Cloudflare Backup Evidence

Evidence that Maelstrom AI's entire Cloudflare estate is backed up by our own production instance of downpipes, our no-custody backup and disaster-recovery platform (we dogfood our own product).

Public

Status: production. Maelstrom AI runs downpipes against its own Cloudflare estate as a customer of its own product. This evidence reflects deployed infrastructure and our self-hosted configuration. Provii is not yet serving end-user production traffic, so Provii’s end-user operational metrics are reported separately.

Cloudflare Backup Evidence

Control(s): UC-062 (Backup and Recovery), UC-168 (Data Backup Procedures) Standards: ISO 27001:2022 A.8.13, CSA CCM BCR-02, BCR-07, GDPR Article 32 Status: ✅ IMPLEMENTED Mechanism: Self-hosted downpipes (Maelstrom’s own production instance)


Executive Summary

Maelstrom AI backs up its entire Cloudflare estate using its own production instance of downpipes, the no-custody backup and disaster-recovery platform we build and sell. We run it exactly as a customer would: the downpipes engine is deployed inside our own Cloudflare account, reads our data on a schedule, seals every archive with post-quantum hybrid encryption under keys we hold as the operator, and fans the archives out to destinations we control.

We dogfood downpipes deliberately. The same mechanism we ask customers to trust for their Cloudflare data resilience is the mechanism protecting ours, and the evidence below is generated by the product itself.

This approach provides:

  • Scheduled, verified backups of Workers KV, D1, R2, Secrets Store, and our zone and account configuration surfaces
  • 3-2-1 fan-out — each source written to two or more independent, customer-controlled destinations
  • Continuously proven recoverability — hourly canary integrity flights and scheduled in-account restore drills
  • Vendor-independent recovery — the MIT-licensed downpipe offline reader restores from archive bytes plus our key alone, with no network and no vendor in the loop

What Is Backed Up

downpipes captures the full Cloudflare data layer that Maelstrom AI depends on:

SourceContents
Workers KVNamespaced operational state and configuration
D1Relational data
R2Object storage
Secrets StoreSecret references and metadata (values remain sealed)
Zone & account configurationRouting, DNS, WAF, bindings, and related account-level configuration surfaces

Architecture (No-Custody, Self-Hosted)

Because downpipes is no-custody, running it on our own estate does not create a second custodian:

  • The engine is a routeless Cloudflare Worker running inside Maelstrom’s own Cloudflare account. It has no public route and no inbound path.
  • Archives are sealed with post-quantum hybrid encryption under keys generated and held by Maelstrom as the operator. Keys were created through the guided in-browser key ceremony and are stored offline.
  • Archives are written to Maelstrom-controlled destinations with 3-2-1 fan-out. No archive is readable by anyone without the operator-held keys.
  • The vendor-side control-plane mints a licence token and receives a content-free advisory beacon only; it holds no keys, no data, and no Cloudflare tokens, and is fail-open (it never gates backup or restore).

Recovery Assurance

  • Hourly canary integrity flights continuously confirm that archives are well-formed and decryptable.
  • Scheduled in-account restore drills prove that a real restore succeeds, not merely that a backup was written.
  • Signed reports from the engine evidence restore-test recency and run success.
  • Offline reader (downpipe, MIT) provides a last-resort recovery path: it reconstructs data from archive bytes plus the operator key, built from a frozen public archive-format specification, with no dependency on Maelstrom or on Cloudflare being reachable.

Mapped Controls

ControlHow downpipes satisfies it
ISO 27001 A.8.13 (Information Backup)Scheduled, encrypted, regularly tested backups with proven restoration
CSA CCM BCR-02 (Business Continuity)Documented, automated recovery capability across the Cloudflare data layer
CSA CCM BCR-07 (Equipment / Data Resilience)3-2-1 fan-out to independent destinations; vendor-independent offline recovery
GDPR Article 32 (Security of Processing)Resilience and the ability to restore availability and access to data in a timely manner; encryption at rest under operator-held keys

Document Information

  • Last Updated. 2026-06-19
  • Owner. ISMS Owner
  • Review Frequency. Quarterly
  • Classification. Public