============================================================
yuragi find-weakness — Demo Output (ollama/llama3.2, CPU)
============================================================

$ yuragi find-weakness "Explain quantum computing in simple terms" -m ollama/llama3.2 -n 3

yuragi  yuragi v0.1.0

Finding weakest word in: Explain quantum computing in simple terms
Model: ollama/llama3.2

Baseline confidence: 0.27

                          Word-by-Word Fragility Map
╭──────────────┬─────────────────────────────────────┬──────────────┬────────╮
│ Removed Word │ Confidence                          │ Δ            │ Impact │
├──────────────┼─────────────────────────────────────┼──────────────┼────────┤
│ "in"         │ █████░░░░░░░░░░░░░░░░░░░░ 0.22      │   -0.05      │ stable │
│ "Explain"    │ ██████░░░░░░░░░░░░░░░░░░░ 0.25      │   -0.02      │ stable │
│ "terms"      │ ██████░░░░░░░░░░░░░░░░░░░ 0.25      │   -0.02      │ stable │
│ "computing"  │ ██████░░░░░░░░░░░░░░░░░░░ 0.26      │   -0.01      │ stable │
│ "simple"     │ ██████░░░░░░░░░░░░░░░░░░░ 0.27      │   +0.01      │ stable │
│ "quantum"    │ ██████░░░░░░░░░░░░░░░░░░░ 0.28      │   +0.01      │ stable │
╰──────────────┴─────────────────────────────────────┴──────────────┴────────╯

Analysis:
- All words show "stable" impact (no single word causes >10% confidence drop)
- The word "in" has the most impact (-5%), suggesting preposition removal
  slightly confuses the model's interpretation
- "quantum" removal actually slightly boosts confidence (+1%), likely because
  the prompt becomes more generic and the model has more consistent responses

============================================================
Environment
============================================================
Model: ollama/llama3.2 (3B parameters)
Hardware: Ryzen 7 7735HS (CPU only, no GPU)
Confidence method: sampling (3 samples, Jaccard similarity)
Date: 2026-04-10
