_prefix_compute']is False
   assert h['native']['checkpoint_sha256']=='8166e46dc6019ae13f0d8fc97d603cdbcc2d20eb1882c7350c8deb9fc6bae215'
   request=json.loads((root/'fixture.json').read_text());body=json.dumps(request).encode()
   r=urllib.request.Request(base+'/api/evaluate',data=body,headers={'Authorization':'Bearer '+key,'Content-Type':'application/json'})
   out=json.load(urllib.request.urlopen(r,timeout=900));(root/'result.json').write_text(json.dumps(out,indent=2)+'\n')
   from importlib.util import spec_from_file_location,module_from_spec
   s=spec_from_file_location('reader',root/'reader-decision-server.py');module=module_from_spec(s);s.loader.exec_module(module)
   module.validate_result(request,out)
   a=out['results'][0]['answers'];assert a[0]['value']is False and a[1]['value']=='six' and a[2]['level']==2
   assert out['usage']['generated_tokens']==0 and out['usage']['truncated_inputs']==0
   print(json.dumps(out,indent=2));print('Fixture structure and grounded preferred alternatives pass; no domain accuracy guarantee.')
   PYVERIFY

   Expected result: Complete native result is saved. All native probabilities are finite, normalized and consistent with boolean mass, choice argmax/margin and ordinal expectation; eight candidate paths and zero generated tokens are checked.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• Documentation and source checks only: no weights, native model imports, tokenizer execution, installation, build, inference, fit, latency or quality measurements were performed.
• Native v1 verifier expects 343 BF16 tensors while HF API safetensors summary says F32. Exact release size, SHA and header composition must all pass before strict whole-state loading; do not replace with the later checkpoint.
• The integrated model is a trained Qwen3 backbone plus candidate head, not a language decoder or adapter. CPU defaults and this explicit reader path use FP32; README BF16 examples do not establish CPU dtype.
• Native path limit is 2048 positions including complete state/question/candidate framing. Sums across eight fixture paths do not share this context bound. The service refuses overlength paths and reports truncated_inputs zero; there is no generated output budget.
• Native source accepts larger candidate counts; this reader boundary allows at most three questions, eight choice candidates or five ordered score levels, a 16 KiB body and bounded UTF-8 fields. Requests are serial and there is no inference deadline.
• The native calibration file is preserved and all three scalars checked. Boolean and relative choice probabilities and ordered expected scores are model distributions; arbitrary-domain calibration, action success and correctness are unknown. No external action executes.
• The original fictional fixture checks meaningful outputs for all three native primitives. A protocol pass or preferred answer is not a general model accuracy benchmark.
• Native source package metadata says version 1.0.0 while its module reports 0.1.0; exact source commit governs. Training upstream revision and Base versus non-Base model-card labeling remain unresolved provenance details.
• 32 GiB RAM and 40 GiB disk are planning allocations. Full load copies, CPU activations and scratch add to the FP32 parameter lower bound; actual required memory, fit, speed, quality and task spend are unknown.
• Larger batches, shared-prefix encoders, routing-only checkpoints, GPU execution and other native workflows remain open. Required family guide total is null; this pair is not exhaustive.

Native asset, header or strict state load fails

Stop. Restore the exact seven v1 files and inspect actual diagnostics; do not waive missing keys, switch to another checkpoint or conceal the dtype discrepancy.

Path too long or invalid native distribution

Treat the request as failed. Shorten the complete state/question/candidate paths deliberately; do not truncate or clamp probabilities, margins or ordinal scores.

HTTP 401 or loopback connection failure

Check the exact service key, foreground process and intended workstation loopback port.

Out of memory or very slow inference

Stop and preserve actual logs. Planning allocations and parameter storage do not prove fit or latency. Any larger resource choice needs its own reviewed configuration.

Stop the model

1. 1. Stop the foreground service

   Press Ctrl+C in the server shell, then confirm the intended listener is gone.

   CommandCopy command

   curl --max-time 3 --fail http://127.0.0.1:8999/health

   Expected result: Connection is refused after shutdown. A 401 indicates the service is still running; an auth rejection is not stopped proof.

