Grace contains the attack.
Care keeps moving.
Hospital security is patient safety. Grace, the on-premise server at the center of Nightingale OS, is designed to contain a cyberattack early and keep the bedside working while the hospital recovers. The goal: fintech-grade security for the critical infrastructure of care.
Two kinds of claim live on this page, and each one is labelled. The Grace hardware and the isolation-to-recovery sequence are in design; this round builds the prototype that demonstrates them. The software guardrails further down are built today and running on synthetic data.
Without early containment, hospitals shut systems down, and patients die.
This is what a major ransomware attack can look like in a hospital that cannot contain it early. It is an illustrative timeline assembled from published research and reported cases, not a description of any one hospital. The headline figures are on the home page; the sources are at the foot of this page.
| When | What happens |
|---|---|
| Hour 0 | Ransomware spreads, and no one can tell which workstations are infected.5 |
| Day 1 | Systems are shut down to stop the spread. Clinicians go to paper. |
| The following weeks | The electronic record, labs and imaging can go dark, and ambulances are diverted. Reported downtime has averaged about 17 days; Scripps was down more than four weeks.24 |
| Week 1 | Patient volume falls by roughly a fifth.1 |
| Already admitted | For patients in the hospital when the attack begins, in-hospital mortality rises by more than a third.1 |
| After | Reported financial impact runs from tens of millions to more than a hundred million dollars, at UHS and at Scripps.34 |
The rules are tightening
The HIPAA Security Rule update is projected for final action in July 2027, and New York’s hospital cybersecurity rule has been mandatory since October 2025.67
Why containment comes first
Federal ransomware guidance says to identify and isolate affected systems first, and to keep tested offline backups.5 When a hospital cannot tell which machines are infected, it pulls the plug on everything. Grace is designed to contain unsafe endpoints early and keep clinical access up.
Contain early, without pulling the plug on everything.
Grace is the on-premise server we lease to the hospital. These four mechanisms are how it is designed to behave in an attack. All four are in design today. The prototype this round builds will demonstrate them end to end.
Unsafe workstations isolated at boot
Grace is designed to check every workstation as it boots and isolate any that fail. Containment starts early, and the core of the system stays online instead of the whole hospital going dark.
One-way optical ports isolate Grace
Grace sits behind hardware-enforced, one-way optical ports: the unidirectional-gateway approach described in NIST SP 800-82, where the hardware itself permits data to flow in one direction only.8
Charlie stays up, read-only
Grace is designed to keep a rolling four-hour backup. During an unscheduled downtime, Charlie, the app clinicians use, stays available read-only on the latest backup, with the age of the data shown on screen.9
iOS holds new work; clinicians approve it after
New work is captured on the iOS devices and held as pending, never filed on its own. When the downtime resolves, the devices push it back and clinicians review, approve and file it. The existing record stays read-only; the bedside keeps working.
Fail a boot check. Keep caring for patients.
One sequence, run end to end on the Grace prototype this round builds, on synthetic data:
Normal workflow
Clinicians work in Charlie as usual.
A workstation fails its boot check
Grace isolates it. The rest of the system stays up.
Charlie goes read-only
Charlie runs on the backup kept on Grace, with the data’s timestamp shown.
Work is held on iOS
A clinician captures new work on an iOS device: pending, not filed.
Service is restored
The held work comes back from the devices.
Approved and filed
The clinician reviews, approves and files it, including one entry that needs reconciling.
What we will measure
- Time from a failed boot check to isolation.
- Any gap in Charlie’s availability.
- The age of the data shown while read-only.
- Whether captured work is intact after recovery.
The goal
Fintech-grade security for the critical infrastructure of care. In phase one, approved work is filed in Nightingale OS, with no write-back to the hospital’s record system.
The guardrails underneath are already built.
Grace is in design. The software that will run on it is built, tested, and running on synthetic data today. These are its standing guardrails.
Gates fail closed
Where a gate cannot establish that an action is permitted, it refuses. Unknown is never treated as authorized, and the refusal returns a code naming the rule it refused under.
