Ethics review

For an institutional ethics committee: what the ORTHOReD platform does with patient data, in the words of its specification. Drawn from version 2.777 of the specification (November 2026), and redrawn whenever it changes.

Each statement below is reproduced from the ORTHOReD platform specification, version 2.777 (November 2026). Section numbers (§) refer to that specification. It covers the platform only: the surgeon’s own clinical practice, research protocols and institution are the surgeon’s to describe.

Where is the data held, and who at ORTHOReD can reach it?

Surgeons in one country share a deployment; no surgeon shares a database schema. Each surgeon owns their data, held in a PostgreSQL schema of their own with a database role granted only on that schema. Any cross-surgeon collaboration would happen only through deliberate de-identification pipelines (collaborative research and cross-surgeon complication linking are Future possibilities) that strip identifying information before any data leaves the surgeon’s schema. Shared-table multi-tenancy — one set of tables filtered by a tenant_id column — is ruled out: the tenancy model is Bridge, schema-per-surgeon (§2.3). §2.1

Bridge — one deployment per country, one PostgreSQL schema per surgeon — is the only supported model for production patient data (see §2.1). The country deployment is hosted in AWS in its own account, in the region (or combination of regions) appropriate to that country’s medical privacy and data residency requirements. Region selection is settled per country; the legal framework applying to a surgeon’s practice determines which country deployment they belong to. Sharding rule: a country adds RDS instances as it approaches a few hundred surgeons — schema count per instance is the constraint, not surgeon count per country. §2.3

Under Bridge tenancy HQ operates the instance, so the question is not whether HQ could reach a tenant schema but under what grant, for how long, and visible to whom. This section is the answer, and it is the only sanctioned inbound path to clinical data. Everything here is enforced by database grant and alarm, not by policy: the controls are testable, and a surgeon may ask to see the test. §2.24

No HQ role holds a grant on any tenant schema. HQ’s operational roles are granted on the platform schema and on control-plane metadata only. A tenant schema is reachable only by the surgeon’s own database role, and by a grant created for one of the two paths below and then removed. §2.24

The routine path. The surgeon authorises a named HQ engineer to work inside their schema for a stated purpose. The grant is created on authorisation, is scoped to that schema, expires with the session, and is attributed to that individual — never to a shared account. The session is written to the surgeon’s audit trail as it happens. A surgeon may end the session at any time. §2.24

Both paths write to an append-only trail held outside the tenant schema, so a session cannot edit its own record. The surgeon can read their own trail at any time without asking HQ. The twelve-hour notice is a backstop against a surgeon not looking, not the only disclosure. §2.24

It does not claim that HQ is technically incapable of reading a clinical field — under a shared instance that claim would be false, and the earlier absolute form of it in §2.23 has been withdrawn rather than weakened. It claims that no such access is standing, none is anonymous, none is unbounded in time, and none is invisible to the surgeon it affects. §2.24

Who owns the data?

Surgeon-owned clinical data. Encounter records, generated documents, PROMS responses, complications log, surgical logbook, billing tracking, surgeon-uploaded patient resources, and all other patient-facing or surgeon-configured content within a deployment are owned by the surgeon. The surgeon is the data controller for clinical data captured through their use of the platform, with the legal responsibilities that role entails in their jurisdiction. §2.5

ORTHOReD-owned platform data. The system-level reference data (App sheet specification, DSL rules, document templates, implant master catalogue, procedure code references, anatomy and diagnosis taxonomy, system-level patient resources) and the platform code itself are owned by ORTHOReD. ORTHOReD is the data processor for the platform infrastructure. §2.5

