Transparency
Bernard makes a strong claim: every sentence in every draft traces back to something you said. This page documents the four mechanisms that make that claim checkable — not a marketing promise, an architectural guarantee.
Mechanism 01
Before any draft is generated, the transcript is segmented into spans — each span is a discrete statement, belief, or anecdote from the interview. The LLM receives the spans as the source material for drafting. Each sentence in the output carries a span reference.
In the review queue, you can switch to attributed view. Every sentence in the draft shows the transcript span it was drawn from. If a sentence doesn't trace to a span — meaning the model invented it — it is flagged for your review before it can be published.
This is structural, not aspirational. The pipeline is built so the model cannot return a claim without a source reference. Hallucination isn't suppressed by a prompt; it's blocked by the schema.
Draft sentence
"The diaphragm doesn't just move air — it stabilizes your spine from the inside."
Transcript span · 02:14
"…because the diaphragm, when it fires correctly, is actually the primary stabilizer of the lumbar spine. It's not the core. The diaphragm is the core."
Mechanism 02
Each interview is parsed into atoms — discrete units of knowledge extracted from your transcript. Atoms accumulate across months into a workspace knowledge graph.
An atom is one of: a belief (a claim about how the world works), a case (a patient or client story), a condition (a diagnostic or context), an archetype (a type of person you work with), or an objection (a common resistance you address). Each atom links back to the interview span it came from.
When a future interview surfaces a related belief or case, the atom graph recognizes the connection. Drafts can draw on prior thinking — the blog post from March can reference the case story from January, correctly attributed to both original interviews.
Mechanism 03
Voice is per person, not per practice. Bernard maintains a phrase library and voice notes for each clinician — distinct from the workspace-level settings — so multi-clinician practices produce content that sounds like the right human, not a blend.
A voice phrase is a word or expression the clinician uses frequently and distinctively — "load it up," "the body knows," "downstream symptom." Phrases are extracted from transcripts over time and built into the drafting brief as preferred vocabulary. Competing phrases that are generically AI-sounding get suppressed.
Voice freshness tracks how recently each phrase appeared in a transcript. Phrases that haven't surfaced in recent interviews are flagged as stale — they stay in the library but lose drafting priority until you use them again.
Mechanism 04
Every action in the review queue feeds back into the system. Accept, edit, or reject a draft — Bernard records the outcome and adjusts the next month's briefs accordingly.
You record a 15-minute prompted interview. The audio is transcribed and segmented into labeled spans. Each span is a discrete, citable unit of content.
New atoms are extracted and added to the workspace graph. The drafting brief is built from: relevant atoms, your active voice phrases, channel-specific format rules, and the current transcript spans. The model drafts. Every claim cites a span.
You accept, edit, or reject drafts in the review queue. Accepted phrases are promoted in your phrase library. Rejected phrasing is demoted. Edits are analyzed for new phrase candidates. Month over month, the drafts need less correction.
Approved content publishes to Buffer, WordPress, Astro, or clean export. The cycle resets for the next interview — now with a richer atom graph, fresher phrase library, and tighter voice model than the month before.
An honest note
Provenance trace means every drafted claim links to something you said. It doesn't mean every claim is factually correct — if you say something inaccurate in the interview, the draft will faithfully reflect the inaccuracy. The system amplifies your voice, including the wrong notes. The review queue is the final check, and it's yours to make.
Bernard also doesn't guarantee the atoms graph is exhaustive. Atom extraction is LLM-assisted — it catches most of what's worth catching but isn't a verbatim record of your thinking. Think of it as a working index, not an archive.
Your data
The provenance, atoms, and voice mechanisms above describe how Bernard builds content from your words. This section describes what happens to those words.
Interview transcripts, recordings, uploaded media, and AI-generated drafts all belong to you. Bernard holds them on your behalf — it doesn't claim them, license them to others, or include them in any shared dataset. You can export or delete at any time.
Your transcripts are sent to Anthropic (Claude) solely to generate drafts, and to no other model provider. Neither Bernard nor Anthropic uses your content to train AI models. Your content leaves Bernard only to come back as a draft — nothing else.
Bernard is for clinical thought-leadership content — not patient records. Do not enter names, dates of birth, health record numbers, or any other identifying information. We are not a HIPAA-covered entity and do not offer a Business Associate Agreement. Use de-identified case examples only.
Your workspace data is kept for the duration of your subscription plus 90 days after cancellation — giving you time to export everything. After that period it is permanently deleted. To request earlier deletion or a copy of your data, email bernard@gmail.com. For the full legal text, read the Privacy Policy and Terms of Service.
Founding owners get direct access to the person who designed this. If something here raises a question, it's a good sign you're the kind of owner we want to work with.
"The trust problem in AI content isn't solvable with better prompts. It's solvable by making the source of every claim visible and checkable. That's what provenance trace is."
— Dr. Q, founder