promptdojo_

Experiment tracker lite — step 6 of 7

Write compare(a, b): print the config diff as {key: (a_value, b_value)} for every differing key (sorted, using .get so one-sided keys work), then f"acc: {a} -> {b}", then "causal claim allowed:" — True only when EXACTLY ONE key differs. Run it on the two runs. The one-change-per-run rule, made mechanical.

full-screen editor opens — close anytime to keep reading.