Drift Compounds: twelve A-grades make a coin flip

An agent runs the same 12-step workflow a hundred times. Every step succeeds 95% of the time, and a failed step doesn't crash anything: it hands a plausible, slightly-wrong artefact to the next step, which builds on it politely. Run the hundred. Then place a checkpoint and see what a human can still catch.

The workflow · click a gap to place a checkpoint (max 4)

A checkpoint is a human review. Its bracket shows the only steps it can still meaningfully verify: the last three. Older drift has been built upon; it reads as foundation, not error.

One hundred runs · all of them finish politely

The tally

–
correct
–
silently wrong
Expected–
Steps no checkpoint can see–
Human reviews this design costs–
"Verify every step of every run" costs1,200
–

Drift compounds because each step trusts the step before. 0.95 twelve times is 0.54, and no bead ever crashed to warn you: the wrong 46 are formatted, polite and sent. A checkpoint is a human asking "is this still what we meant?" while the error is shallow enough to see; one big review at the end inspects a finished artefact whose early mistakes are now load-bearing, which is a rubber stamp wearing diligence's clothes. And the toy flatters you: its reviewers catch everything in their window. Yours won't. Design the checkpoints, don't promise the diligence.