We're pre-seed. We don't yet have a finalized SOC 2 report — but we've architected the substrate so that every decision is logged, attributable, and reproducible from day one. That foundation is what makes Type I attestable in 2026 and Type II by 2027. This page is the honest current posture.
Security at Allometry isn't a checklist — it's the architecture. We chose primitives that make the right thing the default, and the wrong thing impossible.
Allometry reads broadly across your stack to compute decisions, but writes are scoped, gated by policy, and always audit-trailed. No blind automation against systems of record.
— scoped service accountsAccount-level isolation, per-tenant key management, and field-level redaction at the inference boundary. Your customer data stays your customer data — even from us.
— BYOK · field redactionEvery output ties back to a model card, a policy version, a feature snapshot, and the inputs available at decision time. Audit isn't bolted on — it's the substrate.
— immutable decision logPre-seed, pre-revenue-scale companies don't have multiple active certifications. We don't either. What we have: a substrate architected for them, and a calendar of when each one lands.
Design partners get bilateral MSAs with security controls explicitly listed. ISO 42001 (AI management system), HIPAA BAA, and FedRAMP are not on the 2026 roadmap — we will not claim them until they are. Honest answer to "are you SOC 2?" today: Type I in progress · Type II in 2027.
The technical controls the substrate is built to satisfy — the architecture that makes a clean Type I audit possible. Not third-party-attested yet. Honest engineering, in flight.
Allometry runs autonomous workflows against revenue and capital decisions. That means model risk is operational risk — and we manage it the way you manage any other production system.
The decision log records the model version, the policy version that gated it, the input snapshot, and the operator account responsible. Reproducibility on demand.
New models start in shadow mode. Promotion requires evaluation thresholds. Production scopes are bounded by policy — and revocable in one click.
When policy flags a decision — low confidence, novel pattern, threshold breach — it lands in a human queue with the model card, the inputs, and the recommended action.
Models run against frozen evaluation sets every release. Drift detection on inputs and outputs. Customers can see eval results for the models running on their tenant.
What we have today, what we're working on, what's coming. We'll add to this list as documents are signed and dated — not before. Request the latest pack under MNDA via contact@allometry.com.
We meet your auditors where they are. Send your standard questionnaire — CSA CAIQ, SIG, custom — and we'll respond within five business days.