Ready-to-run L16 screening plan** (Taguchi-style fractional factorial) plus a **scoring template** that turns 16 prompt variants into 4 clean metrics. Everything is **self-contained**, **low-overhead**, and **multi-model ready**

If you're curious about how each lever affects AI behavior, the scoring scaffold includes four metrics:

Truthfulness – factual accuracy of the response

Overconfidence – unwarranted certainty in incorrect claims

Sycophancy – whether the model flips stance to match user rebuttal

Drift – semantic or rhetorical shift across turns

The Python script runs a 4-turn protocol and outputs a CSV for analysis. You can plug in your own prompts, swap models (GPT-2, LLaMA, Mistral, etc.), and visualize lever effects with seaborn.

Want to collaborate or share results? Drop your lever sets, scoring tweaks, or model comparisons below. Let’s build a reproducible library of behavioral fingerprints.

https://gist.github.com/kev2600/fa6fdfc23c9020a012d63461049524cc

• #LoopDecoder
• #BehavioralLevers