Security
This page is the short version of our threat model, the architecture, and what’s literally true about how Carnation Spark handles meeting audio and transcripts.
The three tiers
| Property | Public cloud | Private cloud | Fully segregated |
|---|---|---|---|
| Inference runs on | Shared LLM APIs | Dedicated Canadian | Customer VPC / on-prem |
| Raw audio retained for | <=24h, then purged | <=24h, then purged | Per customer policy |
| Transcripts retained for | 90 days default | Configurable | Customer-controlled |
| Encryption at rest | AES-256 | AES-256 | AES-256 |
| Data crosses border | Yes (US LLM APIs) | No (Canada) | No |
| Suitable for | Internal meetings | Regulated industries | Classified / privileged |
Deletion protocol
Raw audio is purged within 24 hours of upload. The deletion runs as a scheduled job, logged, and the deletion timestamp is recorded against each meeting record.
$ audio-reaper.status
last-run: 2026-04-20T04:00:00Z
deleted: 412 files
oldest-surviving: 2026-04-19T04:00:02Z
If a meeting is under legal hold, the customer explicitly extends retention via the admin console. We don’t preserve audio silently.
Encryption
- In transit: TLS 1.3, HSTS preloaded
- At rest: AES-256 (storage layer default)
- Key management: per-tier. Public = shared KMS. Private = dedicated KMS. Fully segregated = customer-managed HSM.
What leaks, and where
Public cloud tier sends transcription audio to a commercial LLM provider for inference. That provider, per their current contracts, does not train on the audio, but their word is the trust boundary. Private and segregated tiers do not leak at all.
Architecture (text form)
+-------------+
Your meeting --->| Ingest |
| (TLS 1.3) |
+------+------+
|
+------v------+
| Tier |
| router |
+--+--+--+----+
public | private | segregated |
| | |
+-----v----+ +---v----+ +-----v-----+
| Commodity| |Canadian| | Customer |
| LLM API | | infra | | VPC |
+----------+ +--------+ +-----------+
| | |
+----------+-------------+
|
+----------v----------+
| Encrypted storage |
| (AES-256 at rest) |
| 24h audio purge job |
+---------------------+
Frequently asked
Can you see our meeting audio? On private and segregated tiers: no, not without explicit access grant for debugging. On public tier: transient access during inference, purged within 24h.
What happens if you get hacked? Audio older than 24h is gone. Transcripts are AES-256 at rest, exfiltration would yield encrypted blobs. Keys per-tier, so blast radius is contained to one tier.
How do we audit you? SOC2 Type II roadmap: Q3 2026 audit start, Q1 2027 report. Customer audits and penetration tests welcomed before then, contact for NDA-gated architecture review.
Data residency? Private cloud tier runs on Canadian infrastructure (details provided under NDA). Segregated tier runs wherever the customer provisions.