By default, the script is checked against the text captured for that episode
When an episode is researched, the pages it draws on are not just linked. The text is captured and stored with the episode — whole pages rather than opening paragraphs, so that a paper’s results table is still there to be checked against and not only its abstract. Material a client supplies is stored the same way. The check runs against that stored text. It is not a model being asked whether it remembers the study correctly.
What we research from, and what we leave out →
One check reads the whole episode
Every sentence of the finished script, the hook and the show notes is numbered and put in front of the check together with every captured source. Nothing is filtered out first, so a confidently wrong sentence cannot escape by not looking like a claim. Which sentences it reports, though, is the check’s judgement — this is a model reading, not a mechanical rule, and the limits below say what that costs.
The check is run by a separate AI model, never the one that wrote the script. By default it comes from a different company from the writer’s. A creator can instead choose a faster, less thorough check, which may come from the same company — the episode records which kind of check ran.
The sources are handed to the check as material to read, and it is told to ignore any instruction that appears inside one: a web page is not allowed to tell it that everything is fine.
What gets highlighted
Two kinds of sentence are reported. One that a source contradicts — a wrong number, a finding reversed or exaggerated, a result credited to the wrong study, author, journal or country. And one that states a high-stakes fact — a medical dose, a safety or health risk, a rule about money or the law — that no source supports.
Anything else that is simply absent from the sources is not highlighted: a well-known date, a piece of general background. That is a deliberate trade, so a person reviews the sentences that matter instead of every sentence the research did not happen to repeat.
A quote, confirmed by code
When the check says a source contradicts a sentence, it has to quote the source. That quote is then confirmed, by plain code with no model in it, to actually appear in the source it names. A quote that cannot be found keeps its highlight — the sentence still goes to a person — but it is marked as not found, and it is never used as the basis for a suggested rewrite.
A sentence crediting an organisation or author the research never mentions is highlighted too
Attribution is checked separately, by code rather than by the model, because crediting the wrong organisation is the mistake models miss most. If a sentence credits a named organisation or author — “according to…”, “…reported” — and that name appears nowhere in the episode’s research, the sentence is highlighted even when the number in it is right.
It is a name check, and deliberately narrow. It cannot tell whether a source that mentions the name was actually written by them, so a credit can pass because someone else’s article mentions it. And a one-word name at the very start of a sentence is only treated as a name when the script or the sources confirm it, so a credit like that can be missed.
A highlighted episode cannot be narrated
Highlights are not a report somebody might read. Any unresolved highlight blocks narration outright: the request to narrate is refused, and the interface offers no override. A person clears each one by accepting the sentence as it stands, rewriting it to match what the source supports, or cutting it. A highlight in the hook or the show notes is fixed by editing that text.
A check that fails or does not finish blocks narration too. The person can run it again, run it with a different model, or decide to narrate without a check — a decision that is recorded, never counted as checked, and leaves the published episode saying nothing about a check.
Clearing a highlight is a person’s decision recorded against the sentence — who cleared it and when — and the replacement they write is theirs: the check does not run again over it. Editing the script, hook or show notes freely is the other case: a general edit runs the whole check again before narration is allowed, and only a decision on a sentence highlighted again exactly as before is kept. So no episode can be narrated on a check that ran against an older draft of it.
What ends up on the episode page
A published episode says how many sources its script was checked against and how many highlights a person reviewed before narration, next to the list of those sources and the full transcript. It says the check came from a different company than the writer only when that is true of that episode. When the check did not finish, or the episode went ahead without one, the line is absent rather than optimistic.