The evidence contract

Podcast claims with the evidence attached.

High Signal Podcasts indexes what was actually said on a podcast and links every claim back to its source: the episode, the speaker, the date, and a verbatim transcript excerpt. It is not a summarizer - no claim here is model-written prose about an episode. A claim is published only if its quote validates verbatim against the stored transcript segment; if it does not match, it is never published, and no confidence score can override that. As of 5 September 2026 the index holds 30,562 published claims across 1,208 episodes from 23 trusted shows.

One real claim, end to end

What a claim is here.

Verified live on 2026-09-05. Every field is checkable via the public API.

"It's unfortunate that they work so well. It's unfortunate that scaling works so well, because it's too easy for people to just sweep these problems under the carpet."

- Llion Jones, on Machine Learning Street Talk, said 23 November 2025. Attribution: verified speaker.

Speaker: Llion Jones, verified_speaker
Show: Machine Learning Street Talk
Date said: 2025-11-23
Quote: byte-identical to the stored evidence row
Transcript kind: publisher_html
Episode: resolves to a page on this site with the excerpt

The claim's own record is public: GET /api/claims/27d6cf86-8403-4185-8c1d-056448739ecd returns the claim, its evidence row, and its references. The quote field on the claim and the quote field on the evidence row are the same string - that identity is what "verbatim" means here.

The refusal conditions

The gate, exactly.

Conditions under which the pipeline refuses to publish. Enforced in judgeClaim in the API worker.

Condition	Outcome
Quote does not validate verbatim	Never published - `quote_not_verbatim`
Speaker is unknown	Never published - `unknown_speaker`
Both confidences >= 0.85	Published
Both >= 0.65, not both >= 0.85	Held
Either below 0.65	Draft
Unverified speaker, extraction >= 0.85	Published, labelled `high_confidence_unverified_speaker`

A high score cannot rescue a bad quote. The verbatim check runs first.
Unverified speakers do not inflate the people index. Person aggregates count only verified_speaker rows - covered by a test.

The negative space

What we do not do.

The negative space is load-bearing. It is what makes "evidence" a category rather than a slogan.

-
No summarization.
Nothing on this site is model-written prose describing an episode. Every published string in a claim is a quote.
-
No inference about what a speaker meant.
The index records what was said, not its intent.
-
No fact-checking verdicts.
We verify attribution - that this person said this, in this episode, on this date. We do not adjudicate whether they were right. That is why this site does not use ClaimReview markup anywhere.
-
No ranking of speakers.
There is no authority score and no leaderboard.
-
Absence is shown as absence.
An empty search returns explicit insufficient evidence rather than a generated answer.

Honest boundaries

What this does not guarantee.

Transcript errors survive. An excerpt can faithfully reproduce a mistake made by a publisher or caption provider. The original episode remains the final authority.
The date is the episode's publication date. Not the same as the date the words were spoken.
Timed links are conditional. A timestamped link appears only when the excerpt came from captions on that same video.

Live figures

What the corpus contains.

Read live from the public API on 5 September 2026. Each figure names its scope.

Figure	Value	Scope
Published claims	30,562	trusted shows only
Episodes with claims	1,208	site-wide
Attributable people	956	verified-speaker only
Episodes catalogued	8,509	includes no-claim episodes
Episodes with transcripts	1,226	resolved, extraction pending
Trusted shows	23	active and trusted
Published references	1,141	individual reference rows
Shows withheld	2	unsafe diarization

Why the catalogue is seven times the published set. 8,509 episodes catalogued; 1,226 have transcripts; 1,208 have published claims. The gap is the gate doing its job - an episode with no usable transcript produces no claims, and summarising from a description is the thing this index exists not to do.

Check it yourself

How to check any claim.

The public API is read-only. The website and API currently require no account, API key, subscription or checkout - which describes the product today and is not a promise of a permanent pricing model.

GET https://api.podcasts.highsignal.app/api/stats
GET https://api.podcasts.highsignal.app/api/claims/{id}
GET https://api.podcasts.highsignal.app/api/search?q=agents

How episodes are discovered, transcribed and extracted

The index

Browse the index.

>
Episodes
Every episode with published evidence
>
People
Speakers with verified attribution
>
Books and tools
Named references pulled from claims
>
Public API
Machine access to all of it