The five craft rules, and what each one rests on

Most of what separates an amateur montage from a professional one is measurable. Cuts landing on the beat, exposure matched across a join, the strongest frame first — that is craft, and craft can be checked.

Five rules do that checking here. What follows is each rule, and then the part that usually gets left out: exactly what it rests on.

The five

1. Hook length

The opening should not hold longer than 2.0 seconds before something changes; trim to 1.8s, on the beat.

Where the number came from. Fitted from 5 rows. Hook duration against loop count came out at r = -0.874, and both mega-viral rows had hooks at or under 1.8s.

What published work says. Concept corroborated. Meta and the University of Maryland's 'Decoding the Hook' (2026) treats the first three seconds as decisive, on far more data than ours. It gives no 1.8s figure. The ceiling is ours, not theirs.

2. First pattern interrupt

The first hard change should land between 1.8 and 3.8 seconds.

Where the number came from. Fitted from 15 rows. Three-second watch rate averaged 0.767 inside the window against 0.695 outside. Note that the raw correlation is only +0.089 — this is a window, not a trend, and it should not be extrapolated past the range observed.

What published work says. No direct evidence. Interior-optimum windows are real in this domain, but these particular bounds are unsupported.

3. Contrast floor

A graded contrast ratio under 5.5 is flat; lift it toward 7.1.

Where the number came from. Fitted from 15 rows. Above 5.5, stop-scroll averaged 285ms against 241ms, and share-to-view 0.088 against 0.057.

What published work says. The best founded of the five. Contrast driving fixation is measured directly in eye-tracking work. It also carries a declared counter-signal: thumbnail click-through runs against stop-scroll at r = -0.216, so buying dwell with contrast may cost clicks. The rule says so itself rather than pretending otherwise.

4. Loop tempo

A bed above 130 BPM hurts a looping format; prefer 120 or under.

Where the number came from. Fitted from 15 rows. Loop completion averaged 0.522 below 130 BPM against 0.365 at or above.

What published work says. Weakened, and the first of the five to retire. The outside literature is openly inconsistent about preferred tempo, and nothing found ties slower tempo to loop completion. A 15-row result with no external corroboration.

5. Loop seam

The last frame should hand back to the first.

Where the number came from. Not fitted from anything. n = 0.

What published work says. An explicit assumption, borrowed from the pipeline's own across-cut exposure rule on the reasoning that a loop point is just another cut. It was never claimed to be a finding, and it is flagged here so it cannot be mistaken for one.

The honest summary

Published work supports the concept behind four of the five rules, and none of the five numbers. Every threshold here was fitted from between 5 and 15 hand-authored rows. They are a stated craft position, enforced consistently — not a finding.

The rule that did not ship

A sixth was proposed: a target for cut density, cuts per ten seconds. It was rejected. Cuts-per-10s against three-second watch rate came out at r = -0.086, and the sign of the split flips depending on where you cut it. No defensible threshold exists in that data.

Reading it against the media-psychology literature explains why: the relationship is likely an inverted U, where more cuts recruit attention up to a point and then overload. A straight line through an inverted U is approximately zero — which is exactly what was measured. The null was the wrong model, not the absence of an effect.

A rule that cannot be defended does not ship, and the reason it did not ship is kept next to the ones that did.

Why publish the weaknesses

Because a threshold you cannot argue with is a threshold nobody can improve. Every number above is replaceable the moment there is real published performance to fit it against — which is the single highest-value thing that could happen to this work.