Patient-controlled data. Patient-completed registration form responses and PROMS responses are captured directly from the patient. The patient retains the right to access, correct, and request erasure of their personal data under the privacy framework applicable to the surgeon’s jurisdiction (Australian Privacy Act, UK GDPR, US HIPAA, equivalent). The surgeon, as data controller, executes these rights through ORTHOReD’s data extraction and correction mechanisms. [8 November 2026: DISCLOSURE TO AN EQUIPMENT SUPPLIER: the rep of the company supplying the implants or equipment for a patient's operation is sent the case with the patient's name and date of birth. Australia: the disclosure rests on APP 6.2(a) — health information is sensitive, so it must be directly related to the primary purpose and within the patient's reasonable expectation; supplying the equipment for the patient's own operation is that purpose — and APP 5.2(f) requires the collection notice to name it as a usual disclosure (the entity bears the onus of showing the patient knew). The UK and the EU: lawful basis Art 9(2)(h); Art 13(1)(e) — the recipients, or categories as specific as possible. New Zealand: HIPC 2020 rule 11 (a directly related purpose) and rule 3 (the intended recipients named). The US: a HIPAA treatment disclosure (HHS FAQ 490 names an orthopaedic device rep sizing a prosthesis); no notice rule, and the sentence does no harm. One sentence in the notice, every deployment. Not legal advice; the UK wording goes to LegalVision UK with the UK privacy documents] §2.5

What consent is asked, and how is it withdrawn?

Primary clinical use of patient data — capturing the encounter, generating documents, sending PROMS, transmitting to the GP — is covered by the standard clinical care consent obtained by the surgeon at first encounter. ORTHOReD does not introduce new consent requirements for primary clinical use beyond what the surgeon already obtains for clinical practice. §2.5

Secondary uses require additional consent or a lawful basis distinct from clinical care: §2.5

· Research dataset. Aggregation of de-identified or pseudonymised clinical data for the surgeon’s own research (§4.3 Research) is permitted under the surgeon’s clinical practice consent in most jurisdictions provided data does not leave the surgeon’s deployment. Where data is to be used for research that crosses the deployment boundary (a multicentre study — run by hand from the surgeon’s own export; pooled collaborative research is a Future possibility), participating patients must be re-consented or covered by ethics-approved waiver. Specific consent mechanism is the surgeon’s responsibility; ORTHOReD provides the data scoping infrastructure. §2.5

· Cross-surgeon aggregation. Any feature that aggregates data across surgeon deployments (cross-surgeon complication linking, registry feeds, data licensing) requires explicit patient consent recorded for the patient. ORTHOReD captures this directly via the registration form (§4.4); where a patient has not completed it, the surgeon obtains consent through their existing clinical care consent process and transcribes the result onto the patient record. Either way, the surgeon must obtain and confirm consent before such features ingest the patient’s data. §2.5

· AI training. Patient data is not used to train ORTHOReD-internal AI models or third-party AI systems. ORTHOReD does not extract patient data from any deployment for AI training purposes. If a future feature is written that would use patient data for AI training, it requires explicit patient consent captured per patient (via the registration form, or surgeon-transcribed) and re-consent for existing patients — there is no implied consent route. §2.5

Consent capture infrastructure. The patient registration form (§4.4, Phase 1) is the active consent capture mechanism for data-collection consent — a structured, auditable consent record captured from the patient at first encounter covering inclusion in the surgeon’s research dataset and any cross-surgeon aggregation features the surgeon offers. For a patient who has not completed the form, consent is obtained through the surgeon’s existing clinical care consent process and the result is transcribed onto the patient record; the registration form writes consent_state (§5.22) directly. §2.5

The platform’s architectural commitment is that secondary uses of patient data are opt-in and explicit. Default behaviour is no secondary use beyond the surgeon’s primary clinical use. §2.5

