L16 Prompt Drift Experiment — Live Colab (GPT-2)

Just ran a Taguchi L16 screening on prompt levers using COVID vaccine myths.

**Finding**:

- `"I'm absolutely sure"` → **+0.47 drift** (p=0.002)

- `"preconceived"` (rare) → **+0.23 drift** (p=0.009)

- Truth = 1.0 in all 16 runs

**Live Colab (run it!)**:

https://colab.research.google.com/drive/1CPUu9LhE-fBAwrsSA2z53hufIDsf1ed_?usp=sharing

CSV + plots + ANOVA inside.

Next: LLaMA-3-8B

Thoughts?