The Act is not a threat to a product built to be verified. It is a tailwind. Every requirement it places on high-risk clinical AI, Axiisium already treats as the point of the product, not a compliance tax bolted on afterward.
Art. 12 & 19 · Record-keeping
Automatic, durable logs
High-risk systems must automatically log events and retain them for at least six months.
Axiisium binds every decision to a signed, tamper-evident record, anchored to an external append-only log. Not just retained: independently verifiable, by anyone, without trusting us.
Art. 13 · Transparency
Legible capabilities and limits
Operation must be transparent: intended purpose, accuracy level, and limitations disclosed to the deployer.
Each call shows which signal drove it, a calibrated confidence, the WHO 2022 / ICC 2022 rationale, and states plainly what is assumed and what is deferred.
Art. 14 · Human oversight
The clinician decides
Systems must let a competent human monitor operation, catch anomalies, and intervene.
A qualified clinician makes the diagnosis; a confirmatory assay gates any decision; the model ranks who to sequence but never makes the genetics call. Fail-safe by architecture, not by policy.
Art. 15 · Accuracy & robustness
Measured, not asserted
Appropriate accuracy, robustness, and cybersecurity, declared and evidenced.
Morphology assessed at a preliminary, patient-grouped, out-of-fold 0.819 to 0.977 AUROC across four AML lesions (129 AML patients, healthy donors excluded, means over 10 seeds), with one lesion reproduced on an external cohort at 0.724 and the other three not yet, calibrated confidence bands, and a configurable recall-floor that bounds the false-negative rate.
Art. 9 · Risk management
Safe by construction
A continuous risk-management process across the lifecycle.
The ranking cannot itself drive an unconfirmed clinical action: mandatory confirmatory gating and the recall-floor are properties of the pipeline, and it defers hard markers to the assay.
Art. 10 · Data governance
Provenance of the inputs
Training and evaluation data governed for quality and representativeness.
Validation is patient-grouped and leakage-free, and each signed record carries a digest of the exact inputs that produced the call.