In plain language
The core protocol, in plain language.
What the first paper proposes, without the mathematics.
A plain-language reading of a published proposal. Where the two differ, the paper is the authority.
On this page
What it is for.
The paper describes a shared record of what a medical model did. Each entry ties an action to a named model in a particular version, to the person who authorised its use, and to the moment it happened. The aim is that nobody, including the company that made the model, can alter that history afterwards.
The proposal keeps patient records outside the shared ledger. It uses fingerprints and other verification material instead. That boundary still needs a privacy design: metadata, predictable contents, access, and implementation details can affect what is revealed.
Who keeps the record.
Somebody has to keep a shared record, and whoever does could be tempted to change it. Systems such as Bitcoin let anyone take part and make cheating expensive. The paper takes another route and calls it proof of clinical authority. Only verified healthcare institutions may keep the record.
Their credentials would come from a group of accrediting bodies in different countries, several of which must sign before an institution is admitted, so that no single body can let in an impostor. Each institution's voice is weighted by what it has staked, how much care it delivers, and how reliably it has behaved. The weight is capped. The paper suggests about five per cent, so that no hospital group can dominate however large it grows.
How they agree.
An entry becomes final when institutions holding more than two thirds of the total voting weight have signed it. The paper’s safety argument assumes that dishonest participants control less than one third of that weight. Under the stated model and assumptions, two conflicting versions of history cannot both become final.
If the network splits, because of an outage or a dispute, it pauses. When the connection returns it carries on from the same point, with no rival histories to reconcile.
A fence around the model.
A neural network cannot be proven safe in every case. The paper does not try. It treats the model as the unpredictable part and draws a fence around it.
The paper proposes a simpler fallback method beside each model, with formal constraints defining an allowed operating envelope. A monitor would detect a limit being approached or breached and switch to the fallback as specified. The guarantees depend on the model, monitor, environment, and assumptions described in the paper.
The safety limit is the theorem. The model is the variable inside it.
Built to last for decades.
A record of a decision about someone's genes may matter in fifty years. The codes that protect records today could be broken by quantum computers long before then. An attacker could copy the data now and wait.
The paper proposes signing entries with both a conventional method and a method designed to resist quantum attacks, alongside a transition plan. Its example includes a roughly thirty-minute exposure window under specified assumptions. This is a modelled interval, not a measured security result.
Changes inside a corridor.
Models are updated. The paper borrows an idea from regulators: agree in advance which kinds of change are allowed. An update within that corridor is recorded and accepted. An update outside it is rejected by the network itself.
What the paper does not show.
These are proofs about a design. They say what must be true if the system is built as described and its assumptions hold. Nothing has been built and run in a hospital, and no proof on paper can stand in for that.
A record of where an answer came from is also no evidence that the answer was right. The paper is about being able to examine what happened. Whether the medicine was good is a separate question, and it needs separate evidence.
Where this connects.
Good questions move the work forward.
Share one workflow problem, one protocol objection, or one evaluation requirement. Your email goes to Berkcan to begin a conversation.
Sources and scope
This summary describes the published proposal under its stated assumptions. A proof about a model does not establish clinical safety or the security of an implementation.
Original paper in English