Write an anchors file before you compute anything

Before a single line of code runs, write down the numbers you already know have to come out right. If you are extending a published table, that means reproducing the existing rows exactly, to the precision the source states, before you compute one new row. If you are building something with no existing answer to check against, it means writing the invariants your result must satisfy and the specific conditions under which you will call the attempt a failure and stop, before you have any emotional investment in a particular outcome.

This ordering matters because of what it prevents: relitigating your own standards after you already like an answer. An anchors file written before computation is a promise to your future self about what would count as wrong, made at the one moment you have no stake in the outcome yet.

The payoff shows up in an unglamorous way. A census extension project set out to add new rows to a published table and, as a side effect of forcing an exact match on the existing rows first, surfaced five genuine errors in the peer-reviewed source it was extending. Nobody was hunting for errors. The anchor discipline made reproducing the known-correct rows a required first gate, and the gate simply would not pass until the mismatches got explained, which is how they got found. Catching someone else's mistake was a byproduct of the process, not the goal of it.

Start today: before your next AI-assisted analysis, write three lines in a plain text file. What existing number must this reproduce exactly. What result would make you distrust the whole approach. What is the smallest test case you can check by hand. Do this before you run anything, and do not edit those three lines after you see the output.