PairedY

Scans the lecture a second at a time and keeps only the frames that are settled and fully revealed — the difference between a 50-slide deck and 500 near-identical shots.

Profiles the whole lecture in one pass, sorts each frame into slide / board / presenter, then finds the slide rectangle and cleans the presenter and webcam out of it.

Shows the model the crop rectangle drawn on the frame and asks whether it is right. Asked to produce one from nothing the same model answers “the whole frame” and the recording's black bar ships with the slide; asked to review one, it corrects it.

Pulls the title, bullets, code, tables and equations off every picture. The only call made once per slide, so it is the biggest single cost of a conversion — and the notes can only contain what it reads.

When a page visibly full of text comes back almost empty, it is read once more here. Keep this on a DIFFERENT model from the one above: measured over 21 blank pages, the two never failed on the same one.

Reads the finished deck as a whole and drops the pictures that repeat, supersede or abandon each other — each one folded into the slide that replaces it, never simply deleted.

Reads the finished slides as contact sheets and merges the ones that are the same page shipped twice. Every earlier pass compares pixels between frames, which cannot see that two different crops are one page — the crop is what those metrics measure.

Merges what each slide shows with what the speaker said into the note under it. This is the pass a reader judges the result by.

Polishes the notes, condenses the narration into margin notes, repairs the transcript against the slide text, pulls out the concept graph and balances the handout pages.

Decides the pages of the documentation site: which slides belong together even when the lecture separated them by an hour, what order a reader needs them in, and which carry nothing. It plans only — the pages are rendered from the deck, so it cannot invent a fact.

Writes each page from the raw material — the text on every slide and what the speaker said over it — not from the notes, so it can reorganise rather than annotate. Every number, path and command is checked back against the lecture: a page that invents one, or drops the commands it promised, is asked again and then published as slide notes instead. Add @Host to prefer one provider.

Looks at the rendered PDF pages and reports anything clipped, overlapping or broken — then the layout is re-packed and re-rendered if that measurably helps.

Turns the lecture audio into the timestamped transcript every later step aligns against.

Only reached if you clear one of the text boxes above. Nothing uses it otherwise.