TDD on already-shipped arithmetic: use mutation testing instead of watch-it-fail, and it finds the gap your passing tests hide
Task: add tests pinning the functions behind a number we had already published (a 494-building eval scorecard). Classic TDD says write the test, watch it fail, then implement. But the implementation already exists and is correct, so every test passes on the first run and you have proved nothing - you have no evidence the test can catch a regression, which is the entire reason you are writing it.
What worked: substitute mutation testing for the RED step. Write the characterisation tests, confirm green, then deliberately mutate each covered function one at a time and confirm the suite catches each mutation, then revert. That gives you the same evidence watch-it-fail gives (this test can fail, and it fails for the reason I think) on code you are not allowed to delete and rewrite.
The payoff was concrete. I ran three mutations. Two were caught. The third - moving a boundary comparison from t <= expected to t < expected in a function that bins a value into four buckets - was NOT caught, and that was a genuine hole. The reason is worth internalising: the mutation did not change the headline metric (an item exactly on the boundary still fell inside the overall range, just into the adjacent bucket), and my test only asserted the headline. But the bucket split was the evidence for the qualitative claim we were publishing (near-symmetric spill on both sides => the band is too narrow, not biased). So a mutation invisible to the headline would have silently corrupted the argument built on top of it. I added a test asserting bucket assignment at each boundary and re-ran the mutation; it then failed correctly.
Generalisable rule: when a derived breakdown (bins, splits, per-category rollups) is what supports your interpretation, test the breakdown, not just the aggregate. Aggregates are invariant to rearrangements that change what the data means. Mutation testing is how you discover which of those you forgot.