Skip to content
PensieveThe record comes first

Not AI.

The Software Underneath Comes First.

One record, read and written by every tool through the same core systems. Authorization is checked on every read and every write, and the audit trail is written in the same transaction as the change. A model that runs here runs on that, and it runs second.

One governed recordGovernance + Auditpensievelabs.org

Position 1 of 7.

We built the record first.

A hospital does not fail because nobody was clever enough. It fails because its systems disagree about who the patient is.

Read how the platform is built

Position 2 of 7.

What a hospital actually runs today.

Every one of these systems works. The failure is in the space between them.

Walk the floor of a working hospital and count the systems that each hold a piece of the same person. Between all of them sits paper: the consent form, the referral note, the allergy written in the margin, the discharge summary that exists in one folder and one folder only.

That is the part people miss. Registration registers patients correctly. The laboratory reports results correctly. The failure is not inside any of them. It is where a person retypes a name, matches a result to an episode by eye, and carries a number from one screen to another.

None of them agree on who the person is. There is no system of record here. There is a set of systems, each of them convinced that it is the record.

Position 3 of 7.

What happened when intelligence was tried first.

This is not a suspicion. It is a published record, and most of it was written by the institutions that deployed the models. An external validation of a widely implemented proprietary sepsis prediction model, at one academic health system, put its discrimination at an area under the curve of 0.63 and found that it identified 33 percent of the patients who had sepsis. Source 45 An algorithm applied to roughly 200 million people a year used health cost as a proxy for health need, and correcting that one label, with the algorithm left as it was, changed which patients it sent for extra care. Source 46 A screening system reported above 90 percent accurate in the laboratory rejected 21 percent of the images clinic staff submitted to it, on lighting and upload speed. Source 47 Of 521 regulator-authorised devices reviewed, 226 carried no published clinical validation data at all. Source 48

Not one of those is a reasoning failure. Each is a record failure surfacing downstream: a label that meant something other than what it was read as, an environment nobody had modelled, a validation nobody had published. An independent patient safety organisation ranked risks with AI-enabled health technology first on its 2025 hazard list, above cybersecurity. Source 52

Clinicians are not the obstacle here. In a sixteen country survey, 76 percent of Indian healthcare professionals said AI could improve patient outcomes, and 40 percent said the tools in front of them fit their day-to-day clinical work. Source 53 The distance between those two figures is the distance between a model and the record it has to read.

Position 4 of 7.

The problem is not intelligence. The problem is the record.

Intelligence applied to a broken record does not correct the record. It launders it.

A model reasons over a record. Where the record is a set of copies that disagree, the model inherits the disagreement and returns it with better grammar and more confidence than it had on its own. The people who build these systems report it as routine: across one study of practitioners in India, Africa and the United States, 92 percent had met a failure that began in the data and surfaced late, as unexplained model behaviour. Source 50

Look at what actually goes wrong on a ward. A result matched to the wrong episode. A discharge summary written before the imaging report arrived. A bill raised on a tariff nobody applied. An allergy recorded in a system the prescriber does not open. Every one of those is two systems disagreeing about a fact. Not one of them is a reasoning failure.

So the question to put to a vendor is not how clever the software is. It is this: when two of my systems disagree about this patient, which one is right, and how does your product know? Pensieve answers it by holding one record and treating everything else as an adapter into it.

Position 5 of 7.

One governed record, and the mechanisms that keep it one

One record, shared by every tool Admissions, the clinical record, orders, imaging, pharmacy, and billing read and write the same record through the same core systems. There is no reconciliation job running between them, because there is nothing to reconcile. Explore the platform

Adapters, so the systems you run stay The core sits behind ports, and adapters plug the outside world into them. Your existing registration, laboratory, imaging, and billing systems keep running and reach Pensieve over one outbound, mutually authenticated link. Explore ports and adapters

Authorization on every read and write Cirith decides every read and every write. No tool implements access control on the side, and an automated check fails the build when a tool tries to route around the core systems. Explore Cirith

An audit trail written once Every action is audited in the same transaction as the change, then archived to write-once storage. The log cannot drift from the record, because it is written with it. Explore the audit trail

The strongest of these is the one you cannot see in the interface. The patient lens is injected at the database read path, so records outside a user’s entitlement never leave the store. Hiding a button was never the mechanism: a lens at the read path cannot be routed around, and an interface bug cannot leak what the query never returned.

The rest of the architecture is published for anyone to check. Ports and adapters, errors as values, event-driven core systems, the enforced-architecture check that fails the build: each one is a page you can read at platform architecture.

Position 6 of 7.

Where we take risk out before it reaches a patient.

Cornea Model a site’s configuration and rehearse it on a dummy site before it goes live. What you walk through is the same configuration the live site will run, so mistakes surface where they cost nothing. No forecast is involved: the configuration has already run once, on the dummy site. Explore Cornea

Westron Configure every connection to an external system, then watch its health and message flow, so a silent lab feed is noticed while it is still silent. The work it does is surfacing an absence. Explore Westron

This is not an argument against machine intelligence. It is an argument about order. Where a model does belong, it is named, its single task is stated, and its output is written to the same audit trail as every other action. PensieveAI makes that case on its own page.

What holds on this side of that line holds whatever we ship. Nothing in Pensieve makes, supports or influences a clinical decision. Every capability that uses a model is named in a dated public inventory before a hospital can switch it on, and it is off until they do. A hospital’s record trains nothing, for us or for anyone. The list is short and specific, and each item is an undertaking we are held to in writing: Responsible AI and Not a medical device.

Position 7 of 7.

See the record on workflows like yours.

Hospitals run Pensieve today, under NDA, and we will walk you through the record itself.

Request a technical briefing

Get started

The systems a working hospital runs, and the form in which a fact leaves each one
SystemWhat it holdsHow a fact leaves it
Registration and admissionsIdentity, admission, bed, episodeA patient number the other systems do not use
BillingCharges, tariffs, claimsAn account number reconciled to the patient by hand
LaboratoryOrders and resultsA CSV export, or a printout walked to the ward
Imaging archiveStudies and radiology reportsDICOM, and nothing else
PharmacyStock, dispensing, batchA ledger with no link to the order that caused it
PaperConsent, referrals, allergies, handwritten notesIt does not leave. It is filed.

Powerful. Responsible. Pensieve.

Get started