Most digital-health sites tell you everything the technology can do. This page tells you exactly which claims are engineering fact, which are tested, and which are still hypotheses awaiting independent evaluation. Assurance should be visible to the people being asked to trust it.
| Reference Model | v1.0-RC (frozen) |
| Register version | 2026.08.2 |
| Rules | 35 — every rule versioned, owned, sourced |
| Schema | s-2026.08.2 |
| Build | B-2026-08-24-02 |
| Canonical SHA-256 | 25aeca5e4e8cf68123d0be81095df68baa41dc47d622abbe442d8f2f07e78d87 |
Every finding envelope carries the register version, the full hash, the schema version and the build ID — so any output can be reconstructed, months later, against exactly one executable artefact. A build gate (skeina-build.js) fails automatically on any drift between the page, the engine, the harness and the published validation pack.
| 26 defined scenarios (positive, negative, boundary, missing-data, conflict) | PASS |
| 7 hard safety invariants (incl. unknown ≠ negative; precedence preserves evidence) | PASS |
| 1,000 external stress cases (independent AI red-team) | PASS — 2 defects found & corrected |
| 10,000 adversarial cases after corrections (independent rerun) | PASS — 0 invariant failures |
| Determinism across replays | PASS |
| Build-gate drift check | NO DRIFT |
The 1,000-case run found two release blockers — an emergency action-precedence conflict and unknown data behaving as a negative. Both were corrected, both corrections are verified in the current suite, and the defect history is published rather than hidden. That is what a safety-testing process is for.
What this testing does not establish: sensitivity, specificity, predictive value, clinical effectiveness, patient benefit or cost-effectiveness. The frozen reference engine passed 10,000 generated adversarial cases against its defined technical and safety invariants — nothing more is claimed.
| Stage A · Synthetic technical & safety validation | COMPLETE |
| Stage B · Retrospective blinded validation against real records (dual independent clinician review, adjudication, silent-case audit) — subject to governance | NEXT |
| Stage C · Prospective silent mode (findings generated, not shown; compared with usual care) | NOT STARTED |
| Stage D · Human-factors study (comprehension, time, alert burden, send-back behaviour) | NOT STARTED |
| Stage E · Assisted pilot — professionals see findings; no autonomous action | NOT STARTED |
| Stage F · Independent comparative neighbourhood evaluation | NOT STARTED |
| External ICB / think-tank / evaluator demonstration | PERMITTED |
| DCB0129 clinical-safety development work | IN DEVELOPMENT |
| Live actioning of findings in real care | NOT YET |
| Autonomous decision-making | NEVER — by design |
| Clinical-effectiveness or "world-class proven" claims | NOT YET — evidence must earn them |
Clinical acuity assessment and acceptance of clinical proposals remain with appropriately qualified professionals; Skeina does not autonomously make or execute clinical decisions. Whether any Skeina function qualifies as software as a medical device will be determined from its intended purpose and functionality, with formal MHRA advice sought before deployment. Clinical safety engineering (DCB0129 for development; DCB0160 for each deploying organisation) begins during development, not at go-live. This is a safety architecture — not a claimed regulatory exemption.
SKEINA-DATA-01 — Unknown is a first-class data state. Missing information must never silently select a negative, lower-risk or less urgent pathway. Implemented, tested, and part of the standard.
Every major component has precedent — Frome, Findhelp, Foundry, Kinzigtal, Bradford. The combination is the hypothesis. The reference engine is technically reproducible and has passed its synthetic safety test suite; whether it improves care is the next question, and that will be independently evaluated. The validation pack, the runnable harness and the build gate are available to any reviewer — do not endorse it, try to break it.