Security & data handling

Your evidence is processed in memory, in one region, and not kept.

This page states what the service does with what you send, who can see it, and which controls exist today versus which are planned. Where a certification is not yet held, it says so.

Where a request goes

One path, three parties, nothing stored by default.

1 · your client
Project API key
TLS to the service. Your key is hashed at rest and scoped to one project.
2 · the service · GCP us-central1
In memory only
Authenticates, meters, assembles the prompt, validates the response. Content is not written to disk or logs.
3 · model provider
Inference, then discard
Sent to Anthropic under our account and key. Your client never holds a provider key.
What we keep

Metadata: request id, timestamps, token and search counts, cost, the reasoning version and model used, and the participation trace. Enough to bill and to reproduce provenance — not the content.

What leaves the region

Inference requests to the model provider, and — only when you enable discovery — web search queries about the counterpart you named. Nothing about Party A is searched.

Subprocessors: Google Cloud (hosting), Anthropic (inference and web search), Stripe (payments). [Link the subprocessor list and DPA when published.]

Account controls

Available today, and what isn't yet.

Marked plainly so a security review doesn't have to guess.

available
Project-scoped, reveal-once API keys

Shown in full exactly once; stored hashed; listed by prefix, creator, and last use; revoked instantly.

available
Five roles

Owner, Admin, Developer, Billing, Viewer. Developers see only their assigned projects; Billing never sees keys.

available
Sign-in with Google, GitHub, or email

Email requires verification before anything else. Passwords are stored as salted PBKDF2-HMAC-SHA256 hashes. Sessions are server-side and revocable.

available
Audit log of account events

Organization creation, key creation and revocation, payment method changes, and sign-ins — with actor, IP, and timestamp.

available
Card data never touches us

Cards are entered on Stripe's hosted page and stored with Stripe. We hold a payment method reference, not a number.

available
Abuse limits on new organizations

Verification before payment, a rate limit for the first 24 hours, a small pre-charge ceiling, and a $1 first charge that validates the card before compute burns.

planned
MFA enforcement, SSO, SCIM

Not available at launch. Compensating controls: allowed-email-domain invitations, one-click member removal that revokes sessions and lists that member's keys for rotation, and the audit log. [Target dates.]

planned
IP allowlists and per-project budgets

Per-project key allowlists, monthly budgets with alerts, and a hard cap. [Target dates.]

Residency & compliance

Controls now. Certifications when they're earned.

region
United States
Every organization is homed in one region and its data stays there. One US region today; an EU region is next on the roadmap, [uncommitted date].
soc 2 · gdpr · hipaa
Controls in place, reports pending
Designed against SOC 2, GDPR, and HIPAA control sets. No SOC 2 report is held yet. [Audit target date.] No BAA is offered yet.
agreements
Terms, Privacy, DPA
Acceptance is recorded with version and timestamp. A Data Processing Addendum will be acceptable by an Owner in the console. [Publish date.]
Reproducible by design

Every assessment names its reasoning revision and model. Revisions are immutable, so a reading can be attributed months later — without the content, which was never kept.

Watched continuously

Synthetic, non-sensitive canary assessments run against production on a schedule, with uptime checks and alerting. Real customer content is never sampled for validation.

For reviewers

Running a vendor review? Ask for the security overview and the subprocessor list.

Grant your reviewer the Viewer role and they can read the audit log and settings without changing anything. Report a vulnerability to [security@domain].

Request the security overview Get an API key