Why Conclave

The one-liner

Conclave is private transcription that can't read your meetings, never names a voice without permission, and has verifiable redaction.

Every meeting notetaker keeps a fingerprint of your voice, and nobody asked you. So in the meetings that matter most, people keep the recorder off and lose the record instead. Conclave removes that trade: transcription runs inside confidential hardware the operator can't read, a name is attached to a voice only with that person's consent, and regulated data is stripped before it is ever stored in the clear.

We are private transcription, not a notetaker. "Private" is the word we own.

The four things it does

• Operator-blind. Transcription and enrichment run inside an Intel TDX enclave; the read path is pure SQL + local embeddings, so the operator can't casually read your data. The sealing key is born inside the enclave and bound into a hardware attestation quote — cryptographic evidence a record was produced in a genuine sealed enclave. (What that does and doesn't prove is spelled out in the trust model.)
• Consent-gated naming. A voice is named only when that person agrees. The identity layer is VFTEE, with a real consent plane (claim / pause / be forgotten).
• Delete with proof. Erasure produces an offline-verifiable Ed25519 receipt.
• Verifiable redaction. Detects and strips PCI/PHI/PII inside the enclave and signs a completeness receipt anyone can verify offline (v1 live). Bring your own policy and the receipt binds its hash, so an auditor confirms your rules ran over this transcript.

The moat

A trust boundary is an architecture: logins decide who gets in, but only encryption and a measured enclave decide what the vendor can read.

Where it's going

Start where confidentiality is mandatory and regulated (legal, health, finance), then expand to every team handling data it cannot afford to expose. Cross-meeting memory and recall are capabilities built on the same operator-blind substrate, not the headline.