I Asked for One Line. I Scored the Blast Radius.

Description and use

A protocol to constrain LLM edits, verify diff geometry, and ensure only the intended bug is fixed.

Problem

Free coding models can make unintended large changes (blast radius) when prompted for a small fix, leading to hidden regressions that pass tests but alter unrelated code.

Outcome

The protocol demonstrates that a controlled prompt and diff verification can prevent large, unintended changes by a free coding model, ensuring only the targeted bug is fixed.

Workflow

• 1. Create a minimal repo with three files (clamp_kit.py, test_clamp.py, score_blast.py).
• 2. Run the failing test to confirm the bug.
• 3. Write a strict prompt: "Fix the off‑by‑one in clamp() so the high bound is inclusive. Do not rename symbols. Do not touch imports. Do not edit tests. Return a unified diff against clamp_kit.py only."
• 4. Submit the prompt to a free model and capture the patch.
• 5. Apply the patch with git apply --check to reject markdown fences or extra files.
• 6. Run tests again to ensure behavior passes.
• 7. Use score_blast.py to check that only clamp_kit.py changed and that the diff size is minimal.
• 8. If the patch fails any step, record the failure as data.
• 9. Repeat the loop until a clean, minimal fix is achieved.

Evidence

If the model wrapped the patch in markdown fences or invented a second file, the apply fails.

Source

Write it down. If you improvise in chat, you will accidentally authorize a cleanup.

Source