The source ladder: how a quote loses its conditions on the way to you
Almost nothing you believe about what someone said comes from hearing them say it. It arrives after three or four hops, and each hop drops something — usually the conditions attached to the sentence. Here is how to see the drop, and how to check.
What this article is about
This is about sourcing discipline — telling primary recorded material apart from descriptions of it, and knowing what has been removed at each step in between. It is not investment advice, recommends nothing, and takes no position on any company, market or policy. We build transcription tools, so the honest subject for us is how information gets handled, not what to do with it.
A sentence, and what happens to it
Consider an exchange in the question-and-answer half of a company's earnings call. This is a constructed example, not a real company's words, and the wording is chosen to be entirely ordinary — nothing here is a gotcha, and no one in the chain is behaving badly.
ANALYST. If input costs stay roughly where they are into next year, would you revisit the pricing structure?
CFO. If costs held at these levels through the first half — and I want to be clear that is not our base case — then yes, we'd have to look at pricing.
The recording exists but cannot be searched. To find this exchange you must know it happened, or listen to the whole call.
01:04:22 — the exchange above, verbatim, with the question attached to the answer and the speaker identified.
Tone, pause and emphasis are gone. Everything propositional survives.
"…then yes, we'd have to look at pricing."
The question is gone, so it now reads as volunteered rather than solicited. The conditional's antecedent is gone. The "not our base case" disclaimer is gone.
Company says it would have to look at pricing if costs stay elevated.
Accurate, and still lossy: "through the first half" has become "stay elevated", so a bounded window became an open one.
They're raising prices.
Four hops from a hedged conditional to a flat assertion of fact, with no single step doing anything outrageous. The clip is a real quote. The report is a fair summary of the clip. The restatement is what a reasonable person concludes from the report. The failure is cumulative, not local, which is exactly why it is hard to catch by being careful about any one source.
The tracing exercise
Do this once, properly, on something you already believe. It takes about twenty minutes and it is more informative than reading another article about sourcing, including this one.
- Pick a claim you hold about a company, a sector or the economy that originates in someone speaking — not a number from a filing, but something a person said. Write it down in one sentence, in the form you actually believe it.
- Write down where you got it. A specific post, article, video, podcast episode, or conversation. If you cannot name the source, stop — that is already the finding.
- Open that source and find its source. Does it link, name the event, or give a date? Repeat until you reach the event itself or run out of links. Count the hops.
- At the event, find the actual words — the official transcript if there is one, otherwise the recording. Read the paragraph before and the paragraph after, not just the sentence.
- Compare your one-sentence version against the words. Specifically: is your version conditional where theirs was conditional, bounded in time where theirs was bounded, and attributed to the right person answering the right question?
The common outcomes, in rough order of frequency: the chain runs out before the event and nobody links the source; the quote is accurate but the condition attached to it was dropped somewhere; the claim is fine and you now know it firsthand. All three are worth the twenty minutes. The second is the one that changes how you read.
Five tests that fit in your head
- Was it asked, or was it offered?
- A statement made in prepared remarks was written, reviewed and chosen. The same statement given in answer to a pointed question is a different act. Any excerpt that has separated an answer from its question has removed the thing that makes it interpretable — and clips do this by construction, since the question is the boring part.
- Does the claim have a timestamp?
- A quote with a source, a date and a position in the recording can be checked in a minute. A quote without those is an assertion about what someone said, which is a different kind of object. This is not a demand for formality; it is that "he said it about forty minutes in" is enough to make you findable and "he said" is not.
- Who is speaking, and about their own knowledge?
- An operator describing their own business, an analyst describing that business, and a commentator describing the analyst are three different distances from the fact. All three can be worth reading. Collapsing them into "sources say" is where the distance disappears.
- Date the claim, not the article.
- A piece published this week can be built entirely on remarks from fourteen months ago. The publication date tells you when someone chose to write about it, which is not information about the underlying fact. Find the date of the event.
- Does the writing want you to check?
- This is the single most useful test, because it is fast and it is rarely wrong. Good analysis links the primary source, names the event and the date, and gives you enough to disagree with it. Writing that describes a source without letting you reach it is asking to be believed rather than read, and that is a property you can detect without knowing anything about the subject.
The ladder is not a purity test
The point is not that commentary is bad and primary sources are good. Commentary is how you learn an event happened at all, and a genuinely well-informed analyst will notice things in a call that you would read straight past. Nobody has time to attend everything firsthand, and pretending otherwise produces a research process that is pure and useless.