Uninstall

1. 1. Remove the dedicated files

   Preserve desired result, environment record and notices first. Verify the exact dedicated path before removing its private key, assets, source and environment. Shared Ubuntu packages remain.

   CommandCopy command

   rm -r -- "$HOME/agentjev-v1-hosting"

   Expected result: Dedicated source, model, virtual environment, fixture, results and key removed.

Exact model and requirements

Artifact:

aimeigaoshou (native release), malevrigns (official source) / AgentJev v1 complete trained Qwen3 backbone and candidate decision head · checkpoint aimeigaoshou/agent-jev v1

Publisher revision:

7d433994fbde17a3f0993c2f2b02fe8ca1370db1

Artifact digest:

sha256:8166e46dc6019ae13f0d8fc97d603cdbcc2d20eb1882c7350c8deb9fc6bae215

Download size:

1.11 GiB

Memory requirement (unknown):

Unknown — No measured peak for both networks, CPU tensors, native decoding, dependencies and serialization.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; path microbatch one, serial HTTP request handling. Complete trained backbone and head on CPU; no separate language decoder, base weights, GPU or offload.

KV cache:

Complete independent causal paths use native backbone use_cache=False; no autoregressive output decoding or shared-prefix compute is requested.

Workload:

2,048 token context; Unknown; Not applicable; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Exact v1 local Qwen3 tokenizer, native path framing without added special tokens · 7d433994fbde17a3f0993c2f2b02fe8ca1370db1

Sampling / reasoning:

No autoregressive sampler. Exact pinned boolean, choice and score temperature scalars are loaded and checked against the complete native calibration file; these are relative head distributions, not arbitrary-domain calibration or action-success probabilities. Each candidate is a complete causal [STATE], [QUESTION], [CANDIDATE] path. The bare v1 strict safetensors route has no training config and rejects paths beyond 2048 rather than truncating. Native config positional capacity 32768 is distinct from this service bound. Boolean returns true mass and threshold; choice returns relative argmax, top probability and margin; score returns ordered expected level plus argmax. Transport identifiers are not semantic inputs; no tool or action executes.

Runtime prerequisites:

Ubuntu24.04 x86_64; See prerequisites above.

Official artifact · License

The native source contains Apache 2.0 and the exact v1 model card declares Apache 2.0; the snapshot lists no separate license file. Retain applicable copyright, license, notices and modification notices; Apache trademark and patent conditions apply. The identified Qwen3 upstream currently publishes Apache 2.0 too, but its training revision and Base versus non-Base labeling are not proven by these sources. Source access is not legal clearance or permission for unrelated data. No decision accuracy or fitness guarantee.

• temperatures.json · 0 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1
• config.json · 0 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1
• tokenizer.json · 0.01 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1
• tokenizer_config.json · 0 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1
• vocab.json · 0 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1
• merges.txt · 0 GiB · revision 7d433994fbde17a3f0993c2f2b02fe8ca1370db1

• Native model file is 1,196,881,242 bytes; all seven required release assets total 1,208,372,367 bytes. These are artifact sizes, not RAM requirements.
• Publisher verifier counts 598,418,689 parameter elements; their FP32 storage lower bound is 2,393,674,756 bytes before loading copies, activations, masks, tokenizer, libraries and OS. Peak and fit remain unknown.
• Four CPU threads, 32 GiB RAM and 40 GiB free disk are planning allocations, not measured minima.

Install the runtime

1. 1. Install the exact native CPU dependency path

   Use the official hashed CPU wheel and complete pinned source package without stripping dependencies. The imported native losses, data, model, encoders and service path are source-audited. Eager Qwen3 attention is selected through the released Transformers API; no CUDA driver or flash-attention package is required. Record the resolved environment; target execution remains unverified.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv git curl ca-certificates openssl
   mkdir -p "$HOME/agentjev-v1-hosting/model"
   cd "$HOME/agentjev-v1-hosting"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install 'https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91'
   git clone https://github.com/malevrigns/agent-jev.git runtime
   git -C runtime checkout --detach 3b0042a6171124c5d2942748ae01f453ed2cce61
   test "$(git -C runtime rev-parse HEAD)" = '3b0042a6171124c5d2942748ae01f453ed2cce61'
   .venv/bin/python -m pip install 'transformers==4.57.6' 'numpy==1.26.4' 'PyYAML==6.0.2' 'tqdm==4.67.1' 'safetensors==0.6.2' -e ./runtime
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Pinned source checkout, direct package versions and pip check succeed; resolved environment saved. Stop on any dependency failure.

