veto-open-jev-v4-qwen3.5-0.8b)Compact 516.8 MiB (Q4_K_M) System-1 & Micro-Blink Triage Model (32.1 ms p50 gateway engine / 90.1 ms WAN Customer mean latency on Cloud Run NVIDIA L4) adapted from Qwen/Qwen3.5-0.8B via Synthetic Rank-16 LoRA training + Post-LoRA Zero-Gradient Roof Calibration + Jump-Forward CUDA Serving.
Qwen/Qwen3.5-0.8B (873,438,784 parameters, Apache-2.0)main / v4): veto-open-jev-v4-qwen3.5-0.8b-Q4_K_M.gguf (516.80 MiB, SHA-256: 67e14a5d1c5307c843d4cdd1e4daa11ba5f9f394f940b801a15fcf96c7daa307)4B Model: AIHeikki/brainsurged-open-jev-qwen3.5-4b (36/36 Customer, 22/24 Business, 2/3 Rescue on frozen Jevlab)v4):1,320 synthetic counterfactual state/objective examples (2 epochs, 166 AdamW steps, 9.34M adapter scalars = 1.083%, 5.4 min 0.8B LoRA phase within a 33.2 min Round-4 campaign on a Spot NVIDIA A100-40GB). Overlap check: no exact string matches were found against the 214 holdout strings checked; this is not a semantic-contamination or pretraining audit.torch.set_grad_enabled(False)): Inverse temperature scaling (beta_temp = 0.85) + <think> loop verification (0.0% unclosed loops on calibration checks).main and v1–v5 branches):v5 branch & standalone repo AIHeikki/brainsurged-chessmaster-qwen3.5-0.8b): Surgical Mid+Upper Cortex (Layers 8..21) Rank-16 LoRA + Roof Calibration + serving_v6 Exact-Leaf Branch-and-Bound decoder (75.83% on Holdout-Chessmaster-120, and ½–1½ against Jev 1.13.0 in the live Cloud Run L4 rematch at 54.8 ms p50 engine — illustrating how single-position benchmark gains and full-game play diverge at 0.8B scale).main / v4: Round-4 flagship Q4_K_M GGUF + v1–v4 .safetensors surgical/LoRA delta patches (patches_and_attestations/) + Round-5 C++/CUDA Jump-Forward serving sources & build guide (serving_v5/) + internal round reports (reports/) and Jevlab/Microgrid evaluation records (reports/jevlab_evaluations/).v1, v2, v3, v5: Historical and domain-specialized weight artifacts, patches, and attestations archived for inspection (specify the generation's exact .gguf filename when downloading; see layout table below).llama.cpp / serving_v5)Timing & Runtime Note: Brainsurged results labeled
Round-5used the customserving_v5/stack on Cloud RunNVIDIA L4. Historical rows labeledold stackused the earlierllama-serversetup.Jevused its own hosted API. These are not Ollama or local-device latency benchmarks.
v4 Default)ollama run hf.co/AIHeikki/brainsurged-open-jev-qwen3.5-0.8b
llama.cpp (llama-server) + Request-Side Schema Controlsllama-server \
--model veto-open-jev-v4-qwen3.5-0.8b-Q4_K_M.gguf \
--gpu-layers 99 \
--flash-attn on \
--jinja \
--reasoning-budget 0
Send a Per-Request Constrained-JSON Call (enable_thinking: false + response_format: json_schema):
(Stock llama-server usage example — separate from the custom serving_v5/ Cloud Run L4 timings below.)
curl -s http://127.0.0.1:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"temperature": 0,
"max_tokens": 256,
"chat_template_kwargs": {"enable_thinking": false},
"response_format": {
"type": "json_schema",
"json_schema": {
"name": "customer_triage",
"strict": true,
"schema": {
"type": "object",
"properties": {
"buying_intent": {"type": "boolean"},
"urgency": {"type": "boolean"},
"team": {"type": "string", "enum": ["sales", "support", "billing", "other"]}
},
"required": ["buying_intent", "urgency", "team"],
"additionalProperties": false
}
}
},
"messages": [
{"role": "system", "content": "Return only the requested JSON."},
{"role": "user", "content": "Customer reports duplicate invoice #4821 and threatens immediate cancellation unless refunded today."}
]
}'
serving_v5/)See serving_v5/README.md for the complete staging and Docker build guide (libjev_jump.so + fast_gateway.py + models.json). serving_v5/Dockerfile embeds both the 4B and 0.8B v4 GGUFs inside serving_v5/models/:
cd serving_v5
python3 stage_models_from_hf.py
docker build --platform linux/amd64 -t jevlab-serving-v5 .
NVIDIA L4, 516.8 MiB Q4_K_M)Source records: LIVE_RESULTS_ROUND5_2026-11-07.md, RESCUE_TRACES_ROUND5_2026-11-07.md, ROUND5_COMPARISON_2026-11-07.json.
