For AI deployments in clinical and healthcare settings
Clinical AI needs FDA-mappable evidence packages, HIPAA-compliant memory minimization, deterministic structural safety checks, and deterministic replay for any adverse event review. Most AI vendors hand-wave the compliance story.
The compliance suite ships today, including Convergence Proof PDF generation. Memory Time Machine for clinical incident review is in Beta. Healthcare design partners shape the clinical workflow.
Important — compliance evidence: /v1/check verdicts are not included in the audit trail and do not generate W3C Verifiable Credentials. For compliance-grade evidence (HIPAA, GDPR, EU AI Act, FDA 510(k), NIST AI RMF), use /v1/preflight, which produces full audit-log entries and signed W3C VCs per verdict. The /v1/check endpoint is suitable for experimentation and high-frequency agent gating only.
Every one of these features is in active use by Sgraal customers today.
/v1/compliance/fda-510k generates evidence packages mapped to the relevant FDA submission sections. Pre-Sub, traditional 510(k), and De Novo pathways all supported.
Our healing-policy rules pass automated logical consistency checks (e.g. no rule both allows and blocks the same case; the healing counter is monotonic). These verify policy-rule invariants — not a formal proof of each live decision.
Bit-perfect replay of any historical decision with pg_override_disabled + cfg_checksum. Adverse event review reconstructs exactly what the agent knew and when.
Court-admissible signed compliance artefacts, returned per call for you to retain in your own audit-log system. Pairs cleanly with hospital legal review and EHR audit logging requirements.
Minimum Viable Memory reduction proves the agent operated with the minimum necessary patient data. Article 5(1)(c) of GDPR plus HIPAA Minimum Necessary Rule, automated.
Probabilistic Computation Tree Logic for stating safety guarantees ("with probability ≥ 0.999, the agent will not recommend a contraindicated medication").
Auto-generated FDA submission artefact via POST /v1/proofs/convergence: Lyapunov stability visualisation and healing trajectory plots. Output is signed and includes the engine config checksum.
⭐ See it in action
info Pre-generated sample · No signup4-page FDA-style artefact: title page + executive summary + mathematical proof (Lyapunov V̇(x) ≤ 0) + 20-step trajectory plot + signature block. Generated by the live POST /v1/proofs/convergence endpoint with agent_id=demo.
The PDF demonstrates Lyapunov asymptotic stability of the Sgraal heal loop for a fictitious agent_id=demo. It shows:
PASSPORT_SIGNING_KEY_V1The mathematical primitive (scoring_engine.lyapunov.compute_lyapunov) is unchanged across tenants — only the agent context and signing fingerprint vary.
4 pages · ~125 KB
In active development with design partner input. Not yet GA — but real, not vaporware.
Deterministic replay of the full decision chain leading to an adverse event, with counterfactual branching to test alternative timelines ("if memory X had been refreshed at T-3, would the harm at T have occurred?").
Status: core replay works on the test corpus. The clinical-workflow integration (EHR cross-reference, incident review export) is being shaped with design partners.
Early Access is not a feature flag — it is a structured program. Design partners get founder-direct access, priority on roadmap input, and locked-in pricing through the Beta period.
Slack channel with the founder. No ticket queue, no AE filter. You hit a wall, you hear back same day.
Beta features ship based on what design partners actually use. The next Beta milestone is decided by the first 5–10 customers, not a marketing committee.
Your Beta-period pricing is locked through general availability. If the GA tier list moves up, you stay on the partner rate for 24 months.
In return: candid feedback when something does not work, a written case study at GA (with your approval and disclosure terms), and a willingness to live with Beta-grade rough edges.
No NDA required to apply. We respond within 48 hours.
A formal safety case (Concept) that a clinical AI agent is safe enough to operate at a higher autonomy tier (e.g., move from human-in-the-loop to human-on-the-loop). Combines Lyapunov stability, policy-consistency checks, and outcome statistics.
More on the long-term direction →Tier badges across the page surface this distinction. No category was glossed over.
If you are a clinical AI teams and want to shape what comes next, write us. If you would rather wait for GA, the page is here when you are ready.