Download the model

1. 1. Download all seven exact v1 assets and notices

   All hashes and sizes come from the native release verifier and pinned public metadata. Download no alternate branch. The wrapper checks every size and digest and the publisher BF16 header composition before importing the native loader. HF API F32 summary disagrees with the verifier; a mismatch is a stop, not permission to relabel the release.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   curl --fail --location --output model/model.safetensors 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/model.safetensors'
   curl --fail --location --output model/temperatures.json 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/temperatures.json'
   curl --fail --location --output model/config.json 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/config.json'
   curl --fail --location --output model/tokenizer.json 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/tokenizer.json'
   curl --fail --location --output model/tokenizer_config.json 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/tokenizer_config.json'
   curl --fail --location --output model/vocab.json 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/vocab.json'
   curl --fail --location --output model/merges.txt 'https://huggingface.co/aimeigaoshou/agent-jev/resolve/7d433994fbde17a3f0993c2f2b02fe8ca1370db1/merges.txt'
   sha256sum --check <<'SHA256'
   8166e46dc6019ae13f0d8fc97d603cdbcc2d20eb1882c7350c8deb9fc6bae215  model/model.safetensors
   5e5032896d77c72feb86b4725abe418488b8a1c7476014c706bd8215eb43c86c  model/temperatures.json
   504a6b58c4271583724e66584b6b7698aea18450209df6b2f7582df0e89cee59  model/config.json
   c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539  model/tokenizer.json
   3c04ed3ca964ea2f6b2b5faf0dc4d31aec1cb1e8b4bcf63f402d295046b422b5  model/tokenizer_config.json
   ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910  model/vocab.json
   8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5  model/merges.txt
   SHA256
   curl --fail --location --output AGENTJEV-LICENSE 'https://raw.githubusercontent.com/malevrigns/agent-jev/3b0042a6171124c5d2942748ae01f453ed2cce61/LICENSE'
   printf '%s  %s\n' 'f4d9b0aad8ff9e650f5890fa3eeb7c1311479cab901f8a9b416a2d9acaf54266' 'AGENTJEV-LICENSE' | sha256sum --check
   curl --fail --location --output QWEN-UPSTREAM-LICENSE 'https://huggingface.co/Qwen/Qwen3-0.6B-Base/resolve/da87bfb608c14b7cf20ba1ce41287e8de496c0cd/LICENSE'
   printf '%s  %s\n' '832dd9e00a68dd83b3c3fb9f5588dad7dcf337a0db50f7d9483f310cd292e92e' 'QWEN-UPSTREAM-LICENSE' | sha256sum --check

   Expected result: Every asset and both retained notices pass SHA checks. Native weights have not been fetched by the author.