| Model / Stack | Customer (36) |
Business (24: Rev / Trust / Spend) |
Single-Step Subtotal (60) |
60-Step Rescue (3) |
Customer Mean E2E (p50 Gateway Engine) |
Business / Rescue Mean E2E (p50 Engine) |
|---|---|---|---|---|---|---|
Jev 1.13.0 (Native hosted API, paired w/ R5) |
36/36 |
24/24 (8 / 8 / 8) |
60/60 (100.0%) |
2/3 |
180.5 ms (not exposed) |
191.0 ms / 184.1 ms |
0.8B v1 (0.0127% 0-Grad Roof Surgery) |
23/36 |
9/24 (2 / 4 / 3) |
32/60 (53.3%) |
0/3 |
299.9 ms (old stack) |
— |
0.8B v3 (0.0222% 0-Grad 3-Phase Roof) |
21/36 |
8/24 (2 / 3 / 3) |
29/60 (48.3%) |
0/3 |
279.1 ms (old stack) |
— |
0.8B v4 (1.08% LoRA + Roof, llama-server) |
34/36 |
19/24 (8 / 7 / 4) |
53/60 (88.3%) |
1/3 |
242.2 ms (old stack) |
245.6 ms / 223.4 ms |
0.8B v4 + Round-5 (libjev_jump.so Serving) |
34/36 |
19/24 (8 / 7 / 4) |
53/60 (88.3%) |
1/3 |
90.1 ms (32.1 ms) |
129.0 ms (50.0 ms) / 116.1 ms (45.7 ms) |
Scope & Limitations (LIVE_RESULTS_ROUND5_2026-11-07.md):
516.8 MiB, 0.8B v4 + Round-5 matched 39/39 (100.0%) saved v4 replay decisions while averaging 90.1 ms end-to-end on Customer (32.1 ms p50 gateway engine, vs Jev's 180.5 ms end-to-end) and scoring 34/36 (94.4%) on Customer labels and 15/16 (93.8%) on Revenue + Trust choices.0.8B scored 4/8 on Spend and exited immediately on both rescue missions requiring evacuations (1/3).0.8B was not evaluated in the new Microgrid campaign. Earlier Jevlab findings informed later development rounds.v0 → v1 → v3 → v4)| Benchmark & Metric | Untuned v0 |
v1 / v2 (0.0127%) |
v3 (0.0222% Roof) |
v4 (1.08% LoRA + Roof) |
|---|---|---|---|---|
Holdout-240 Combined (Zero-Think JSON) |
59.58% |
60.00% |
60.41% |
90.41% (+30.00 pp) |
↳ Minimum-Spend (Zero-Think JSON) |
60.42% |
62.50% |
62.50% |
100.00% (+37.50 pp) |
↳ Sequential Agentic 120 (Zero-Think JSON) |
64.17% |
65.00% |
65.00% |
93.33% (+28.33 pp) |
Holdout-240 (Micro-Blink, <think> on) |
Loops | 60.00% |
61.66% |
87.09% (+25.43 pp) |
Generative Jev-180 (<think> on) |
0.00% (100% loop) |
33.33% (30% loop) |
62.22% (0% loop) |
76.11% (+42.78 pp vs v1) |
Enterprise Jev-180 System-1 (360 probes) |
66.81% |
67.92% |
67.92% |
76.11% |
Live llama-server Dual-Mode Probes |
— | — | 2/3 |
4/4 Zero-Think & 4/4 Blink |
| Branch / Path | Artifact | Size | SHA-256 |
|---|---|---|---|
main / v4 (root) |
veto-open-jev-v4-qwen3.5-0.8b-Q4_K_M.gguf |
516.80 MiB |
67e14a5d1c5307c843d4cdd1e4daa11ba5f9f394f940b801a15fcf96c7daa307 |
v3 branch (root) |
veto-open-jev-v3-qwen3.5-0.8b-Q4_K_M.gguf |
516.80 MiB |
fdf63cc5aa7b9ffe3317034e9c98f237888738eae4a2c85029adc552a89c3354 |
v2 branch (root) |
veto-open-jev-v2-qwen3.5-0.8b-Q4_K_M.gguf |
516.80 MiB |
d3170f33df63f48b8f371a490cbf1cb77fe02ac565566fb26261220e49a491ba (identical bytes to v1) |
v1 branch (root) |
veto-open-jev-qwen3.5-0.8b-Q4_K_M.gguf |
516.80 MiB |
d3170f33df63f48b8f371a490cbf1cb77fe02ac565566fb26261220e49a491ba |
main (patches_and_attestations/) |
v1–v4 .safetensors deltas & JSON attestations |
~32.8 MiB |
Archived bfloat16 surgical & LoRA deltas |
main (serving_v5/) |
jev_jump_engine.cpp, fast_gateway.py, models.json, Dockerfile |
~345 KiB |
Jump-Forward CUDA serving sources & staging guide |
main (reports/jevlab_evaluations/) |
LIVE_RESULTS_*.md, traces, microgrid_manifest.json, microgrid_simulation.py |
~295 KiB |
Archived Jevlab & Microgrid evaluation records |
Specify the exact .gguf filename when downloading a historical generation (v1–v3 branches were branched from main and also retain inherited v4 files):
hf download AIHeikki/brainsurged-open-jev-qwen3.5-0.8b veto-open-jev-qwen3.5-0.8b-Q4_K_M.gguf --revision v1
hf download AIHeikki/brainsurged-open-jev-qwen3.5-0.8b veto-open-jev-v2-qwen3.5-0.8b-Q4_K_M.gguf --revision v2
hf download AIHeikki/brainsurged-open-jev-qwen3.5-0.8b veto-open-jev-v3-qwen3.5-0.8b-Q4_K_M.gguf --revision v3
Released under Apache-2.0, inheriting the license of Qwen/Qwen3.5-0.8B by the Qwen Team (Alibaba Cloud).
patches_and_attestations/v4/veto-open-jev-v4-qwen3.5-0.8b-lora-and-roof.safetensors 33 MBpatches_and_attestations/v3/veto-open-jev-v3-qwen3.5-0.8b-wsp-delta.safetensors 429.6 kBpatches_and_attestations/v2/veto-open-jev-v2-qwen3.5-0.8b-wsp-delta.safetensors 263.2 kBpatches_and_attestations/v1/veto-open-jev-qwen3.5-0.8b-wsp-delta.safetensors 263.2 kB