Narrate

A citation is not a fact-check

By default, every factual claim pulled out of an episode script is checked against the source text captured for that episode, by two independent checks that have to agree. Anything they cannot support is flagged, and a flagged episode cannot be narrated until a person has dealt with it.

  • Every claim that is checked is checked twice, independently — and a disagreement has no tiebreak.
  • Support found only in a write-up of the record is flagged, not passed: a report about a study is not the study.
  • Any unresolved flag stops the episode — it cannot be narrated until a person has dealt with it.
  • What this does not catch is set out on this page, not buried in a footnote.
Open a published episode

Its panel lists the sources, the transcript, and — where a check ran — how many claims were verified.

An episode can cite a source and still be wrong. One of ours was: the script quoted a figure from a write-up of a report, the write-up had misread the report it was summarising, and every link on the page went somewhere real. Nothing about the citation was fake — it just was not the same thing as being right. Everything below exists because of that episode, and none of it is a promise that a model behaves. It is what happens to a sentence before anyone is allowed to record it.

The check, end to end

  1. Capture

    Research stores the text of each source with the episode — whole pages, so a results table is still there to check against and not only an abstract.

  2. Extract

    A pattern pass with no model in it, then a reading pass, pull the factual sentences out of the finished script.

  3. Verify

    Two independent checks read each extracted sentence against that stored text, and each has to quote the passage it relied on.

  4. Gate

    Anything they cannot support is flagged, and a flagged episode cannot be narrated until a person has dealt with every flag.

This runs by default on every episode. The person creating one can switch it off for that episode, and then there are no claims and nothing to flag — which is why the finished page says nothing about checking rather than something weaker.

What happens to a claim

  1. Which sentences get checked is not left to a model

    The first pass over a finished script is a plain pattern match, with no model involved at all. Every sentence carrying a number, a percentage, an amount of money, a date, a comparison of size or direction, or the name of an organisation is pulled out as a claim that needs a source. It has no opinion about which of those sentences look risky, and that is the point: the sentences a model is most confidently wrong about are exactly the ones it would not volunteer.

    A second pass, one paragraph at a time, picks up the factual assertions a pattern cannot see — a causal statement carrying no figure, for instance. The two lists are merged, so a sentence caught by both becomes one claim rather than two.

    Neither pass is a promise of completeness, and it would be a strange page that claimed one here. The pattern pass is mechanical and cannot decline to pull a sentence out; the second reads each paragraph and can come back with nothing, and when it does, an assertion only it would have caught never becomes a claim — so there is nothing to check and nothing to flag. Everything below describes what happens to the claims these two passes do find.

  2. By default, each claim 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. The check runs against that stored text. It is not a second opinion on whether a sentence sounds plausible, and it is not a model being asked whether it remembers the study correctly.

    That has a consequence worth stating plainly: a claim with nothing captured to check it against is flagged, not waved through. If the sentence cites nothing, or cites something the research never managed to capture, it comes out of the process unsupported and stays that way until a person decides what to do with it. On a real episode this is the most common flag of all, and it is the behaviour we want — the bias runs towards stopping, not towards shipping.

  3. Two independent checks, and disagreement counts against the claim

    Every claim that is checked is checked twice, separately, by two systems that do not see each other’s answer. Each one has to point at the specific passage in the captured source that settles the claim — a quote, not a verdict — and that quote is confirmed to actually appear in the document before the judgement counts for anything.

    If the two disagree, the claim is flagged. There is no casting vote and no third opinion brought in to break the tie: two checks reaching different conclusions about the same sentence is itself the finding, and it goes to a person.

  4. Where the support comes from decides whether it counts

    A report about a study is not the study. Sources are sorted, as they are captured, into those that are the primary record of something — a journal, a registry, a government or institutional publication — and everything else. A claim passes on the strength of the check alone only when one of the two can quote a primary source supporting it. Support that exists only in someone’s write-up is flagged for a person to look at, even when both checks agree the write-up says what the script says it says.

    It is also why the research step goes looking for primary sources rather than hoping one turns up: a separate search runs against primary publishers alone, and places are held in the episode’s sources for what it finds, so a journal article cannot be crowded out by faster-moving coverage of it.

    The lists themselves, publisher by publisher

  5. A sentence crediting a publisher we never read is flagged too

    Attribution is checked as well as arithmetic. If a script says a named organisation reported something, and nothing that organisation published is in the episode’s research at all, that sentence is flagged even when the underlying number is correct. Borrowed authority is its own kind of wrong, and it is the kind a listener is least able to catch.

  6. A flagged episode cannot be narrated

    Flags are not a report somebody might read. Any unresolved flag — contradicted, unsupported, disagreed over, credited to a source we do not hold — 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 actually supports, or cutting it.

    Clearing a flag that way is a person’s decision recorded against the claim — who cleared it and when — and the replacement they write is theirs: the two checks do not run again over it. That is deliberate rather than an oversight. A resolution is bounded to the sentence it resolves, and treating each one as a fresh edit would restart verification of the whole script after every flag and never converge. What stands behind a rewritten sentence here is the person who wrote it, which is why the record keeps their name against it.