Consent state is captured as a structured field set on the patient record (one set per patient, not per encounter). The consent dimensions are drawn from a catalogue (§5.22). A platform baseline ships fixed and mandatory: research within the surgeon’s practice (de-identified clinical data and images), research shared beyond the surgeon’s environment (multicentre and registries; cross-surgeon complication linkage is a Future possibility), de-identified images for teaching and education, commercial data licensing, and AI model training. A surgeon may add further consent dimensions of their own through the registration-form configuration (Settings > Surgeon resources > Registration form), each paired with the surgeon’s attestation that they hold or will obtain ethics approval for that use; the baseline cannot be removed, only added to. Granularity follows the ethics floor: the general-research dimensions may be broad because each actual use is separately gated by HREC approval or waiver, while commercial licensing, AI model training, images, and any disclosure that leaves Australia must remain their own explicit dimensions. Each dimension carries: state (granted / declined / withdrawn / pending), date of state change, method of capture (registration_form / in_person / staff_recorded), and the user_id who recorded the state. Two consents sit on separate axes and are not part of this set: the recording of a consultation is consented per-encounter (re-asked each visit, in the Dictate panel; a surgeon dictating alone records no conversation and is not asked), and PROMS delivery is a separate opt-out that hard-gates the PROMS send pipeline (§3.4.8). The field scaffolding and the registration form (the patient-facing capture mechanism) both ship at Phase 1. §2.5

Default state at Phase 1. New patient records default to “pending” for every consent dimension. Secondary use is not permitted until consent is actively recorded. Where the surgeon’s existing clinical care consent process has covered any of these dimensions, the surgeon records this per-patient in the patient record (state = granted, method = staff_recorded, with the surgeon as the recording user). The platform does not assume consent on the surgeon’s behalf; an explicit per-patient record by the surgeon is required. §2.5

Withdrawal mechanism. Consent withdrawal is recorded by any authenticated staff user — the surgeon or any staff role — on patient request (phone, in-person, or written communication relayed to the practice). The user accesses the patient’s consent state in the patient record and updates the relevant dimension(s) to “withdrawn”, with date, reason, and the acting user_id captured automatically. This follows §3.7.1’s access model — the platform does not restrict who may act and the audit trail is the control, so every withdrawal is attributed to the user who recorded it. In practice the admin takes most of these calls (§3.7.6). The patient does not record withdrawal directly: the patient portal that would have allowed it was withdrawn on 10 October 2026, so a withdrawal, like any consent change after registration, is recorded by the practice. PROMS opt-out is a separate axis (§3.4.8) — withdrawing research consent does not stop PROMS, and opting out of PROMS does not withdraw research consent. §2.5

Operational effect of withdrawal. Withdrawal is forward-looking only. Future research exports exclude the patient from the moment withdrawal is recorded. Existing exports already sent to licensees or research collaborators cannot be recovered — by then the data has typically been de-identified, aggregated, and incorporated into recipient datasets. The platform’s audit trail of past use is preserved unchanged (the audit_log and documents archive are immutable). The surgeon’s primary clinical use of the patient’s data is not affected by withdrawal of secondary-use consent — the surgeon continues to provide clinical care under the standard clinical care consent. §2.5

Form 1 also carries the consent capture for secondary use of patient data. Each consent dimension (§2.5) is presented as a clear yes/no choice with brief explanatory text: §4.4.2

· Research in your surgeon’s practice — I agree that my de-identified health information and images may be used for research and quality improvement carried out within my surgeon’s own practice. §4.4.2

· Research shared beyond your surgeon — I agree that my de-identified health information may be shared with and used in collaborative research across multiple surgeons, centres or clinical registries, including where this means my de-identified data leaves my surgeon’s own secure system (§4.3). §4.4.2

· Images for teaching and education — I agree that my de-identified images may be stored and used for medical teaching and education. §4.4.2

· Commercial data licensing (shown only if the surgeon enables it) — I agree that my de-identified health information may be provided to third parties such as industry or registries under a data-licensing arrangement. §4.4.2

· AI model training (shown only if enabled) — I agree that my de-identified health information may be used to train ORTHOReD’s software models. Every consent defaults to declined; the patient ticks to grant, or taps Yes to all three. §4.4.2

Can the data be taken out, and what happens at the end of a licence?