The workable version is narrower: know which rung a given claim is standing on, and go to the source for the ones that carry weight. Most of what you read does not need checking, because most of it will not change what you do. The small number that would change something is where the tracing habit earns back its cost — and the whole reason to make verification cheap is not to verify everything, but to remove your excuse for never verifying the things that matter.
Why this got practical recently
For most of the time this discipline has been recommended, it has been advice people could not follow. Checking one sentence against a two-hour recording meant finding the recording, then listening for the sentence. That cost — call it an hour to check a claim you probably had right — is why "go to the primary source" stayed a thing people agreed with and did not do.
What changed is that the audio can become searchable text in the time it takes to make coffee, at which point checking a quote is: transcribe once, search the phrase, read the surrounding two hundred words. Minutes, not an hour. The discipline did not get more virtuous; it got affordable, and a habit that is affordable is a habit you will actually keep.
Turn the recording into text you can search in seconds
Paste a media link or upload the file. Timestamped transcript with speaker labels, exported as TXT, DOCX, SRT or JSON — so finding the exact sentence, with its question and its conditions intact, is a search rather than a listen. First 60 minutes free.
Start a transcript →Be honest about what you can actually reach
A sourcing method that assumes every source is available will collapse the first time one is not, so it is worth being precise about the terrain.
Openly published. A great deal of the material worth tracing is free and permanent by design. Regulatory filings on EDGAR are full-text searchable. Central bank statements, press conference transcripts, minutes and speech texts are posted openly on federalreserve.gov without registration. Many companies keep long investor relations archives, and many industry conferences post full sessions publicly. Start here, because the hit rate is high.
Restricted, legitimately. Some transcripts are a paid product, and paying for one is a perfectly reasonable answer when the claim justifies it. Some investor relations replays sit behind a registration form or expire. Some events are members-only. None of this is a scandal; it is just a boundary, and the correct move at a boundary is to note the claim as unverified or pay to cross it.
Technically closed. DRM-protected audio cannot be downloaded by any tool. Spotify is the routine case, and we say so at the point someone pastes the link rather than letting them wait for a failure — the honest answer is that if you have your own copy of the file you can work with it, and if you do not, find the episode where it is published openly. The same applies to anything behind a login or a platform's terms. Do not route around a protection; find the openly published version, or accept that this particular claim stays on rung 3 and mark it as such in your notes.
The tracing exercise costs nothing
Be clear about this, because the previous section mentioned paying for things: the exercise at the centre of this article requires a browser and twenty minutes. Following a chain of links backwards is free. EDGAR is free. Central bank material is free. Checking whether the writing you rely on links its sources is free, and it is the single most useful of the five tests.
Transcription enters only at the last rung of the last step, and only when the primary source is audio with no published text. Most traces do not get that far — they terminate at a filing, an official transcript, or a dead end where nobody linked anything, and the dead end is a finding you get for free. If you run the exercise once and never pay anyone a penny, you have still got the entire benefit this article is arguing for.
That last part is the piece people skip. "I could not verify this" is a legitimate result, and writing it down next to the claim is more valuable than quietly filing the claim as true because checking was inconvenient.
Common questions
What counts as a primary source when the source is speech?
The recording of the event, and a complete transcript of it. A clip is not primary — it is an edit, and the edit encodes somebody's judgment about which part mattered. A report that quotes the event is secondary even when the quote is word-perfect, because the surrounding context was chosen by the writer rather than the speaker.
What gets dropped first?
The conditions. "If X held through next year, we would consider Y" arrives as "they're considering Y". Conditionals, timeframes and explicit disclaimers go first because they lengthen a sentence without sharpening it. The question that prompted the answer usually goes at the same hop.
Isn't an official transcript also an edit?
To a degree — punctuation is a choice, and false starts and crosstalk are handled inconsistently. It is still the best available text record, and its edits are disclosed by convention rather than selected for interest. Where the exact phrasing decides something, go back to the audio at the timestamp; that is what the timestamp is for.
Do I have to verify everything?
No, and attempting it is its own failure mode. Verify what is load-bearing: the claims that would change what you do if they were wrong. Everything else can stay on the rung it arrived on, as long as you know which rung that is.
What about a machine transcript's own errors?
They exist, mostly on proper nouns, numbers and crosstalk — which is an argument for keeping the audio, not for skipping the transcript. Treat a machine transcript as a searchable index into the recording: it gets you to the right thirty seconds, and you confirm the words that matter by listening to those thirty seconds. That workflow is robust to transcription errors in a way that quoting from the text without checking is not.
Checking a quote should cost minutes, not an hour. Paste a link or upload a recording and get searchable, timestamped text you can trace against. Credits never expire.
Start a transcript →