From the first sentence to a reviewed film

An AI is asked to import a customer list. An ordinary file works. Then an empty file arrives — and the program fails. How do we know the assignment is really finished?

This small example appears in our HLE/METR film. It makes the difference between a correct answer and a connected task tangible. The viewer can check the outcome: an empty input must be handled without an error. A list of technical names could not show that difference.

Production therefore starts with a learning goal, a concrete example and a sequence of pictures. For every new tool, we ask what it is for, what it receives, what it does and what comes out.

The narration and drawings use the same example: receive a customer list, run the import, read the error, correct the code, test again. The check mark appears with the successful test. It should not reveal an outcome before the voice explains it.

For current claims, we read original sources and preserve the conditions beside the claim. An HLE score needs its tool setting and experimental setup. Our example arithmetic and possible future applications are explained as such. A striking number deserves a story without invented certainty.

ElevenLabs receives the prepared text and selected voice and produces speech clips. We measure each file’s actual duration. Those measurements determine chapter starts, short pauses and a quiet ending. A spoken sentence finishes before the next chapter starts.

Forced alignment associates the words actually spoken with their timestamps. These timestamps control important reveals and subtitles. When the voice says “64.7 percent with tools,” that value appears with its condition.

The example calculator above shows how seconds become video frames. It is a checkable calculation for one hypothetical speech clip; real production uses measured audio files.

Remotion describes pictures and scenes using React components. It receives drawings, text, audio and an editing plan and produces a film. Our shared design system supplies the website and videos with colours, fonts and approved characters. Numbers retain the same meaning and hierarchy in light and dark themes.

Johannes briefly opens and closes the story. In between, the explanation gets the canvas. A few complete hand drawings change calmly, held for about one or two seconds. The mouth stays closed; the drawings do not simulate lip sync.

Small held offsets make other drawings feel like paper pieces placed on the canvas. A chart moves as a whole; individual data points do not shift. Real source screenshots use readable browser crops with their origin and capture date. A tiny screenshot would not help a phone viewer.

Rendering runs on our Hetzner video worker. We then review all actual speech clips and the complete film with sound. We also inspect individual frames: scene openings, important reveals, cuts, gesture changes and the final image.

We look for arrows that accidentally form black shapes, labels covering symbols or a clipped final syllable. Automatic reports can find technical defects. A positive model assessment does not replace direct frame inspection. A defect means a correction, a new render and another review.

After approval, the film, poster and subtitles are stored in R2. A file hash identifies the exact bytes. We read the uploaded files back and compare them with the reviewed result. A release manifest connects these files to their article.

German and English have their own speech, timelines and visible text. A new film needs review in both themes. The published page also gets a playback, subtitle and phone-layout check.

The story ends with the viewer: an understandable film, its sources and — for explainers — an optional understanding test in your own words. The next film starts with another question.

HLE and METR: the boundary beyond the boundary puts the import test in context. Our design system documents the shared style. Technical background: Remotion, ElevenLabs Forced Alignment and Cloudflare R2. This sequence describes our actual production workflow; fully automatic Daily video production is still being built.