What Yentl does
Yentl checks what is being said. It can work from a live tab, microphone audio, uploaded media, a YouTube link, a media URL, or pasted text, then keeps the source, transcript, claims, evidence, and rhetoric markers together. It is designed to support critical thinking during talks, debates, classes, and media consumption. Its verdicts are advisory: AI-generated, potentially incomplete, and always subject to your own judgment.
Accounts and saved sessions
The current v1 experience is guest-first. Saved sessions are stored in this browser first so guests can return to work on the same device. Deployments with Clerk and the Yentl database configured can also sync saved sessions to a signed-in account for restore, rename, delete, and export on another device.
Processing services
- Deepgram — speech-to-text transcription for live audio and uploaded or linked media.
- Anthropic — claim checking, bias identification, fallacy detection, and source citation analysis from transcript text.
- Vercel — application hosting, API routing, and temporary server processing needed to run the app.
Taxonomy source
The bias and fallacy taxonomy used by Yentl is derived from Cognitive Biases & Logical Fallacies Used by Antisemites by Israel B. Bitton (2024), supplemented with additional entries covering general cognitive biases, logical fallacies, and rhetorical patterns. The full taxonomy (123 entries) is available at taxonomy.json.
Funding model
Self-funded; seeking responsible partners. Yentl does not sell user data. No advertising. No third-party data brokers. See Pricing for the current public-preview posture.
Known limitations
- AI verdicts may be wrong. Language models hallucinate, misread context, and have knowledge cutoffs. Treat every verdict as a starting point for your own inquiry, not a conclusion.
- Sources may be incomplete. Web search coverage varies by topic, recency, and language. Absence of a source is not confirmation.
- Transcription is imperfect. Deepgram Nova-3 is highly accurate but makes errors — especially with names, technical jargon, accents, and overlapping speech. Errors in transcription propagate into fact-check analysis.
- Context is limited.Yentl analyzes individual claims in isolation. It does not understand the full arc of a conversation or the speaker's intent.
- Not a legal or professional opinion. Nothing Yentl produces constitutes legal, medical, financial, or other professional advice. See Terms of Service.
- Scope limits. Yentl declines to adjudicate certain claim types (private individuals, harassment vectors, extremism). See Methodology for the full scope policy.
Accessibility
Target conformance level: WCAG 2.2 Level AA (required by the European Accessibility Act, enforcement date 28 June 2025).
Current status: Automated checks on / and /session as of 2026-05-18 support the documented implementation work: skip-to-content navigation, focus ring tokens, mobile-sized primary touch targets, prefers-reduced-motion support, and aria-live transcript regions. This is not a full-product or manual assistive-technology conformance claim.
Known gaps: Audio playback controls are not yet implemented (v1 does not save replayable audio in the local library). Complex data visualizations will require additional table/chart accessibility work.
Date of last documented audit: 2026-05-18. The CI workflow defines an axe-core audit step that runs when RUN_A11Y_AUDIT is enabled.
Contact for accessibility issues: Use the accessibility contact for barriers, assistive-technology bugs, or accommodation requests.