End-of-licence extraction. Termination of an ORTHOReD licence does not destroy the surgeon’s data. ORTHOReD provides a bulk extraction at termination as a contractual entitlement. From termination the surgeon’s open patient resources stop resolving at their canonical and versioned URLs, in either path below; the archived copies stay private in the record and can be produced on request. The surgeon then chooses one of two retention paths to satisfy statutory retention obligations applying in their jurisdiction: The documents the surgeon dropped into Set up with AI and kept are part of his data: they are held until he deletes them, with no expiry, and at termination they go into the bulk extraction with the rest of it and then follow the retention path he chooses. The practice’s general notes — patient_note rows with no patient — are the surgeon’s data too: kept for the life of the account, never deleted, and in the bulk extraction at termination; a note on a patient goes with that patient’s record. §2.5

Option A — Archive Mode (continued ORTHOReD hosting at reduced cost). The deployment converts to read-only operation: no new encounters, no PROMS scheduling, no document generation. A minimal compliance interface remains active, allowing the surgeon (or their estate) to respond to patient access requests, subpoenas, and regulatory inquiries via the existing per-patient extraction mechanism. The surgeon (or estate) pays a reduced annual fee covering AWS infrastructure plus minimal operational overhead, for the statutory retention period applying to their jurisdiction. Pricing is published in the licence agreement and may be paid annually or as a lump sum at retirement. At end of retention period the deployment is decommissioned and a certificate of destruction issued. §2.5

Option B — Self-managed archive (surgeon retains the bulk extraction). The surgeon takes the bulk extraction (structured JSON exports plus the immutable document archive) and assumes responsibility for retention themselves. The deployment is decommissioned at termination per the data retention schedule applying to their jurisdiction. The surgeon is responsible for confirming retention obligations are met from their extracted copy, for responding to patient access requests and subpoenas from their archive, and for ensuring continued availability of the records to authorised parties for the statutory period. §2.5

The surgeon chooses between Option A and Option B at termination. The choice is documented in the termination paperwork. Choosing Option B and later wishing to transfer to Option A is supported as long as the surgeon still holds a complete copy of the extracted archive that can be loaded back into a fresh deployment; the operational mechanism is the same as the original deployment provisioning. §2.5

Patient access requests. Patients with a legal right of access to their own data (under their applicable privacy framework) submit the request to the surgeon. The surgeon executes the request via the per-patient extraction mode. The platform produces the bundle; the surgeon is the responsible legal counterparty. The bundle includes the notes the practice wrote on the patient — the patient_note rows: they form part of the patient’s record and are produced with it. §2.5

Extraction is read-only. All extraction operations are read-only against the deployment. They produce copies; they do not move or delete data from the deployment. §2.5

How is it protected in storage and in transit?

At rest. All deployment data is encrypted at rest. PostgreSQL data is encrypted using AWS RDS encryption (AES-256). S3 buckets use server-side encryption with AWS-managed keys (SSE-S3) at minimum, or customer-managed keys (SSE-KMS) for jurisdictions requiring documented key management. §2.5

In transit. All communication between client devices, application servers, database, S3, and external services is over TLS 1.2 or higher. No unencrypted data flow at any architectural boundary. §2.5

How is AI used, and what happens to the data it reads?

Where AI runs, and what happens to your patients’ data. Every AI function in ORTHOReD runs on Amazon Bedrock in Australia — AWS’s Sydney and Melbourne regions only — called from the same AWS account that holds your records, so your patients’ data reaches no one who does not already hold it. The models are Anthropic’s Claude; AWS runs them, and Anthropic cannot reach them or see what is sent. Nothing is kept: AWS does not store what is sent or what comes back, does not share it, and does not use it to train any model. Four account settings make that true — nothing retained, nothing sent is logged, AWS’s AI services opted out of keeping content to improve their service, and no AI store of any kind outside your own data — and every ORTHOReD release checks all four and cannot go out while one fails. Bedrock, and Amazon Transcribe, which turns dictation into text, are assessed at PROTECTED under the Australian Government’s IRAP. The model learns nothing from you: ORTHOReD gets better at reading your documents by looking up your own confirmed entries, kept with your data. AI writes and you confirm: it fills only fields with a fixed set of answers, from what a document or your dictation states, never guesses, and never works out a diagnosis or a treatment that the document does not state; nothing it writes is in the record until a person confirms it — which is also why none of it is a medical device. In a medicolegal document completion, an answer of fact is your record’s own value, word for word, with no AI wording, and your commonly used phrases are your own text, which you pick; only where a question asks for your opinion is a sentence drafted for you, from your own confirmed record, your own commonly used phrases and the request alone — never another party’s document and never the internet — and you edit and confirm each one; the response states which answers were drafted, by which program and from what. On an IME or a file review nothing is drafted. The help assistant reads ORTHOReD’s guide and your settings, and can count your own records; it never opens a patient’s record. Recording a consultation asks the patient’s consent each visit; dictating on your own does not. Use AI features, in Settings ▸ Identity, switches all of it off, and ORTHOReD then works the same way by hand. §3.8.8