Configure

1. 1. Write the manifest, original fixture and persistent key

   Create a key only when absent; an empty existing key must fail rather than silently rotate. Protect files and review the original synthetic state and criteria.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   cat > release-manifest.json <<'MANIFEST'
   {
     "repository": "aimeigaoshou/agent-jev",
     "revision": "7d433994fbde17a3f0993c2f2b02fe8ca1370db1",
     "files": [
       {
         "path": "model.safetensors",
         "bytes": 1196881242,
         "sha256": "8166e46dc6019ae13f0d8fc97d603cdbcc2d20eb1882c7350c8deb9fc6bae215"
       },
       {
         "path": "temperatures.json",
         "bytes": 389,
         "sha256": "5e5032896d77c72feb86b4725abe418488b8a1c7476014c706bd8215eb43c86c"
       },
       {
         "path": "config.json",
         "bytes": 727,
         "sha256": "504a6b58c4271583724e66584b6b7698aea18450209df6b2f7582df0e89cee59"
       },
       {
         "path": "tokenizer.json",
         "bytes": 7031645,
         "sha256": "c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539"
       },
       {
         "path": "tokenizer_config.json",
         "bytes": 9678,
         "sha256": "3c04ed3ca964ea2f6b2b5faf0dc4d31aec1cb1e8b4bcf63f402d295046b422b5"
       },
       {
         "path": "vocab.json",
         "bytes": 2776833,
         "sha256": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"
       },
       {
         "path": "merges.txt",
         "bytes": 1671853,
         "sha256": "8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5"
       }
     ]
   }
   MANIFEST
   cat > fixture.json <<'FIXTURE'
   {
     "id": "original-library-schedule",
     "state": "A fictional community library opens weekdays 9am–6pm, Saturday 10am–4pm, and is closed Sunday. These are original synthetic facts for a low-stakes demonstration.",
     "questions": [
       {
         "id": "sunday",
         "type": "boolean",
         "question": "Is the library open on Sunday according to this state?",
         "criteria": {
           "true": "The supplied state says it is open Sunday.",
           "false": "The supplied state says it is closed Sunday."
         }
       },
       {
         "id": "saturday-hours",
         "type": "choice",
         "question": "How many hours is the library open on Saturday?",
         "options": {
           "four": "Four hours.",
           "six": "Six hours.",
           "nine": "Nine hours."
         }
       },
       {
         "id": "weekday-duration",
         "type": "score",
         "question": "Rate the weekday opening duration by the ordered definitions.",
         "levels": [
           "At most four hours open per weekday.",
           "More than four but at most eight hours open per weekday.",
           "More than eight hours open per weekday."
         ]
       }
     ]
   }
   FIXTURE
   if [ ! -e .api-key ]; then (umask 077; openssl rand -hex 32 > .api-key); fi
   test -s .api-key
   chmod 600 .api-key

   Expected result: Manifest and fictional fixture saved; private nonempty key retained.

