Safety and governance
Clinical safety is a design input, not a document.
In clinical software, interface and workflow decisions are safety decisions. This page explains how Reedley approaches clinical risk, which standards apply, and exactly how far we have got.
Statements reviewed 14 November 2026. Reedley is in development; nothing here describes a live service.
The standards that apply
In England, clinical risk management for health IT is set out in two information standards published under section 250 of the Health and Social Care Act 2012:
- DCB0129 applies to the manufacturer of a health IT system — Reedley. It requires a clinical risk management file and plan, a hazard log, a clinical safety case and clinical safety case reports, approved by a Clinical Safety Officer: a suitably qualified and experienced clinician with current professional registration.
- DCB0160 applies to the deploying organisation — for example a GP practice. It covers the risks of deploying and using the system in that setting, and relies on the manufacturer’s safety case report.
NHS England has been reviewing both standards; a public consultation closed in September 2026. We will follow the outcome and update our approach and this page.
Sources: NHS England Digital, DCB0129; NHS England, national review of DCB0129 and DCB0160; NHS England Engage, the consultation (29 June to 11 September 2026).
Where we are
We do not claim DCB0129 compliance. An engineering draft of a hazard log and a safety-case outline are maintained alongside the build, with provisional severities. No Clinical Safety Officer has reviewed them, so they are not a safety case.
Before any use with real patient data, Reedley will:
- appoint a suitably qualified Clinical Safety Officer;
- establish the clinical risk management file and plan under their approval;
- review and approve the hazard log, and produce a clinical safety case report for the scope of any pilot;
- provide each deploying practice with what it needs for its own DCB0160 assessment.
How hazards are identified
Hazards are identified from the clinical workflows themselves — misidentified patients, missed or misread instructions, stale information acted on, conditional wording treated as certain, work marked complete that never happened — and from the behaviour of AI-assisted components, such as incomplete or incorrect extraction.
Each hazard records its causes, effects, proposed controls (draft, not yet approved by a Clinical Safety Officer), the tests that exercise them on synthetic data, and the residual risk. The draft log already links hazards to automated tests and evaluation cases in the product repository.
Behaviours designed and tested in the development build (not yet safety controls)
- Identity confirmation before any clinical processing, with mismatches blocking review (enforced by Reedley’s own screens in the development build).
- Every finding cited to its source passage; uncertainty and inference shown as such.
- Conditional or tentative wording never converted into an instruction.
- Human decision designed into every proposal, with draft review floors (awaiting sign-off by a Clinical Safety Officer, not yet appointed). In the development build these are enforced by Reedley’s own screens only, there is no authenticated user, and one person can still close an item on their own recorded evidence (under review).
- A person’s report moves an item to actioned; verification comes from the record where Reedley can read it; where it cannot, a named person’s structured statement is accepted today, and that rule is under review. Completion requires verification.
- Derived information marked stale when the record changes; Reedley’s screens block action on it while stale.
- No interpretation of results, no prescribing and no writing to the medication record.
Human oversight and residual risk
Human review is a control, but not a perfect one: people can over-trust a confident interface. Reedley’s design tries to reduce that risk — showing the source rather than a summary alone, avoiding numeric confidence, and making uncertainty visible — but it cannot remove it. Residual risk will be stated in the safety case for each release and pilot.
Validation
The build has a regression suite and a locked holdout set of evaluation cases on synthetic documents. One model-in-the-loop run on synthetic development documents exists (13 November 2026); its thresholds are provisional and carry no clinical weight, and the locked holdout has not been used. No clinician has yet used the build.
Validation will proceed in stages: synthetic walkthroughs with clinicians, reported separately for founder and independent users; then a governed pilot with a protocol agreed with the practice and its Clinical Safety Officer.
Incident escalation and release governance
Before any real-world use, Reedley will put in place a process for practices to report safety concerns, a named route to a Clinical Safety Officer (not yet appointed), defined response expectations and a way to pause affected functionality. Releases will be governed by that Clinical Safety Officer, with the hazard log updated for every change that affects clinical behaviour.
If you believe something on this website, or in a demonstration, could mislead a clinician, please tell us.