A stop condition the maker can assert is not a stop condition

The interesting part of the checker template is the list of things it is forbidden to do. It verifies and does not fix, it never edits code, and it never reports PASS on the maker's say-so: PASS requires that `bash tests/run.sh` and `bash scripts/lint.sh` both exit 0, and FAIL arrives with the first failing line ^[8].

Those prohibitions are the mechanism. A grader allowed to repair what it grades has become the maker again, and its verdict collapses back into a self-report, which is the author's actual objection: a stop condition is a claim about reality, and the agent that wrote the code is the worst available witness to it ^[2]. The article notes the failure mode is already named in Anthropic's writeup on building effective agents, as premature victory and fake-done features ^[3]. The prescription that follows is narrower than the ladder implies: different instructions, optionally a different model, ideally a clean context each turn, with the split between maker and checker called the single highest-leverage move in loop design ^[9].

The bug in the demonstration run shows what the second attribute buys. The article defines falsifiable as being able to fail with a clear failing signal ^[7]. ShipWithAI's `scripts/check-draft-seo.sh` compared every word of a keyword against only the first three words of the title, so any keyword longer than three words got flagged even when it led the title verbatim ^[13]. Ask a maker whether its SEO check is correct and you will get an argument. The test prints the offending keyword and exits 1 ^[14], and an exit code does not negotiate.

The stop condition written for that fix is worth copying for its second half. It required the new multiword-keyword test to exit 0 and `scripts/test-rubric.sh` to still exit 0, which the article frames as guarding against collateral damage rather than only the target bug ^[15]. The cheapest way to make a keyword-placement test pass is to loosen the check that raises it, and the second clause closes that route. One clause tells you the maker hit the target. Two tell you what else it hit on the way.

Which leaves the number the piece does not have, and it is the one that decides which rung an operator can actually run. Token and dollar cost were not instrumented, and the author declines to quote a figure he does not have ^[12]. That is more honest than the alternative, but a design that puts a grader on every turn is a design whose price per lap is the whole argument, and the reader is left to work it out from their own bill.