ai-coding-minesIndexGitHub

A hypothesis without a control group was always wrong

Claude Code and agents

Case A — after deploy, only the new route 404s. Five hypotheses (keystore, source maps, manifest, infra outage, lost commit) were all assertions beyond the observed range, and all wrong. The real cause was one line in the dependency list.

Case B — of three deployments, two return 503. Days spent on "the bundles differ." A three-way comparison showed the three bundles were identical from the deployment's point of view; the only difference was an unrelated feature. The real cause: the affected version had never been registered in the console.

Get the working one in hand first and put it side by side.

Every hypothesis formed without a working control was wrong. The comparison table only got built after two consecutive misdiagnoses, and the moment it existed the answer fell out.