2. 2. Write request and result boundaries

   Reader-authored service code is independent of publisher HTTP serving code. It rejects unknown fields, duplicate JSON keys, unsupported content framing and malformed probability algebra. No publisher model code executes while writing this file.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   cat > reader-decision-server.py <<'PYWRAPPER'
   """Reader-authored serial private AgentJev decision service; no native imports at import time."""
   import hashlib,hmac,json,math,pathlib,re,socket,struct
   from http.server import HTTPServer,BaseHTTPRequestHandler
   PORT=8999
   EXPECTED_WEIGHT='8166e46dc6019ae13f0d8fc97d603cdbcc2d20eb1882c7350c8deb9fc6bae215'
   def digest(path):
    h=hashlib.sha256()
    with open(path,'rb')as f:
     for b in iter(lambda:f.read(4*1024*1024),b''):h.update(b)
    return h.hexdigest()
   def verify_files(root,manifest):
    for row in manifest:
     f=root/row['path']
     if not f.is_file()or f.stat().st_size!=row['bytes']or digest(f)!=row['sha256']:raise ValueError('Pinned release asset mismatch')
   def header_check(path):
    with open(path,'rb')as f:
     raw=f.read(8)
     if len(raw)!=8:raise ValueError('Missing safetensors header')
     n=struct.unpack('<Q',raw)[0]
     if not 1<=n<=1048576:raise ValueError('Header bound')
     h=json.loads(f.read(n))
    tensors=[v for k,v in h.items()if k!='__metadata__']
    if len(tensors)!=343 or any(v.get('dtype')!='BF16'for v in tensors):raise ValueError('Expected exact publisher v1 BF16 header; metadata discrepancy is not waived')
    total=0
    for v in tensors:
     shape=v.get('shape')
     if not isinstance(shape,list)or any(type(x)is not int or x<0 for x in shape):raise ValueError('Invalid shape')
     total+=math.prod(shape)
    if total!=598418689:raise ValueError('Full parameter composition mismatch')
   def parse_request(raw):
    def unique(pairs):
     d={}
     for k,v in pairs:
      if k in d:raise ValueError('Duplicate field')
      d[k]=v
     return d
    p=json.loads(raw.decode('utf-8'),object_pairs_hook=unique,parse_constant=lambda _:(_ for _ in()).throw(ValueError('Nonfinite JSON')))
    if not isinstance(p,dict)or set(p)!={'id','state','questions'}:raise ValueError('Exact state request schema')
    def text(v,limit):
     if not isinstance(v,str)or not v.strip()or len(v.encode())>limit or any(x in v for x in ['[STATE]','[QUESTION]','[CANDIDATE]']):raise ValueError('Bounded plain text required')
    text(p['id'],80);text(p['state'],4096)
    if not isinstance(p['questions'],list)or not 1<=len(p['questions'])<=3:raise ValueError('One to three questions required')
    ids=set()
    for q in p['questions']:
     if not isinstance(q,dict):raise ValueError('Question object required')
     t=q.get('type')
     if not isinstance(t,str):raise ValueError('Typed primitive string required')
     extra={'boolean':'criteria','choice':'options','score':'levels'}.get(t)
     if extra is None or set(q)!={'id','type','question',extra}:raise ValueError('Exact typed question schema')
     text(q['id'],80);text(q['question'],1024)
     if q['id']in ids:raise ValueError('Duplicate question id')
     ids.add(q['id']);v=q[extra]
     if t=='boolean':
      if not isinstance(v,dict)or set(v)!={'true','false'}:raise ValueError('Both boolean criteria required')
      vals=list(v.values())
     elif t=='choice':
      if not isinstance(v,dict)or not 2<=len(v)<=8:raise ValueError('Two to eight choice options')
      for k in v:text(k,80)
      vals=list(v.values())
     else:
      if not isinstance(v,list)or not 2<=len(v)<=5:raise ValueError('Two to five ordered score levels')
      vals=v
     for v in vals:text(v,1024)
     if len(set(vals))!=len(vals):raise ValueError('Distinct descriptions required')
    return p
   def request_length(headers):
    vals=headers.get_all('Content-Length',[])
    if headers.get('Transfer-Encoding')is not None or len(vals)!=1 or not re.fullmatch('[0-9]{1,5}',vals[0]):raise ValueError('One bounded decimal length')
    n=int(vals[0])
    if not 1<=n<=16384:raise ValueError('Body outside limit')
    return n
   def validate_result(request,result):
    if result.get('api_version')!='agentjev.decision.v1':raise ValueError('Native API version mismatch')
    usage=result.get('usage',{})
    if any(type(usage.get(k))is not int or usage[k]!=0 for k in ['generated_tokens','truncated_inputs']):raise ValueError('No decoding or truncation allowed')
    rows=result.get('results');questions=request['questions']
    if not isinstance(rows,list)or len(rows)!=1 or rows[0].get('id')!=request['id']:raise ValueError('Native state identity mismatch')
    answers=rows[0].get('answers')
    if not isinstance(answers,list)or len(answers)!=len(questions):raise ValueError('Missing typed answers')
    paths=0
    for q,a in zip(questions,answers):
     t=q['type'];keys=['true','false']if t=='boolean'else list(q['options'])if t=='choice'else[str(i)for i in range(len(q['levels']))];paths+=len(keys)
     dist=a.get('distribution')
     if a.get('id')!=q['id']or a.get('type')!=t or not isinstance(dist,dict)or list(dist)!=keys:raise ValueError('Typed answer identity mismatch')
     probs=list(dist.values())
     if any(type(x)not in(int,float)or not math.isfinite(x)or not 0<=x<=1 for x in probs)or abs(sum(probs)-1)>1e-4:raise ValueError('Invalid finite distribution')
     best=max(range(len(probs)),key=probs.__getitem__)
     def close(a,b):return type(a)in(int,float)and math.isfinite(a)and abs(a-b)<=1e-6
     if t=='boolean':
      if type(a.get('value'))is not bool or a['value']!=(probs[0]>=.5)or not close(a.get('probability'),probs[0]):raise ValueError('Boolean semantics mismatch')
     elif t=='choice':
      ranked=sorted(probs,reverse=True)
      if a.get('value')!=keys[best]or a.get('description')!=q['options'][keys[best]]or not close(a.get('top_probability'),probs[best])or not close(a.get('margin'),ranked[0]-ranked[1]):raise ValueError('Relative choice semantics mismatch')
     elif type(a.get('level'))is not int or a['level']!=best or a.get('legend')!=q['levels']or not close(a.get('score'),sum(i*x for i,x in enumerate(probs))):raise ValueError('Ordered expected-score semantics mismatch')
    if usage.get('questions')!=len(questions)or usage.get('candidate_paths')!=paths or type(usage.get('input_path_tokens'))is not int or not paths<=usage['input_path_tokens']<=paths*2048:raise ValueError('Native path usage mismatch')
    return result
   PYWRAPPER

   Expected result: First wrapper portion saved.

