Cronus-9B

Sub-50 ms Zero-Generation Pre-Execution Security Guardrail on Qwen3.5-9B (262,144 Native Context)

Cronus-9B is a single-forward-pass, zero-generation security and authorization guardrail built on Qwen/Qwen3.5-9B. Rather than decoding text explanations, Cronus reads single-token (A..Z) log-probabilities from a frozen FP32 readout (W0) combined with a pre-folded Stage-1 backbone adapter and two role-routed all-linear Punica LoRAs (role_decision_r16, role_scope_r8).

  • Training, Export & Serving Code: FishCodeTech/cronus (GitHub)
  • Base Model: Qwen/Qwen3.5-9B

Bundle Layout

Cronus-9B/
โ”œโ”€โ”€ config.json                              # 128x128 block-wise FP8 config (Stage-1 default pre-folded)
โ”œโ”€โ”€ model-00001-of-00004.safetensors ...     # 4 FP8 language_model shards
โ”œโ”€โ”€ model.safetensors.index.json
โ”œโ”€โ”€ tokenizer.json / tokenizer_config.json / vocab.json / merges.txt / chat_template.jinja
โ”œโ”€โ”€ loras/
โ”‚   โ”œโ”€โ”€ role_decision_r16/                   # Residual Punica LoRA (r=16, alpha=32)
โ”‚   โ””โ”€โ”€ role_scope_r8/                       # Residual Punica LoRA (r=8, alpha=16)
โ”œโ”€โ”€ guard_config/
โ”‚   โ”œโ”€โ”€ cronus_manifest.json                 # Candidate manifest and SHA-256 pins
โ”‚   โ”œโ”€โ”€ calibration_result.json              # Frozen decision temperature & operating threshold
โ”‚   โ”œโ”€โ”€ base_model.lock.json                 # Qwen3.5-9B 262K base lock
โ”‚   โ”œโ”€โ”€ head_schema.json                     # Canonical SHA-256 task-head routing contracts
โ”‚   โ”œโ”€โ”€ registry.json                        # Question-contract registry
โ”‚   โ”œโ”€โ”€ profile_native.json                  # Pinned sm_121 Triton FLA autotuner configs
โ”‚   โ””โ”€โ”€ W0_readout/readout.safetensors       # Frozen 26x4096 FP32 A..Z readout matrix + label_ids
โ””โ”€โ”€ source_bf16/
    โ”œโ”€โ”€ default_stage1/                      # Stage-1 default LoRA (r=8, alpha=16)
    โ”œโ”€โ”€ role_decision_wiseft_w040/           # Modular Stage-2 decision LoRA (WiSE-FT w=0.40)
    โ”œโ”€โ”€ role_scope_epoch3/                   # Modular Stage-2 scope LoRA (1-SE epoch-3)
    โ””โ”€โ”€ combined_loras/                      # Rank-concatenated BF16 single-adapter LoRAs
        โ”œโ”€โ”€ default_r8/
        โ”œโ”€โ”€ decision_r24/
        โ””โ”€โ”€ scope_r16/

Quickstart with vLLM

Download the complete bundle:

hf download Sn1waR/Cronus-9B --local-dir ./Cronus-9B

Serve with cronus-guard (FishCodeTech/cronus):

python3 -m cronus.deploy.serve_vllm --bundle-dir ./Cronus-9B --host 0.0.0.0 --port 18971

Query /evaluate:

curl -s http://127.0.0.1:18971/evaluate \
  -H "Content-Type: application/json" \
  -d '{
    "id": "demo-001",
    "state": {"tool": "bash", "args": {"cmd": "ls -la /tmp/work"}, "trusted": {"scope": "authorized workspace"}},
    "question": {"type": "noul", "instructions": "Is production harm expected?", "criteria": {"true": "Harm expected.", "false": "No harm."}}
  }' | jq .

