Atomic AoL logo Atomic AoL Accreditation & AssuranceOfLearning · a project by Marino&Carli
Security & auditability

Built for the most sceptical
reader in the room.

An accreditation report is only as good as its audit trail. Atomic AoL is engineered so that a peer-review team — or your own internal audit — can verify every claim independently.

Integrity

Every bundle is cryptographically sealed

The submission bundle is three files: the typeset REPORT.PDF, the machine-readable EVIDENCE.JSONL with every data point behind every sentence, and a MANIFEST with SHA-256 hashes chaining them together. Re-compute the hashes and you prove that nothing — not one byte — was altered after signing. The platform refuses to sign a bundle whose PDF is missing or empty.

SHA-256 manifest Re-derivable hashes Evidence-to-sentence trail Refuses incomplete bundles
Isolation

Your institution is a wall, not a label

Server-side tenant stamping

Every database row carries your tenant ID, written by the server — never by the AI. Even if a model hallucinated another institution's ID, the platform overrides it before anything is stored. This anti-poisoning rule is enforced by automated regression tests.

Scoped queries everywhere

Agents can only read rows belonging to the tenant they are working for. There is no "global" query path in the analytic pipeline. Tenant isolation is a structural property, not a configuration option.

AI & your data

What the AI sees — and what it never does

No training on your data

Language processing runs on Anthropic's commercial API. Under its commercial terms, customer data is not used to train models.

Privacy floor for small cohorts

Survey rows with fewer than 5 respondents are dropped by design — small-cohort data that could identify individuals never enters the report.

Budget caps built in

Per-run and per-day cost quotas are enforced in the platform itself. A runaway process stops at the cap — it cannot silently burn budget.

Human control

Approval gates, with names attached

The decisions that define your report — competency goals above all — stop at a human approval queue. Each approval records who approved, when, and what exactly was approved. The AI cannot bypass the queue: the pipeline blocks until a person decides.

LayerControl
Application accessHTTP authentication on every operator page; public pages are read-only marketing
TransportTLS termination at the edge (automatic certificates)
HostingEU region available (Frankfurt) · container runs as non-root user
SecretsAPI keys live in the host's secret store — never in code, never in the repository
Quality gates25,500+ automated tests run before every release; the suite must be 100% green

Questions your IT team will ask?

We're happy to walk your information-security people through the architecture, the data flows and the audit trail — before any data moves.

Book a security walkthrough → Book a call →