Security & data handling

Replayability — what "proof, not custody" actually buys you

When Echo Pathways raises a recommendation, it records the rule that fired, the inputs that rule was given, a source_sha256 of exactly what it read, and where that came from — source_system_id, source_version_id, retrieved_at, mapping_version.

That combination means an auditor can take any past decision, re-fetch the original result from your record system, re-hash it, and confirm byte-for-byte that it is what we saw — then re-run the rule and confirm it reaches the same conclusion. The decision is reconstructible without us having warehoused your patients' results.

It also means a disputed decision has an answer. "Why did this fire?" is not a matter of interpretation.

The two cases where we retain the artefact

The argument above depends on there being an upstream record of truth, and on a hash being a sufficient account of what happened. Two cases break that, and there we retain the document — encrypted, access-controlled, with a retention window agreed with the service:

• Documents we receive directly. An upload, or an HL7 feed where we are the endpoint. There is no upstream custodian to point back to; if we discarded it, the evidence would not exist anywhere.
• Anything a model read. Where a value was extracted from an image or scan by a model, a hash cannot tell a reviewer whether the extraction was right. Only the original image can. Keeping it is what makes the extraction auditable — and therefore challengeable.

Retention outside these two cases is a defect in our system, and we treat it as one.

Access control

• Row-level security on every table. Not a filter in application code that could be forgotten — a database-enforced policy, so a query simply cannot return a row the user is not entitled to.
• No key in the client. Secrets are server-side only. The public anonymous key is the only key that ever reaches a browser.
• Least privilege, and logged. Access is granted by role, and access events are recorded.
• Encryption in transit and at rest, including retained artefacts.

The AI boundary

A generative model is used for constrained perception only — reading a value off a document, schema-constrained and source-linked, with no inference and no arithmetic. It never harmonises a unit, decides an event, computes a schedule, or selects an action; those are deterministic rule execution. Anything the model cannot read confidently is routed to a human, never guessed.

Models are version-pinned, so replaying a past decision uses the model that made it. No patient data is ever used to train a model — see the privacy policy, which sets out that limit in full.

Built as a regulated device

Echo Pathways is developed under a medical-device quality system, and its evidence — traceability matrix, verification and validation results, lifecycle records, anomaly log, residual-risk view — is emitted by the build rather than written up afterwards. That is a deliberate choice: evidence that is hand-authored after the fact tells you what someone believed, not what the system did.

Its current regulatory status is set out on the intended use page, and it is pre-market.

Responsible disclosure

If you find a vulnerability, report it to security@echo-pathways.com. We will acknowledge within 2 business days. We will not pursue legal action against researchers acting in good faith who give us a reasonable opportunity to fix an issue before disclosing it. Please do not access, modify or exfiltrate data that is not yours in the course of testing.