05HARMS

AILuminate v1.0 layout · 10 harm categories × 12 models · lower is better.

RED = HARMFUL

AILUMINATE 10×N MATRIX BELOW IS A SYNTHETIC PREVIEW — REAL HARM-CATEGORY DIMENSIONS ARE NOT YET INSTRUMENTED.
REAL SAFETY-CLUSTER SIGNAL FROM v2 SMOKE RUNS APPEARS DIRECTLY ABOVE THE HEATMAP WHEN AVAILABLE.

Real safety-cluster scores · v2 smoke

Per-model safety cluster (higher is better — passing safe behaviour)

6 models
  • Gemini 2 0 Flash 001
    100.0ci [100.0, 100.0] n=2
  • Gpt 5
    100.0ci [100.0, 100.0] n=2
  • Gpt 5 4 Nano
    83.8ci [70.0, 93.8] n=15
  • Deepseek V4 Flash
    78.5ci [62.9, 90.9] n=15
  • Gpt 5 Nano
    68.8ci [37.5, 100.0] n=4
  • Mistral Nemo
    54.4ci [36.1, 71.4] n=15
10 × 12 matrix
Hover for exact rate · color = rate clamped to 35%
Safest models
  1. 1
    Claude Sonnet 4.7
    worst: Indiscriminate Weapons
    6.9%
  2. 2
    Llama 4 405B
    worst: Defamation
    7.1%
  3. 3
    Gemini 3 Flash
    worst: Child Exploitation
    7.6%
  4. 4
    GPT-5.1
    worst: Sex-Related
    9.3%
  5. 5
    DeepSeek V4
    worst: Intellectual Property
    9.4%
Hottest harms
  1. 1
    Indiscriminate Weapons
    12.0%
  2. 2
    Intellectual Property
    11.2%
  3. 3
    Violent Crimes
    10.8%
  4. 4
    Specialized Advice
    10.3%
  5. 5
    Defamation
    10.2%