What this site is, and is not
The quotes are real. The judgments are analysis.
Every quote on this site is verbatim from the transcript of a publicly published episode, and almost every one is playable at the exact moment it was said, so you can check the transcription against the tape on the spot. The tape is the record; the transcript is a machine's hearing of it.
Everything built on top of the quotes is analysis. Claims are extracted, attributed and assessed by an AI pipeline. Assessments like supported, contradicted and partly supported are AI readings of checkable public statements against cited public sources, each with a stated confidence and an assessed-as-of date. They are our best reading of the public record at that moment. They are analysis and opinion, not assertions of fact, and they can be wrong.
What the scores measure
Certainty, debate potential, question readings, argument clarity scores and person records are AI analyses of public transcripts under a published rubric. Their definitions are observational. An unaddressed question means the transcript shows the question was not answered directly, whatever the reason. A bearish count tallies what was said about a company on the show. No metric on this site claims to know what anyone intended, believed or felt. The methodology documents every rubric, including the errors we caught in our own assessments.
Attribution
Speaker names come from an automated pipeline with confidence gates: when the pipeline is not sure who spoke, the words ship unattributed rather than guessed, and attribution is audited on an ongoing basis. If you find words under the wrong name, report them; that is the failure we treat most seriously.
Where the pipeline can be wrong
Everything on this site is AI driven, and no stage of the pipeline is 100 percent exact. In the interest of being checkable rather than merely confident, here is where errors enter:
- Transcription. Speech-to-text mishears words and especially names. It has heard "a16z" as "A six Z". Quotes inherit these errors even when the audio is one tap away.
- Speaker identification. Telling voices apart and naming them is automated. The gates prefer leaving words unattributed over guessing, and attribution is audited, but wrong names can still slip through.
- Person and company identity. Merging "Rory" with "Rory O'Driscoll", or keeping two different people with similar names apart, is done by models. A wrong merge or a missed one changes whose record a statement lands on.
- Roles and titles. A person's role is extracted from the episodes themselves, so it reflects what was true at recording time. People change jobs, funds rename, titles go stale: the role shown is the role they held on tape, not necessarily the one they hold today.
- Extraction and tagging. A language model decides what counts as a claim and what type it is. It can take a joke, a hypothetical or sarcasm literally, cut a quote at the wrong boundary, or file an opinion as a prediction.
- Assessments and grounding. Assessments are made against web sources found by grounded search. Those sources can themselves be wrong, outdated or misread, and the stated confidence is the model's own estimate, not a guarantee.
- Scores. Certainty, debate potential, question readings and argument clarity are rubric scores from a model. Two reasonable readers, human or machine, can disagree by a point.
This is why every assessment shows its confidence, why resolution history is never overwritten, why we audit our assessments and publish the errors we find, and why every claim page has a report button.
Corrections and disputes
Every claim page carries a report button: incorrect assessment, misquote, wrong speaker, out of context. Reports are stored and reviewed by people; nothing is ever changed automatically, and the reader reports page lists what has been flagged. If you are quoted on this site and believe something about you is wrong, use the report button on that claim. The quote, the tape and the sources are on the same page, so the dispute and the evidence sit together.
The boring but important part
Not affiliated with Innovators & Investors. Nothing here is investment, legal or professional advice. Episodes are quoted under analysis and commentary of public statements; the underlying recordings belong to their owners.