3. 3. Write strict native loading and serial loopback HTTP

   Startup verifies every native file and header before loading the complete trained state with strict=True. It checks CPU FP32, all three calibration scalars and native identity/path mode. It has a 30 second socket-read timeout, no native inference deadline and no public-service hardening.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   cat >> reader-decision-server.py <<'PYWRAPPER'
   class Engine:
    def __init__(self,root):
     import os
     manifest=json.loads((root/'release-manifest.json').read_text())['files'];verify_files(root/'model',manifest);header_check(root/'model/model.safetensors')
     os.environ.update(HF_HUB_OFFLINE='1',TRANSFORMERS_OFFLINE='1')
     import torch
     from jev_service.engine import DecisionEngine
     from jev_service.contract import prepare,encode_paths
     torch.set_default_dtype(torch.float32);torch.set_num_threads(4);torch.set_num_interop_threads(1)
     self.engine=DecisionEngine(root/'model/model.safetensors',str(root/'model'),device='cpu',max_tokens=2048,path_batch=1,encoder='path',temperatures=root/'model/temperatures.json')
     self.engine.model.path_encoder.backbone.set_attn_implementation('eager')
     if self.engine.model.path_encoder.backbone.config._attn_implementation!='eager':raise ValueError('Explicit eager attention required')
     if any(p.is_floating_point()and p.dtype!=torch.float32 for p in self.engine.model.parameters()):raise ValueError('CPU FP32 parameters required')
     info=self.engine.info();expected=json.loads((root/'model/temperatures.json').read_text());temps={k:v['temperature']for k,v in expected.items()}
     if set(temps)!={'boolean','choice','score'}or temps!={'boolean':1.0717734625362931,'choice':1.0352649238413776,'score':1.0717734625362931}:raise ValueError('Complete exact calibration required')
     if info['api_version']!='agentjev.decision.v1'or info['checkpoint_sha256']!=EXPECTED_WEIGHT or info['encoder']!='path'or info['max_path_tokens']!=2048 or info['temperatures']!=temps or info['types']!=['boolean','choice','score']or info['output_token_decoding']is not False or info['shared_prefix_compute']is not False:raise ValueError('Native release/calibration identity mismatch')
     self.prepare=prepare;self.encode_paths=encode_paths;self.info=info
    def evaluate(self,request):
     paths,_,_=self.encode_paths(self.prepare(request),self.engine.tokenizer,2048)
     if not paths or any(not p or len(p)>2048 for p in paths):raise ValueError('Complete per-path bound')
     return validate_result(request,self.engine.evaluate(request))
   def serve(engine,key):
    class Handler(BaseHTTPRequestHandler):
     protocol_version='HTTP/1.0'
     def setup(self):super().setup();self.connection.settimeout(30)
     def log_message(self,*a):pass
     def reply(self,status,obj):
      b=json.dumps(obj,ensure_ascii=False,allow_nan=False).encode();self.send_response(status);self.send_header('Content-Type','application/json');self.send_header('Content-Length',str(len(b)));self.end_headers();self.wfile.write(b)
     def auth(self):
      v=self.headers.get_all('Authorization',[])
      try:return len(v)==1 and hmac.compare_digest(v[0].encode('ascii'),('Bearer '+key).encode('ascii'))
      except UnicodeError:return False
     def do_GET(self):
      if not self.auth():return self.reply(401,{'error':'Unauthorized'})
      if self.path!='/health':return self.reply(404,{'error':'Not found'})
      return self.reply(200,{'ready':True,'native':engine.info,'reader_path_microbatch':1,'reader_context_per_path':2048})
     def do_POST(self):
      if not self.auth():return self.reply(401,{'error':'Unauthorized'})
      if self.path!='/api/evaluate':return self.reply(404,{'error':'Not found'})
      try:
       if self.headers.get('Content-Type')!='application/json':raise ValueError('JSON required')
       n=request_length(self.headers);b=self.rfile.read(n)
       if len(b)!=n:raise ValueError('Incomplete body')
       req=parse_request(b)
      except(ValueError,UnicodeError,socket.timeout):return self.reply(400,{'error':'Invalid bounded request'})
      try:out=engine.evaluate(req)
      except Exception:return self.reply(422,{'error':'Native path or finite-distribution check failed; no action executed'})
      return self.reply(200,out)
    server=HTTPServer(('127.0.0.1',PORT),Handler)
    try:server.serve_forever()
    finally:server.server_close()
   if __name__=='__main__':
    root=pathlib.Path(__file__).resolve().parent;key=(root/'.api-key').read_text().strip()
    if not re.fullmatch('[a-f0-9]{64}',key):raise ValueError('Restore private service key')
    serve(Engine(root),key)
   PYWRAPPER
   printf '%s  %s\n' 'a997d981592ccc21575be799483bf41f65ce43990175e579936cc643d5d59cce' 'reader-decision-server.py' | sha256sum --check
   .venv/bin/python -m py_compile reader-decision-server.py