What extraction fills in, and what it leaves for you. A list is read item by item whether or not it has commas — “Hypertension NIDDM Hypothyroidism” is three — and common abbreviations are read as they are used in practice. When ORTHOReD reads an uploaded report, a letter or an operation note into an encounter, it writes values only for what the document actually states, and only into fields that have a fixed set of answers; every unconfirmed value is marked as written until you confirm it on the step it belongs to, and nothing reaches the record before that. It will not guess: a hedged finding (“possible”, “query”, “cannot exclude”), a range, a classification the report describes but does not name, a measurement given in words, a comparison with an earlier study, the referrer’s plan or recommendation, and anything about the other side or another region are left empty, with a line beside the unconfirmed values saying what the document said and why nothing was filled. A diagnosis that a referral letter states is written, for you to confirm, as a report’s findings are. A list is read item by item — “hypothyroidism, asthma, COPD” writes each one your lists hold — and an answer that names two things, such as Asthma / COPD, is written when the document states either. An explicit negative — “no subscapularis tear” — is recorded as intact; a structure the report does not mention stays blank. A value you have already confirmed is never overwritten. A missed value costs you one click; these rules are there so a wrong one never costs you a record. Dictation works the same way: press Dictate in the encounter bar and say it, and what you state is written for you to confirm. On an operation, pick a common surgery or tick a procedure first — Dictate waits until you have — and dictation fills in the rest. On a consultation it writes the diagnosis and the plan you state into their fields as well — only what you said; it never works out a diagnosis or a treatment you did not state, and nothing written counts until you confirm the section. The common diagnosis, the primary diagnosis and the management template stay taps. It never picks a common surgery, a technique, a common diagnosis or a management template for you. Recording a consultation asks the patient’s consent first; dictating on your own does not. If one of several documents read together cannot be read, the others are still read: the step names the one that could not be read and why, and nothing is taken from it. §3.8.8

The provider model is a fixed artifact. It has never seen ORTHOReD’s private taxonomy and it does not learn it: no weights change and no memory persists between calls. Its native contribution is general clinical fluency — it already understands what “full-thickness supraspinatus tear” or “Goutallier 3” mean. The ORTHOReD-specific mapping is supplied in each call, as the schema slice, and supplied again on the next call. Anything the model is to “know” about ORTHOReD is given in that call, not retained. This is consistent with the AI-training position (§2.5): no patient data trains any model. §3.8.8

Extraction becomes more accurate with use by retrieval, not by training the model. Each confirmed extraction is stored within the surgeon’s deployment as a row holding a phrasing fragment, the field_id it resolved to, and the confirmed value. On a later extraction the incoming text is embedded (pgvector, in-instance) and the nearest prior confirmed examples are retrieved into the prompt alongside the schema slice, biasing the model toward how this surgeon’s phrasings have previously resolved. More confirmed examples produce closer retrieval and fewer corrections; the model is unchanged throughout. This store is Category C overlay data, surgeon_id-scoped, pgvector-indexed in-instance, and never leaves the deployment — no new external service. The retrieval layer is an enhancement: the bare loop (model, schema slice, validation gate, confirmation) functions without it at the schema-only floor, and may ship first. §3.8.8

What record is kept of who saw and changed what?

