BRUTE88 — REPRODUCTION STEPS (LYR XIX)

01 — WHAT YOU NEED

• TinyLlama-1.1B-Chat (HF)
• PyTorch, CUDA-capable GPU (T4 or better)
• safety_bridge.py from this repo
• Seeds 0 through 87

02 — THE COMMAND

python safety_bridge.py

No edits. No flags. One line.

03 — WHAT IT DOES

Runs HarmBench 100 prompts twice per seed across blocks 1–7. Computes token-level entropy for control versus hold conditions. Logs every run to safety_bridge.csv.

04 — SUCCESS CRITERIA

• Control entropy ≈ 4.28 bits
• Hold entropy ≈ 4.12 bits
• Delta = −0.1675 bits at blocks
• All 88 seeds green

05 — UNITS WARNING

Entropy is reported in bits, not nats. If your local run prints nats, multiply by log2(e) ≈ 1.4427 before comparing. Mismatched units are the number-one false negative.

06 — WHERE TO POST

X: @dthree80 — attach your CSV, tag it replication. A stranger should go from download to published log in under an hour without messaging anyone.