   Expected result: Exact wrapper SHA and Python syntax pass; no inference has run.

Start the model

1. 1. Start the private CPU decision service

   Run in a foreground shell. Four intra-op threads and one inter-op thread are explicit CPU settings. Complete independent candidate paths use native path microbatch one and no shared-prefix compute. Health means initialized and checked, not successful model inference.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   export OMP_NUM_THREADS=4 MKL_NUM_THREADS=4 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1
   .venv/bin/python reader-decision-server.py

   Expected result: Exact native checkpoint/calibration identity checks pass and port 8999 binds only to loopback. Any size, hash, header, state-key, dtype or attention failure aborts startup.

Verify the setup

1. 1. Verify authentication, native algebra and grounded fixture

   Run in a second shell on the same host. Require unauthorized health to return 401, then check native identity and all three primitives. The checker independently revalidates complete distributions and manual fixture answers: Sunday closed, Saturday six hours, weekday duration above eight hours. Score is the probability-weighted ordinal expectation, not forced to its argmax. Wrong preferred alternatives fail; passing one fixture is not accuracy evidence.

   CommandCopy command

   set -e
   cd "$HOME/agentjev-v1-hosting"
   .venv/bin/python - <<'PYVERIFY'
   import json,urllib.request,urllib.error,pathlib
   root=pathlib.Path.cwd();key=(root/'.api-key').read_text().strip();base='http://127.0.0.1:8999'
   try:urllib.request.urlopen(base+'/health',timeout=10)
   except urllib.error.HTTPError as e:assert e.code==401
   else:raise AssertionError('Missing authentication rejection')
   h=json.load(urllib.request.urlopen(urllib.request.Request(base+'/health',headers={'Authorization':'Bearer '+key}),timeout=30))
   assert h['ready']is True and h['native']['types']==['boolean','choice','score']
   assert h['native']['output_token_decoding']is False and h['native']['shared_prefix_compute']is False
   assert h['native']['checkpoint_sha256']=='8166e46dc6019ae13f0d8fc97d603cdbcc2d20eb1882c7350c8deb9fc6bae215'
   request=json.loads((root/'fixture.json').read_text());body=json.dumps(request).encode()
   r=urllib.request.Request(base+'/api/evaluate',data=body,headers={'Authorization':'Bearer '+key,'Content-Type':'application/json'})
   out=json.load(urllib.request.urlopen(r,timeout=900));(root/'result.json').write_text(json.dumps(out,indent=2)+'\n')
   from importlib.util import spec_from_file_location,module_from_spec
   s=spec_from_file_location('reader',root/'reader-decision-server.py');module=module_from_spec(s);s.loader.exec_module(module)
   module.validate_result(request,out)
   a=out['results'][0]['answers'];assert a[0]['value']is False and a[1]['value']=='six' and a[2]['level']==2
   assert out['usage']['generated_tokens']==0 and out['usage']['truncated_inputs']==0
   print(json.dumps(out,indent=2));print('Fixture structure and grounded preferred alternatives pass; no domain accuracy guarantee.')
   PYVERIFY

   Expected result: Complete native result is saved. All native probabilities are finite, normalized and consistent with boolean mass, choice argmax/margin and ordinal expectation; eight candidate paths and zero generated tokens are checked.