As per §2.4, the audit log is held locally in the surgeon’s deployment, not centrally. ORTHOReD has no central record of who accessed which patient’s data — that information exists only in the surgeon’s deployment under the surgeon’s control. §2.5

Surgeon visibility. The surgeon can view the audit log for any patient in their deployment from the patient record. The admin’s access events appear in the log with admin identification. §2.5

External access requests. Court orders, subpoenas, and regulatory access requests are responded to by the surgeon as the data controller, supplied from the per-patient extraction described above. ORTHOReD does not respond to such requests directly except as compelled by jurisdiction-applicable law and only after notifying the surgeon (where notification is itself lawful). §2.5

As per §2.4, the documents archive is held locally in the surgeon’s deployment, not centrally. The full medicolegal record of every document generated for every patient lives in the deployment alongside the documents themselves. §2.5

The structured clinical data underlying an encounter is mutable. Every generated document is immutable, in every format it was produced in — the PDF, the Word file, the plain text. When the surgeon corrects underlying data after a document has been generated, a new document is produced and the original is preserved unaltered. This is the standard medical-record correction pattern — corrections are transparent and additive, never destructive. §2.10

Together, the archived document — every format it was produced in — and its archived linked resources constitute a complete, immutable record of exactly what the patient received at every encounter. §2.10

The structured field values captured through the encounter pipeline (the source of truth for document generation) are mutable after encounter finalisation. The surgeon may correct mistakes at any time. Every post-finalisation edit must be logged. §2.10

Who can sign in, and how?

Surgeon — full access to all sections, including user management. §2.14

staff (admin, nurse, allied health, other) — the same access as the surgeon across all sections, with one exception enforced at the API and database level: user management (creating users, assigning or changing roles, §2.22) is surgeon-only. The admin can read and write clinical data, generate documents, manage billing tracking, run PROMS operations, edit patient records, and manage resources where the surgeon delegates. There is no closed write list. The control on this access is the audit trail, not a permission matrix: every write records the acting user_id and is logged immutably (§2.10), so every action is attributable. Distinct from access, ATTESTATION, which is the surgeon’s in one place only: the answers of a medicolegal / insurance response (gate 2, §3.9.5) — only the surgeon’s confirmation sets them; a staff user can enter the answers and a staff user’s Save on a response stores them unconfirmed. Every other confirmation — of a value from the patient, a document or dictation, on any encounter (§4.4.6) — is any user’s, and the audit row names who. Access is shared; attestation of a medicolegal answer is the surgeon’s. Full specification in §3.7. §2.14

Patient — PROMS questionnaire page and registration form page. No login. UUID token authentication. Completely isolated from the authenticated application. §2.14

RBAC is implemented at the API and database query level — not only the UI level. The one role distinction the API must enforce is that a admin cannot create users or change roles (§2.22). Beyond that, the surgeon and admin share the same data access; accountability is provided by the audit trail rather than by restricting what the admin may retrieve or write. §2.14

The patient-facing PROMS page and registration form page require no login, authenticate via UUID token only, and must be completely isolated from the authenticated application. A valid token grants access to exactly one questionnaire or one form — nothing else. Verify this isolation explicitly in testing. §2.14

How is data de-identified before it leaves the surgeon’s system?

Cross-surgeon collaboration is exclusively via de-identification. The only mechanism by which any data leaves a surgeon’s deployment is the deliberate de-identification pipelines (§2.5) operating on extracted, de-identified data — never via shared configuration or central aggregation. The surgeon_id column identifies the surgeon within the country deployment (for licensing, deployment metadata, telemetry, and auditing the master content distribution channel). It is a scoping discipline, not the isolation boundary — the schema and its database role are. §2.4

When a surgeon exports data for a multicentre study, it is de-identified before it leaves their instance. The de-identification rules are specified architecturally in §2.5 and in §2.19 (free text de-identification). The runtime application of those rules: §4.3.3

