How a science explainer is built, second by second.

On-screen text repeats the spoken word, at the same instant

When a new piece of text appears on screen, the presenter says the same word within ±3 s far more often than if the text were placed at random. Every one of the 74 episodes is above chance. Three quarters of matched pairs fall within ±1 s, which at 1 frame/s of OCR is effectively simultaneous; when the gap is larger, the text leads.

Spoken numbers behave the same way: 60% get digits on screen within ±3 s (2.7× chance, 74/74 episodes).

Text-to-speech rate ratio, per episode

Each dot is an episode; 1 = chance. Log scale.

Cuts step around words and land just before speech resumes

In running speech, an edit placed at random would split a word three times out of four. The real edit does it less than half as often, in all 74 episodes. Cuts are 1.4× as common inside pauses of 0.3 s or more, and speech picks up a median 0.18 s after the nearest cut.

Share of cuts that fall inside a word

Observed (blue) against the episode's own chance level (gray), one row per episode.

A question is answered across a cut

When a sentence ends in a question and the next begins within 5 s, a cut lands within ±1 s of the answer far more often than chance, about 0.1 s before it. The effect survives a matched control (non-question sentences with the same pause: 64% vs 49%) and holds outside the dramatised cold opens.

Cut-at-answer rate ratio, per episode

1 = chance. Log scale.

Every episode opens with the same three-beat signature

After the cold-open sketch, the greeting arrives on a cut and a music duck, both within half a second, and the music enters or changes a median 3.7 s later (IQR 3.4–4.0 s). When a sponsor message follows the cold open instead, the same bundle appears in about two thirds of episodes. It is the most stable multi-layer sequence in the corpus.

The opening template

Measured timing of the three cues around the greeting.

What is not choreographed

Several expectations failed, and they matter as much as the hits:

Sections are not marked by cuts or music any more than ordinary sentence starts are (cut within ±1.5 s: 51% vs 53%). They are marked by a music duck (22% vs 6.5%).

No "emphasis package". Stress plus gesture plus a text reveal, or cut plus music plus text, occur at chance. Stress with a zoom and a gesture is below chance (0.37×).

Cuts are not on the beat. They cluster 0.10–0.18 s before a tracked beat, which may be an editing habit, beat-tracker latency or an audio–video offset.

Zooms and sound effects avoid stressed words (0.77× and 0.78×).

Registered tests that did not show the expected marking

Rate ratio with interval; 1 = chance.

Each kind of talk gets its own kind of picture

A vision model screened 59,066 keyframes for what is on screen. Crossed with the script labels, the visual material sorts itself by the kind of talk: stories run over archive footage and original illustration, evidence over screen captures and documents, and when the presenter weighs alternatives or repairs a misconception the presenter is on camera.

Two VLM tests were registered: punch-in cuts (to a tighter frame) sit near stressed words slightly more than chance (1.06×), but no more than any other cut between the two main framings; source citations are kept off punch-ins (0.35×). Graphics tend to enter as a new term is first spoken (1.34×, borderline at ±1 s, clear at ±2 s).

Visual material by kind of talk

Rate ratio of each asset type inside each label vs the episode's usual share. Red = more, blue = less.

Two eras, two editing styles

The corpus spans two production homes: 25 episodes from New Media Academy Life (Oct 2025–Jan 2026) and 49 from AJ+ كبريت (Apr–Oct 2026). Channel and period are the same split here, so these are contrasts between eras, not causes.

The later episodes cut less often (13.4 vs 18.1 cuts/min) and hold long takes for more of the running time (45% vs 21%), with more varied shot lengths, more silence and more camera movement. Speech rate, pitch range and the amount of on-screen text did not change.

Style fingerprint: later era vs earlier era

Cliff's delta (−1 to 1). Right = larger in AJ+ episodes. Filled = q < 0.05.

Promise first, questions early, payoff last

Where a move falls in the running time is remarkably fixed. Opening promises, and questions that are left hanging for later, cluster in the first fifth, and most characters are introduced early. Recaps, callbacks to earlier moments and takeaway statements pile up at the end: 89% of takeaways fall in the last fifth.

The working moves in between (questions answered on the spot, causal explanations, analogies, jokes, cited sources) are spread almost evenly through the body. Hedges grow more frequent toward the end.

Reliability. Model-coded positions. Areas 17, 18 and 26 passed an agreement check. Areas 19, 20 and 24 are exploratory (weak agreement). Areas 21, 23 and 28 have one coder and no agreement check yet. Each episode counts once.

Where each move falls in an episode

Share of its moments in each tenth of the running time, averaged over episodes. Darker = more.

Two eras, the same script

The later AJ+ episodes run longer (median 41 vs 32 minutes), and longer episodes are written more sparsely: 73 of the 122 moves tested come less often per 10 minutes as length grows, in both eras. Once length is held equal, questions, explanations, analogies, stories, humor and cited sources come at the same rate in both eras.

The one clear change is packaging: on-screen brand elements are about a third as frequent in AJ+ episodes. Certainty boosters and signposts lean the same way, but do not survive correction for 122 tests.

Reliability. Rates come from model coders; see the agreement notes above. Channel and period are the same split, so this is a description of the two eras, not a cause.

Move rates, AJ+ vs New Media Academy, at equal length

Rate ratio with 95% interval; 1 = same rate. Core moves plus every move with q < 0.1.