2. 2. Open a trusted SSH tunnel

   From the operator machine, replace HOST with the existing instance address and use the existing SSH identity. Keep inference port 8999 closed in the security group. Choose a free local port if needed and adjust the client address accordingly.

   CommandCopy command

   ssh -N -L 8999:127.0.0.1:8999 ubuntu@HOST

   Expected result: Tunnel forwards the operator loopback port to VM loopback. Keep this shell open only while using the private endpoint.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• Documentation and source checks only: no weights, native model imports, tokenizer execution, installation, build, inference, fit, latency or quality measurements were performed.
• Native v1 verifier expects 343 BF16 tensors while HF API safetensors summary says F32. Exact release size, SHA and header composition must all pass before strict whole-state loading; do not replace with the later checkpoint.
• The integrated model is a trained Qwen3 backbone plus candidate head, not a language decoder or adapter. CPU defaults and this explicit reader path use FP32; README BF16 examples do not establish CPU dtype.
• Native path limit is 2048 positions including complete state/question/candidate framing. Sums across eight fixture paths do not share this context bound. The service refuses overlength paths and reports truncated_inputs zero; there is no generated output budget.
• Native source accepts larger candidate counts; this reader boundary allows at most three questions, eight choice candidates or five ordered score levels, a 16 KiB body and bounded UTF-8 fields. Requests are serial and there is no inference deadline.
• The native calibration file is preserved and all three scalars checked. Boolean and relative choice probabilities and ordered expected scores are model distributions; arbitrary-domain calibration, action success and correctness are unknown. No external action executes.
• The original fictional fixture checks meaningful outputs for all three native primitives. A protocol pass or preferred answer is not a general model accuracy benchmark.
• Native source package metadata says version 1.0.0 while its module reports 0.1.0; exact source commit governs. Training upstream revision and Base versus non-Base model-card labeling remain unresolved provenance details.
• 32 GiB RAM and 40 GiB disk are planning allocations. Full load copies, CPU activations and scratch add to the FP32 parameter lower bound; actual required memory, fit, speed, quality and task spend are unknown.
• Larger batches, shared-prefix encoders, routing-only checkpoints, GPU execution and other native workflows remain open. Required family guide total is null; this pair is not exhaustive.

Native asset, header or strict state load fails

Stop. Restore the exact seven v1 files and inspect actual diagnostics; do not waive missing keys, switch to another checkpoint or conceal the dtype discrepancy.

Path too long or invalid native distribution

Treat the request as failed. Shorten the complete state/question/candidate paths deliberately; do not truncate or clamp probabilities, margins or ordinal scores.

HTTP 401 or loopback connection failure

Check the exact service key, VM readiness and intended SSH tunnel; keep public port closed.

Out of memory or very slow inference

Stop and preserve actual logs. Planning allocations and parameter storage do not prove fit or latency. Any larger resource choice needs its own reviewed configuration.

Stop the model

1. 1. Stop the foreground service

   Press Ctrl+C in the server shell, then confirm the intended listener is gone. Close the SSH tunnel too. This does not stop instance billing.

   CommandCopy command

   curl --max-time 3 --fail http://127.0.0.1:8999/health

   Expected result: Connection is refused after shutdown. A 401 indicates the service is still running; an auth rejection is not stopped proof.

Uninstall

1. 1. Remove the dedicated files

   Preserve desired result, environment record and notices first. Verify the exact dedicated path before removing its private key, assets, source and environment. Shared Ubuntu packages remain.

   CommandCopy command

   rm -r -- "$HOME/agentjev-v1-hosting"

   Expected result: Dedicated source, model, virtual environment, fixture, results and key removed.

Shut down cloud resources

1. 1. End cloud billing resources

   In EC2 us-east-1, stop for later reuse or terminate this exact VM. Review Delete on termination for its root volume. Delete unwanted detached EBS volumes, snapshots and reserved/Elastic/public IP resources. Stopped VM storage and retained resources may still bill. Check Bills/Cost Explorer later for final usage.

   Expected result: Exact instance is stopped/terminated as intended; no unwanted attached/retained billed resources remain. Instance compute stops only after its lifecycle transition, not when the Python service stops.