Trust Center

Primary application servers, databases, object storage, and backups for the deployment are located in-region. Administrative records describe regions and change approval; technical deployment pins resources to the regional project.

Public ingress uses modern TLS. Traffic to speech-to-text and LLM inference subprocessors uses TLS to the configured regional endpoints only. This is a technical control verified in architecture reviews and deployment config.

Databases, object storage, and backup objects use industry-standard algorithms (e.g. AES-256-GCM) with keys managed per our cryptography policy. WAL-G backups add client-side encryption before off-site storage.

A managed edge layer fronts public hostnames, absorbing volumetric attacks and applying WAF rules for common web abuse. Provider-level protections and tight firewall posture remain in place behind the edge.

Layered detection: the hosting platform provides network-layer IDS, while each VM runs fail2ban (SSH brute-force prevention), UFW with default-deny posture, and unattended security upgrades. Authentication and access events from the auth-proxy are shipped to a central Loki/Grafana stack so anomalous login patterns surface in dashboards and alerts. Policies assign triage ownership when alerts fire. Kernel-level runtime IDS for containers is a planned enhancement on top of this baseline.

Clinicians authenticate through Firebase (OIDC/OAuth-style flows); every API path validates identity. Session lifetime and lockout behaviour align with our access-control policy.

Clinician, patient, and admin capabilities are separated in the data model and GraphQL layer so least-privilege is the default. Role definitions and review cadence are documented administratively.

Mandatory MFA is enforced for accounts that can change security-sensitive settings or access high-risk data, via Firebase Authentication.

Changes flow through source control, review, and pinned dependencies. Security impact is considered for substantive features as an administrative gate; technical tools (Dependabot, etc.) support ongoing hygiene.

Each release carries a machine-readable inventory of dependencies for vulnerability response. SBOM export is wired into CI as part of our supply-chain assurance programme.

Audio and transcript material defaults to a short retention window with automated deletion; clinical notes follow each organisation’s retention policy. Once integrated with your record system, the authoritative clinical record remains yours.

Retention-driven deletions are recorded in deletion_log, an append-only Postgres table with no UPDATE or DELETE grants to any role. Each row links to the previous one by SHA-256 hash chain (row_hash = sha256(prev_hash ‖ canonical(metadata))), so altering or removing any earlier row breaks every subsequent hash and is detected by the in-database verifier hanah_verify_deletion_chain(). Companion audit tables (ai_audit, secret_access_audit) are also append-only. Quarterly customer-facing rollups are produced by hanah_deletion_report_quarter(year, quarter).

Individuals can request access to or correction of their personal information; Hanah documents timelines, formats (JSON + human-readable exports), and escalation paths. Tooling supports export where data lives in Hanah systems.

Security-relevant events (authentication, privilege use, configuration changes) are queryable for institutional customers and forwardable to your SIEM (for example Microsoft Sentinel) where contractually agreed. Configured per customer.

Our speech-to-text and LLM inference providers' enterprise agreements prohibit training on customer content; Hanah does not operate an internal training pipeline on your health data. Administrative vendor reviews and technical configuration (region-locked endpoints) support the claim.

Generated notes require explicit clinician review and finalisation before they are treated as part of the clinical record. This is both a product control (workflow gates) and a policy commitment in the PIA.

One harness, two evaluation layers, run on every grading pass: Word Error Rate against verbatim ground-truth transcripts (Layer A), and an LLM jury panel measuring both coverage (does the note capture each pre-registered expectation?) and hallucination (does the note contain claims the transcript does not support — i.e. content the model invented?). The full methodology is published, and live results render directly from the harness's aggregate output. The corpus continues to graduate v1 (synthetic) → v2 (role-play) → v3 (real consent-based) → v4 (joint with health-system partners).

System prompts are server-controlled and inputs are bounded by the clinical workflow (no free-form chat surface). Provider safety classifiers run on every generation; when they trigger, the stream surfaces a content-filter finish reason that is captured in the per-call ai_audit row alongside a canonical ai.content_policy error code, latency, tokens, and a Tempo trace ID. Anomalous patterns surface in dashboards over that audit table. Cross-call behavioural anomaly detection for abuse patterns is a planned enhancement on top of this baseline.

When the clinician clicks record, the product opens a consent dialog requiring explicit confirmation that consent has been obtained from the patient and any other participants in the session. Recording cannot begin until that check is acknowledged, and the confirmation is bound to the encounter alongside clinic policy referenced in the privacy impact assessment.

A named executive is accountable for information security; risks are logged with owners and target dates and reviewed on a fixed cadence. This matches what buyers typically expect from an ISO-style management system, even at our current scale.

Formal policies cover access, cryptography, incident response, vendors, and change management. The pack is board-approved and held under version control; full content is available under NDA.

Security events from Hanah-controlled systems are forwarded into your Sentinel workspace (or equivalent) for joint monitoring where contractually agreed. Configured per institutional customer.