Safety and governance
AI is how we build it. It is not what we sell.
Reedley is designed to use language models, alongside deterministic rules, as enabling technology. In the development build the product screen calls no language model: its document findings are pre-written. A model route exists in the code outside the product screen; it has been run once, on 89 synthetic development documents (13 November 2026), through a development harness rather than the intended production provider, and that run is not evidence of accuracy. The demonstrations on this website use pre-written outputs, and Ask Reedley uses a deterministic planner with no language model. Practices care whether the work is right, explained and finished — so we foreground the evidence, the human decision and the limits.
Statements reviewed 14 November 2026. Reedley is in development; nothing here describes a live service.
Where AI assistance is intended to be used
In the current design, AI-assisted components are intended to help with:
- reading clinical correspondence: classification, extraction of findings and proposed record updates, each with a verbatim citation;
- Ask Reedley: finding and quoting entries in the record with their sources, and saying when nothing can be found. It does not give clinical advice. This is in development and its accuracy has not been measured. In the current build the question is read by a deterministic planner and answered from structured record retrieval; no language model is consulted.
Many other functions — routing, review policy, lifecycle and verification — are deterministic rules, not AI. Rules are designed to be versioned and owned by a named clinician, with the rule behind every item visible in the product. Today every rule is an unsigned draft with no named owner.
What AI in Reedley must never do
- Make a clinical decision, diagnose, interpret a result or prescribe.
- Act without a named person’s decision, or mark work verified.
- Present an inference as a fact. Every statement carries its kind: record fact, derived, rule, AI inference or uncertainty.
- Follow instructions contained in a document. Document text is data.
- Be personified. There are no chat personas or “I” statements in clinical surfaces.
Known limitations and risks
Language models can produce output that is incomplete, incorrect or confidently wrong. They can miss a passage, misread a dose or link a finding to the wrong source. These are hazards, and they are recorded as such.
Extraction accuracy has not been established. One model-in-the-loop run on synthetic development documents was recorded on 13 November 2026, against provisional engineering thresholds that carry no clinical weight; the locked holdout has not been sent to any model, and the demonstration on this website uses pre-written outputs, not a live model. Citations are checked to appear word for word in the source; whether a quoted passage actually supports its finding is for the reviewer to judge.
Human verification and source traceability
Every generated statement links to the passage it came from, quoted verbatim, so the reviewer checks it against the original rather than taking dictation. Confidence is expressed as a checkable reason — “dose stated explicitly on page 2” — never as a percentage. Proposals are accepted, edited or rejected by a named person, and rejections need a reason.
Evaluation principles
- When the evaluation suite runs, false negatives and false positives are to be measured separately, with dangerous misses reported on their own. The one run so far, on synthetic development documents, reports error classes separately; no accuracy figure is claimed from it.
- A locked holdout set is kept apart from development and re-locked only with a recorded reason.
- Results are reported by document type and specialty, not just as an average.
- Synthetic results are never presented as real-world performance.
Change management and accountability
Any change to a model, prompt or rule that affects clinical behaviour will be versioned, evaluated against the regression and holdout suites, assessed in the hazard log and approved by a Clinical Safety Officer (not yet appointed) before release. Outputs are designed to be snapshotted with the versions that produced them, so that any decision can be reconstructed later; in the development build that history is held in browser memory and is lost on reload. Accountability for the product sits with Reedley; accountability for clinical decisions stays with the clinicians who make them.
Data protection and model training
Reedley does not currently process any real patient data. No model provider has been contracted for clinical use, so we make no claims yet about where production processing would happen or which provider would be used.
Our intended position, to be confirmed in contracts and published before any pilot:
- patient data will not be used to train models;
- processing will be limited to what is needed for direct care, under the practice’s instructions as data controller;
- providers, processing locations and retention will be listed here and in the practice’s data processing agreement.
Whether parts of Reedley are regulated as a medical device has not been assessed. MHRA guidance says intended purpose is judged partly from promotional material, so this website avoids describing clinical claims Reedley does not make. MHRA guidance on software as a medical device.