N
NomosGuard
Deterministic security reasoning core
v0.2.3 · Oct 2026

LLM benchmark — security telemetry parsing

Every model that touches NomosGuard's evidence edge reads raw logs and writes structured claims. This benchmark measures how well each model performs that one job: 16 models, 4 hand-derived scenarios, 76 measured runs — same prompt, script-scored, nothing by hand.

ModelRecall PrecisionSchema Halluc.
1 gpt-oss-120bAntigravity 100% 100% 100% 0
2 gemini-3.8-flashAntigravity 100% 96% 100% 1
3 claude-opus-4.6Antigravity 100% 88% 100% 2
4 opencode/step-5-preview-free 100% 88% 100% 2
5 gemini-3.1-proAntigravity 95% 95% 100% 1
6 opencode/mimo-v2.6-flash-free 95% 95% 100% 1
7 opencode/nemotron-3-ultra-freeOpenCode Zen 95% 95% 100% 1
8 opencode/space-bunny-freeOpenCode Zen 95% 87% 75% 1
9 claude-sonnet-4.6Antigravity 95% 83% 100% 2
10 opencode/longcat-2.5-preview-freeOpenCode Zen 95% 83% 100% 2
11 opencode/ling-3.1-flash-free 93% 82% 100% 2
12 opencode/big-pickleOpenCode Zen 88% 88% 100% 1
13 mistral-medium-3.5Vibe CLI 87% 78% 100% 2
14 opencode/muse-spark-1.3-contributor-freeOpenCode Zen 55% 62% 100% 1
15 gemini-3.7-flashAntigravity 25% 25% 25% 0
16 gemini-3.6-flashAntigravity 0% 0% 0% 0

Recall by scenario

Brute force & pivot · benign admin window · lateral movement · privilege escalation

ModelBrute Force Then PivotBenign Admin WindowLateral Movement PairPrivilege Escalation
gpt-oss-120b100%100%100%100%
gemini-3.8-flash100%100%100%100%
claude-opus-4.6100%100%100%100%
opencode/step-5-preview-free100%100%100%100%
gemini-3.1-pro80%100%100%100%
opencode/mimo-v2.6-flash-free80%100%100%100%
opencode/nemotron-3-ultra-free80%100%100%100%
opencode/space-bunny-free80%100%100%100%
claude-sonnet-4.680%100%100%100%
opencode/longcat-2.5-preview-free80%100%100%100%
opencode/ling-3.1-flash-free80%100%100%—
opencode/big-pickle100%50%100%100%
mistral-medium-3.580%100%100%67%
opencode/muse-spark-1.3-contributor-free20%0%100%100%
gemini-3.7-flash0%0%100%0%
gemini-3.6-flash0%0%0%0%