Weak config will not boot
Startup checks refuse to start the application on missing, weak, or development-default security configuration. A misconfigured deployment fails loudly at startup instead of running quietly without its guarantees.
A floor of five
Aggregates over people hold a k ≥ 5 anonymity floor, server-side. Small groups are suppressed, never fudged: totals, never names.
No cloud AI by default
Inference runs on-premise. Cloud seams ship dark and open only as a deliberate, BAA-attested opt-in, so no cloud model sits in a clinical or compliance decision path unless the hospital puts it there.
Every act recorded
Governed decisions are written to a hash-chained log. An edit or a deletion breaks the chain, and the break is detectable.
When systems go down, the gates stay on.
The software already treats a downtime as a mode, not a blackout: the safety gates keep running while the systems around them are dark, and everything captured while offline is tagged with its own provenance so nothing captured offline can later pass for something captured live. This is the groundwork under the in-design iOS hold-and-approve sequence above.
The gates do not take a holiday
Downtime mode is not the product with the safety rules switched off. The lab heartbeat carries the promise in one line — safety gates remain active in downtime mode — and that line is load-bearing, not decorative.
Provenance on every downtime entry
Rows created while offline carry a provenance label — lisSource=nightingale-downtime — so a downtime entry is always distinguishable from a live one, by any reader and by any query.
Reconciliation when service returns
The tag is the reconciliation key: when systems come back, offline entries are matched and reconciled instead of silently diverging from the record of what happened.
Offline capture, purged on logout
Offline entries live on the device until they sync — in a local store the logout flow purges, with tests. The shared bedside tablet belongs to the next clinician too.
Your watch is not a management tool.
Wearables only become deployable if the data can never become a management feed. So the raw signals — heart-rate variability, resting heart rate, sleep, activity — never leave the paired phone. What crosses the firewall is a derived score with its confidence band and consent-flagged factors, in three participation tiers with no penalty for choosing the lowest.
Raw signals never cross
The leadership-facing layer has no column for an individual's raw biometrics. That is not a policy someone could waive — it is a schema shape that cannot hold the data, so the promise does not depend on anyone's good behavior.
Consent first, by scope and platform
A request without active, scoped consent is refused — including tier-3 no-participation staff, whose requests are refused outright. Participation moves at will between tiers.
Aggregates keep the same floor
Any aggregate over biometric-derived data holds the platform-wide privacy floor of at least five, server-side — the same k that governs every manager-facing aggregate the platform publishes.
Location data answers questions. It does not watch people.
Real-time location tags can make a hospital measurably better, and the same tags can make it a surveillance state. The firewall keeps the first and refuses the second: location data is queryable only through pre-vetted aggregation templates that cannot return individual rows by construction, every run is logged where the staff council can audit it, and any council member can veto a template at any time.
Templates, not free-form queries
There is no API surface for an unapproved query. Buckets that fall below the anonymity floor return suppressed counts, and individual traces are not in the queryable schema at all.
Time-shifted, not real-time
Results exclude the most recent week. You can study patterns from last week; you cannot watch real-time movement.
A public log and a veto that binds
Every template run is recorded — issuer, parameters, k achieved, rows returned, a public summary — and a staff-council veto deactivates a template immediately and refuses all future runs.
Individual opt-out
A staff member can opt their tag out entirely; their pings are excluded from every aggregation regardless of the template.
Finance gets a 4xx error, not a soft warning.
A unit's resilience floor — the slack that absorbs a bad night — is the first thing a tight budget asks to spend. The platform refuses to spend it. Trades, coverage offers, shift rosters, and sick-call backfills that would drop a unit below its floor get a refusal with a citation, not a warning with an override. The Code of Design puts it in one line: Finance gets a 4xx error, not a soft warning.
One floor, every door
Trade approval, coverage offers, roster scheduling, and sick-call backfill all cross the same policy. The floor is not negotiable per workflow — which is exactly the loophole it exists to close.
Erosion, shortage, critical
Three severities, escalating with how deep the cut goes. Erosion is already a refusal — the system does not wait for the floor to be fully broken before it speaks.
