How we work
How an episode is made, and every place it could go wrong.
Here is exactly what happens between a question and a finished episode, what AI does, what people do, and what catches a mistake before you see it. Everything on this page is a rule we hold ourselves to, written down under Standards.
What AI does
It drafts the research and tags every fact with its source. It writes the script in the character's voice, with a source marker on every claim. It renders what has never been filmed, cuts the episode and lays the sound. And its work is checked against written rules at every step: a second, independent check of every claim, a picture check against the sources, a quality check of every cut and every label.
What people do
Our editorial team decides what we make, and it works unlike a newsroom: AI reads every source, drafts every check and builds every picture; the editors decide. Each show has its own editors, specialists in its field; the editorial board holds the five gates together: the pitch, the script, the pictures, the cut and the release. The rules that matter most stay with them: the characters, the gates and every check, and no check is loosened without their word. Editors read what viewers report and answer when something is wrong.
The seven stages of an episode.
1 The editors say yes
Before anything is madeEvery episode starts as a pitch: the question, what you will see, what has to be true, and how sure we can be. The show's editors and the editorial board read it and say yes or no before anything is made.
2 Research
Primary sources onlyThe topic is worked from three sides: the strongest case for, the history and the context, and the strongest case against. Every source cited is opened, and what does not say what it is quoted for is thrown out. The result is a dossier where every fact carries its source and its grade.
3 Script
In the character's voiceThe script is written in the character's voice, with a source behind every claim. A fact without a source goes back before the script moves on. The show's editor locks it.
4 Fact-check
Two independent checksEvery claim in the script is checked at its primary source twice, by two checks that never see each other's work. A claim is green only when both agree. Disagreements are settled from the sources, never averaged. A central claim needs two independent sources, one of them a study, official data or an original document. The same pass checks the law and the platform's rules: health, drugs, living people, advertising.
5 Pictures
Checked before and afterEvery scene set in a documented place and time is checked against its visual sources before a picture is made. Every picture is checked again after: does it suggest something the claim does not hold? The picture desk approves, and living people are never generated.
6 Finish
Every cut, every labelSound, captions from the locked script, and a quality check of the whole master: every cut, a frame every few seconds, every label and number against the claim sheet. The companion booklet lists every source, claim by claim, and says for every key study whether it was replicated. Then the editorial team approves the finished episode, or it does not go out.
7 Release and review
28 daysWe read the comments 2, 7 and 28 days after release, and a mistake can be reported by email at any time. A reported error goes to the fact-checker, never to the character. A confirmed mistake is corrected in the channel, in the description and the pinned comment, and we say why it happened.
Every place it could go wrong, and what catches it.
| What could go wrong | What catches it | What you see |
|---|---|---|
| A source is misread | Two independent checks at the primary source; a claim is green only when both agree | The source in the description and the booklet |
| A number is read in the wrong unit | Every label and every number is checked against the claim sheet before release | Units spoken and shown in metric, with their source |
| A picture suggests something the claim does not hold | The picture check: every documented scene against its visual sources, before and after it is made | How sure we are, on the scene |
| A quote is invented | Every quote is matched word for word at its source; invented words are marked as invented | The mark on screen |
| A real person is shown who never agreed to it | Living people are never generated; people of the past only when dead for more than 70 years, with a documented appearance | Real people only as objects: a photo, a document |
| The evidence is thinner than it sounds | The certainty is named: the exact verb, the size of the evidence, and "we don't know yet" where it is true | The character says it, out loud |
| A character seems human | Our characters never claim to be human; in the first seconds they do what no person can; every episode carries the AI label | The label on every episode |
| We get it wrong anyway | The comments are read at 2, 7 and 28 days; a reported error goes to the fact-checker; a confirmed mistake is corrected in the channel | The correction, with its cause, where you watched |
Every picture, checked twice. Before it is made, and after.
The picture desk
Before
Every scene set in a documented place and time is checked against its visual sources first: the excavation plan, the fossil, the probe's images. The show's look book sets how every picture of that show is made.
After
Every finished picture is read against the claim it stands next to: does it suggest something the claim does not hold? If it does, it goes back.
On screen
Every picture carries how sure we are: real evidence shown as it is, rendered from the evidence, or we don't know yet. Each show keeps its own words for it.
Never
No picture that could pass as a real recording of a real event. No living person generated. Everything generated is the world of the show; everything real is an object inside it.
One claim, every check it passed.
From an episode of Tycho Takes Off, as it was checked"Light from Io takes 33 to 54 minutes to reach Earth." On screen, the place line: Io · 33–54 light-minutes.
NASA NSSDCA, Jupiter Fact Sheet: distance from Earth, 588.5 to 968.5 million km · grade A, official data
NIST CODATA 2022: the speed of light, 299,792,458 m/s · grade A, official data
588.5 and 968.5 million kilometers at the speed of light give 32.7 and 53.8 minutes. The script rounds to 33 and 54.
Confirmed at both sources by the first check on 24 October 2026. Checked again, independently, from the claim and the locators alone. Both agree, so the row is green.
The show's own label: modeled, computed. The number is not a measurement of a light ray; it is calculated from two measured values.
The claim, its minute in the episode, both sources with their pages, and the range.
How sure we are, on every scene.
Every scene that makes a claim shows how sure we are. The show's own device on screen says what the picture is, where it comes from and how far the evidence carries it. The studio uses the same grammar for every picture on this site:
Examples. Rendered: drawn from the evidence; real evidence: a document, a photo, a measurement, shown as it is; we don't know yet: nobody has measured it, and the episode says so.
| Shown as it is: a measurement, a document, a photo. | |
| High: computed or reconstructed from measured values and established physics, with the sources named. | |
| Best evidence: supported by studies or records that leave gaps; the episode names the size of the evidence. | |
| We don't know yet: no measurement exists, and the character says so. |
Each show keeps its own words on screen: Tycho says observed, modeled or imagined; Harper ends with a verdict. This site uses one grammar for its own pictures.