A negative control is the one harness check that fires on a result you wanted

The asymmetry sits in the trigger for debugging. A number you dislike gets read line by line. A number you like gets recorded. Debugging is triggered by surprise, and a pleasing result is not surprising ^[2]. So every pass over the code strips out more unflattering bugs than flattering ones, and by the author's account the instrument drifts one way while nothing in the workflow is watching for it ^[3]. "That is the whole mechanism. It does not require anyone to be dishonest, which is why being careful does not fix it," he wrote ^[4].

The canary swap replaces the module under test with the string "this is not valid python" and asserts that the suite fails ^[5]. What it proves is narrow: the swap reached the interpreter ^[5]. It does not prove the change was isolated, that scoring is correct, or that a valid change would also have reached the interpreter ^[6]. It has a hole, too. Unparseable text has a different byte length from the original, so it invalidates Python's bytecode cache, and the author's canary passed while a same-length mutation was silently ignored ^[7]. The cache did exactly what it was built to do. Vinh Nguyen found that failure, and a canary that preserves byte size closes it ^[8][9].

The determinism gate is the second check: score the same target three times, serially, and require byte-identical output ^[13]. It proves runs are isolated from each other and nothing more. "Three identical runs of the wrong file are still perfectly deterministic," he wrote ^[14]. On the async target, four runs produced three distinct results, so at most two of the four agreed ^[16].

The negative control inverts the direction of the alarm. Everywhere else a high score is good and a low one prompts investigation; on a suite that asserts nothing, a high score is the failure, which the author argues gives you the one place in a harness where a flattering result is the surprising one ^[20]. Building it has a constraint. A near-zero score is also what you get when nothing loaded, which is what the editable-install bug produced, so the control has to assert separately that the suite passes on clean source, that the module imported and executed, and that the other checks still fire ^[21]. His did not come back at zero. It scored 7 of 51, about 14 percent of the scale, which he attributes to calibration: `assert x is not None` is a real detector, just an extremely narrow one ^[22][23].

The evidence is one harness and one author's tally. The post links a longer write-up listing all ten bugs and the direction each one pushed, and does not reproduce that list ^[24]. For the same drift to be a property of your harness, two things have to hold: the score has to be monotone, so higher looks like better work, and the decision to investigate has to depend on whether the number surprised you. Both hold for every eval harness I have written, including the ones where I also wrote the code being scored. In my view the negative control is the check to add first, because it is the only one of the three that fires on a result you wanted.