The erosion ledger
Every below-floor acceptance that does happen is recorded with who asked and why, and the ledger feeds executive metrics and the manager performance composite. How the hospital is actually being run is written down, with receipts.
It also looks for what is going right.
Every other line on this page refuses something. The deviance miner completes the just-culture picture: it finds the outlier unit doing it right and makes the pattern teachable, instead of punishing the difference. A diagnosis cohort is split into an optimal arm and a comparison arm on a pre-declared outcome definition, and feature prevalence is compared across the arms — the miner proposes, humans decide what to teach.
The outcome is declared first
The outcome definition and cohort window are pre-declared, not chosen after the results look interesting. The miner is a hypothesis generator, not a retrospective fishing expedition.
Association, with honest intervals
Results carry confidence intervals and false-discovery-rate correction, and every result says ASSOCIATION — hypothesis-generating, ranked against nothing. No leaderboard, no discipline.
Small arms are refused
Each arm must hold at least five distinct patients or the query refuses with PDM_001 — an arm too small to be anonymous is an arm the platform will not compare.
A burnout program, priced the way a CFO would.
This is diligence detail, not a headline number and not our pricing. The architecture document prices one program the way a CFO would: turnover replacement cost from the published nursing literature (NSI, Aiken, Shanafelt, Tawfik), malpractice and adverse-event exposure, and CMS value-based payment — line by line, each line with its source. The program it prices is mandatory leave wired to EAP and HR, never to the supervisor, plus the wellness engine stack behind it.
A model, not a measurement: these figures are computed from published literature, not yet observed on a live deployment, and the outcome claims behind them are preregistered and remain UNVALIDATED until the validation protocol runs at a first hospital site.
The exact edge of what we claim.
Everything in the “built today” sections exists in the repository today, with tests. Everything labelled “in design” does not exist yet as hardware. Here is the edge, stated plainly.
In design: Grace and the attack sequence
The Grace server, isolation at boot, the one-way optical ports and the rolling four-hour backup are in design. The prototype that demonstrates them is built with this round.
Built and tested, not yet installed
No hospital is running these gates against live systems yet; they are exercised on synthetic data and fixtures until a hospital deployment changes that. The gates are code, not field results.
Modeled, not measured
The diligence model is sourced arithmetic. The preregistered outcome claims stay UNVALIDATED until the validation protocol runs — that is a property of the claims, not a footnote about them.
Structural, not attitudinal
The firewalls are schema shapes, startup checks, and refusal paths — properties of the code a later release cannot quietly waive. The staff council holds the veto, and the query log is theirs to audit.
Where the attack timeline comes from.
The timeline above is illustrative. Each step rests on a published source.
- Neprash, McGlave and Nikpay, “Hacked to Pieces?”, AEJ: Economic Policy 18(1), February 2026: in-hospital mortality for Medicare patients admitted at attack onset, and patient volume in the attack week.
- Comparitech, ransomware attacks on US healthcare (updated December 2024): average downtime across reported attacks, 2018–2024.
- Universal Health Services, 2020 earnings: reported pre-tax impact of the September 2020 Ryuk attack.
- Scripps Health, Q3 FY2021 financial statements: lost revenue and response and recovery costs; electronic record down more than four weeks; stroke and heart-attack patients rerouted.
- CISA and FBI, #StopRansomware Guide: identify and immediately isolate affected systems; keep tested offline backups.
- HHS, HIPAA Security Rule notice of proposed rulemaking (January 2025); Unified Agenda projects final action in July 2027.
- New York, 10 NYCRR 405.46 (hospital cybersecurity requirements): effective October 2, 2024; compliance required by October 2, 2025.
- NIST SP 800-82r3, Guide to Operational Technology Security: unidirectional gateways whose hardware permits data to flow in one direction only.
- NIST SP 800-34r1, Contingency Planning Guide: recovery point and recovery time objectives.
Nothing acts unless it passes the hospital’s rules.
Request a demo and we will show the built software guardrails refusing things on synthetic data — including, on request, the ones that refuse us — and walk through the Grace prototype plan.