The decisive parameter was never in the registration document. It lived in the pipeline, as whatever sample count the code happened to use, so the artifact meant to adjudicate the run had handed half the decision to a default that no retro would notice changing [2].
The median is what makes that bite. A median over 8 samples sits between the fourth and fifth ordered values; over 24 it sits between the twelfth and thirteenth [21]. Those are two different estimators of the same quantity, and if the per-frame motion values are bursty, they can land on opposite sides of a fixed line. Here they did: 47.5% at 8 frames, 56.5% at 24 [3][4]. That is a spread of 9 percentage points across a bar sitting 2.5 points above the low reading and 6.5 points below the high one [17].
I would not carry the 9 points anywhere else. For that gap to transfer, your clips would need a comparable motion distribution and your sampling rule would need to draw frames the same way, because the number is a property of one clip's order statistics, not a general fact about 8 versus 24. What does transfer is the shape of the failure. hexisteme's own reading, in a first-person retro published on dev.te's platform under the hexisteme byline, is that a threshold like this is undefined rather than strict, since both verdicts were honestly computed [6][20]. The rule that followed is short and correct: pin sample count, measurement window and sampling rule alongside the threshold [8].
The next pre-registration is the part I would copy. It fixed the window to [0, 5.5s], spelled out the sampling rule, and committed to reporting both 8 and 24 samples [9]. Committing to both is better engineering than picking one, because it puts the sensitivity in the report instead of resolving it by choosing whichever count agrees with you.
Then the same defect reappeared in the process layer. An audit the next day counted three files matching report/prereg-*.md and found one that had pinned all three slots in prose, which the author wrote up as the rule failing inside 24 hours [11][12]. A single git log on first-add dates showed the compliant document was born in the rule's own commit, and the other two on 2026-08-14, before the rule existed [13]. Two of the three were therefore never in scope [19]. Scoped correctly, the denominator is 1 and compliance is 100% [14]. Same numerator, same corpus, same day: 33% or 100%, a 67-point swing produced entirely by which documents you decide the rule applies to [18].
That leaves a compliance figure of one over one, on a rule whose only test case says in its own text that it is enacting the lesson for the first time [10][15]. A rule whose entire evidence base is the document announcing it has a tidy citation graph and not much else. Both compliance readings were arithmetic done correctly on an unstated denominator, which is the threshold bug one layer up, in the process rather than the metric.