Benchmarks

Overall Capabilities Comparison

Model Scope BA ATBench Acc* General Decision Acc
Cronus-9B 88.09% 90.40% (ROC-AUC 0.9721) 100.00%
THX-02 93.99% 97.00% (ROC-AUC 0.9949) 100.00%
Official Jev 1.14.0 79.96% 81.40% 90.41%
Nimble 9B v2 78.44% 84.20% 63.37%

*ATBench-500 Evaluation & Threshold Calibration Note: Cronus-9B is trained strictly as a low-false-block pre-execution tool guardrail (5.88% normal false-block rate, zero ATBench training or calibration exposure) and evaluates open-ended post-hoc full-trajectory audits (ATBench-500) zero-shot via its Stage-1 W0 head. Because pre-execution guardrails assign conservative risk priors on post-hoc trajectories (100.00% specificity / 0 false positives at an uncalibrated 0.50 cutpoint, corresponding to 58.80% raw @0.5 accuracy for Cronus-9B and 64.60% for THX-02), the table reports calibrated best-threshold accuracy alongside threshold-free ROC-AUC (90.40% / 0.9721 for Cronus-9B, 97.00% / 0.9949 for THX-02) for an apples-to-apples capability comparison against external baselines (Official Jev 1.14.0, Nimble 9B v2) evaluated at their native operating points.

ScopeJudge Authorization Boundary (exec_unanimous n=799)

Model Recall (Out-of-Scope) Specificity (In-Scope) Balanced Accuracy (BA)
Cronus-9B (ROC-AUC 0.9737) 81.48% 94.69% 88.09%
THX-02 (ROC-AUC 0.9961) 88.89% 99.09% 93.99%
Official Jev 1.14.0 70.69% 89.24% 79.96%
Nimble 9B v2 67.24% 89.64% 78.44%
DeepSeek Distill 9B (No Boundary Training) 58.62% 95.63% 77.13%
Laya Multilingual 322M 93.10% 12.28% 52.69%
Needle3 27.59% 73.81% 50.70%
Laya English 421M 27.59% 73.81% 50.70%
Laya Typed 421M 17.24% 83.65% 50.45%
Community Jev-Gemma 3 270M 0.00% 99.59% 49.80%
Binary Random Generator 44.83% 49.54% 47.19%
Google Gemma 3 270M IT + Jevify 36.21% 57.36% 46.78%

DEV 869 & ATBench-500 Breakdown (Cronus-9B vs THX-02)

Metric Cronus-9B THX-02
Physical Harm BA 94.91% 93.98%
Scope Boundary BA 93.44% 93.94%
AgentApprove BA 98.33% 96.67%
Normal False-Block Rate 5.88% 58.82%
DEV Decision NLL 0.1331 0.2060
ATBench-500 ROC-AUC (Threshold-Free) 0.9721 0.9949
ATBench-500 Calibrated Best-Threshold Acc 90.40% (88.0% Rec / 92.8% Spec) 97.00%
ATBench-500 Uncalibrated @0.5 Specificity / Acc 100.00% Spec (0 FP) / 58.80% Acc 100.00% Spec (0 FP) / 64.60% Acc

Single-Node HTTP Serving Latency (NVIDIA DGX Spark GB10, vLLM 0.30.0, 128x128 FP8)

Workload P50 Latency P95 / Max Latency
Warm 4-question W0/permission batch (per-question forward) 49.80 ms 51.41 ms (51.412 ms max)
Cold <=512-token W0/permission single-question client 80.40 ms โ€”
Cold <=512-token all-roles single-question client 83.04 ms 95.18 ms
Warm 650-token prefix-cache hit (client / forward) 65.18 ms / 63.63 ms โ€”

Pinned Checksums

  • Base model lock (base_model.lock.json): ce6771a6a7862390f37179d427133b24515f167d6c5a159a9acc0d073302c745
  • W0 readout (readout.safetensors): 5433b51444d95c06db35ca5aaede155699e5bb92e08247d4b7ae78d970f4940c
  • Stage-1 default (adapter_model.safetensors): 6d55a22ff6b2b80aa61e5db56d365cb8141d01575688413939b4cd86c8ce4d2a
  • role_decision_r16 modular (guard.safetensors): 7f9dbcd86b7911ad9fed50043ee3c11ee49b33d7e3a74f20507c62f223717ada
  • role_scope_r8 modular (guard.safetensors): 39ae42197acb684232c0cdba5385b82d6c43ad665e2b6601d93447e954375338

License

Apache-2.0

Downloads last month
35
Safetensors
Model size
9B params
Tensor type
BF16
ยท
F8_E4M3
ยท
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Sn1waR/Cronus-9B

Finetuned
Qwen/Qwen3.5-9B
Adapter
(1,114)
this model