    Editing the script freely afterwards is the other case, and there the earlier verdicts do not stand: a general edit puts the whole claim set back in question and the check runs again before narration is allowed. So an episode cannot be narrated on the strength of a check that ran against an older draft.

  7. What ends up on the episode page

    A published episode states how many of its factual claims were checked against their cited sources, next to the list of those sources and the full transcript. That line comes from the check that actually ran on that episode, not from the fact that we have a process — and when a run verified too little of the script for the number to mean anything, the line is absent rather than optimistic.

Two ways out, and only one of them is automatic

This is the part worth reading twice, because it is where a sentence would slip through if anything were going to. A claim that has been checked leaves the process one of two ways.

A claim pulled out of the script

Clears without anyone reading it

  • Both checks agree, and one of them can quote a passage supporting the sentence from a primary source — the record itself, not somebody’s write-up of it.

Stops the episode and waits for a person

  • Nothing was captured that the sentence could be checked against.
  • The captured source does not support it, or contradicts it.
  • The only support is a write-up of the record rather than the record itself.
  • The two checks reached different conclusions, and nothing breaks the tie.
  • The sentence credits an organisation whose material the research never held.
  • The check could not be completed at all.

The three ways a person clears a flag

  1. Accept the sentence as it stands
  2. Rewrite it to match what the source supports
  3. Cut it
There is no third exit and no override in the interface. What clears a flag is the person’s decision, recorded with the claim: the replacement they write is not sent back through the two checks, which is why the name kept against it matters. Editing the script freely is the other case — a general edit puts the whole claim set back in question and the check runs again before narration.

What the check keeps, one claim at a time

For each claim it checked, the episode keeps the sentence, the source it was checked against, what kind of source that is, the passage a check had to quote, and what happened next. That record stays with the episode and can be produced for a client on request. It is not published claim by claim on the episode page: a listener gets the sources, the transcript, and the count.

Illustrative. The three records below show the shape of the thing and the three outcomes it can carry; the sentences and passages in them are written for this page, not lifted from a published episode.

  • Cleared — nobody had to read it

    The sentence
    The trial reported a 27 per cent fall in hospital admissions over two years.
    Checked against
    The trial’s own published results, captured whole when the episode was researched.
    What that source is
    The primary record — the report itself, not coverage of it.
    The passage a check had to quote
    “Admissions in the treatment arm fell 27 per cent against control across the 24-month follow-up.”
    What happened
    Both checks quoted that passage and agreed it supports the sentence. No flag, and the episode went on to narration without anyone opening the claim.
  • Flagged — supported, but not by the record

    The sentence
    Emissions from the sector fell by a fifth over the decade.
    Checked against
    A news article summarising the agency’s report. The report itself was not among the captured sources.
    What that source is
    A write-up of the record.
    The passage a check had to quote
    “The agency’s figures show a drop of around 20 per cent since 2014.”
    What happened
    Flagged even though both checks agreed, because a report about a study is not the study. A person found the agency’s own publication, rewrote the sentence to cite it, and that rewrite is recorded against the claim as their decision — a person standing behind the replacement, not the two checks running a second time.
  • Flagged — the credit, not the number

    The sentence
    According to the national statistics office, the figure has doubled since 2019.
    Checked against
    Nothing published by that office was in the episode’s research. The number itself does appear in a captured source, which does not name the office.
    What that source is
    Not in the research at all.
    The passage a check had to quote
    Nothing to quote, which is the finding.
    What happened
    Flagged although the number is right. Borrowed authority is its own kind of wrong and the kind a listener is least able to catch, so the sentence was rewritten to credit the source the episode actually holds.

What this does not do

  • It checks a script against the research that episode was built on, not against everything known about the subject. A claim simply absent from the captured sources is flagged for a person, not chased down elsewhere.
  • It is not a judgement about whether a source is right. If the research captured a study and the script reports it accurately, the claim passes — the check has confirmed the reporting, not the finding.
  • It checks the claims extraction finds, which is not the same thing as every factual sentence a script contains. The mechanical first pass cannot decline to pull a sentence out, but the pass that reads each paragraph can fail or return nothing, and extraction can fail outright — the episode then reaches review with fewer claims than it should have, or with none, and nothing flags the shortfall. The gate stops a flagged claim; it cannot stop a claim nobody found.
  • The check runs on every episode by default, and the person creating an episode can switch it off for that episode. When it is off there are no claims and nothing to flag, so the finished episode page says nothing about checking rather than saying something weaker.
  • When a check on an extracted claim cannot be completed at all, that claim is flagged rather than passed. A bad run therefore costs a delay and someone’s afternoon, which is the trade we chose and the reason flags are common enough to be worth a page. That fail-closed behaviour covers the claims that reached the check — not the step that decides which claims reach it.
  • A person is still the last step before anything is recorded, and again before anything is released. None of this replaces that. It decides what that person is looking at.

Go and check one

The panel at the top of a published episode lists the sources it cites, the transcript it ships with, and — where a check ran and verified enough of the script to report — how many of its factual claims were checked against those sources. It is the same panel on every episode, including the ones where that line is missing.

Open a published episode: Berberine, Blood Sugar, and the "Nature's Ozempic" Problem

The rest of the process, step by step