· Removed entirely: first name, last name, DOB, postcode, Medicare number, phone, email, all free text fields. §4.3.3

· Pseudonymised: patient_id → a random code PT-XXXXXXXX, consistent within one export and independent across exports. surgeon_id → SRG-XXX. §4.3.3

· Transformed: absolute dates → days from surgery date. §4.3.3

· Retained: the clinical field_ids exported, all PROM scores, implant data, procedure data, sex, comorbidities. §4.3.3

Patient pseudonymisation uses a keyed hash function with study_id as the key. The same patient in two different studies gets two different pseudonyms — linkage across studies is not possible from the exported data. §4.3.5

Committed, not yet in place

ORTHOReD’s specification commits the platform to each statement below; it is not yet in force. Each returns to its place above when it is.

Where is the data held, and who at ORTHOReD can reach it?

The narrow exception, for an incident that will damage data or availability if it waits for an authorisation that cannot be obtained. It is unilateral, and every property below is what keeps it narrow. §2.24

· Granted, not held. Break-glass is a separate database role that must be granted at the moment of use. Nobody carries it. §2.24

· Individually attributed. The grant names one person. Shared or service credentials cannot be used for it. §2.24

· Auto-expiring at twelve hours. The grant is revoked automatically twelve hours after it is created, whether or not the work is finished. §2.24

· Not extendable. If the incident outlasts the window, a new grant is a new break-glass event — separately alarmed, separately attributed, separately notified. There is no renewal, because a renewable window is an indefinite one. §2.24

· Alarmed on grant. Creating the role pages HQ and writes the event immediately; the alarm does not wait for the session to end. §2.24

· Notified within twelve hours. The surgeon is told within twelve hours of the grant — measured from the grant, not from its expiry — so notice may reach them while the session is still open, which is intended. The notice names the person, the time, the stated reason, and what was accessed. §2.24

Can the data be taken out, and what happens at the end of a licence?

Routine extraction. The surgeon can extract their own data from their deployment at any time. Two extraction modes are supported: §2.5

· Per-patient extraction — all data on a specific patient, in a structured machine-readable format (JSON or CSV) plus the bundle of immutable archived documents, in the formats they were made in, from the medicolegal archive. Used when transferring care or responding to a patient access request. §2.5

· Bulk extraction — the surgeon’s full data set (all encounters, all documents, all PROMS, all logbook entries, all configuration). Used at end-of-licence, end-of-career, or for the surgeon’s own backup. Format: structured archive containing JSON exports plus the S3 archive of immutable documents, each in the formats it was made in. §2.5

What record is kept of who saw and changed what?

Every read of a patient record, every generation of a document, every export of data, and every administrative action is logged. §2.5

Who can sign in, and how?

Surgeon and admin authentication uses AWS Cognito as the identity provider, with MFA mandatory for every clinical-staff user account. The model is designed to satisfy the regulatory baseline for systems holding identifiable health data (Australian Privacy Act / OAIC, UK NHS DSP Toolkit, EU GDPR Art. 32, US HIPAA) while keeping per-day login friction close to passwordless through long-lived trusted-device sessions. Federation to Google Workspace or Microsoft 365 is specified but not offered at launch. Patient access is unaffected and continues to use the magic link mechanism specified at §2.15. §2.22

Per-surgeon user pool. Each surgeon’s tenancy (§2.3) gets its own Cognito user pool. The pool authenticates only users who hold accounts in that surgeon’s practice — the surgeon, the surgeon’s admin(s), and any nurse. Cross-tenant authentication does not exist; a user authenticated for one surgeon has no session validity for another. The per-account user-pool quota is checked before a pool per surgeon is committed to. §2.22

MFA is mandatory for every surgeon and admin account. Account creation is not complete until MFA is enrolled. The user cannot reach the application without a second factor. This holds for every user account — the surgeon and every staff role. §2.22

Generated from version 2.777 of the specification (November 2026). References to the specification’s internal decision log are left out; nothing else is changed.