01Mbrane · the protocol

Train AI on datathat cannot move.

Plenome builds the infrastructure for private model training, and the applications that run on it.

01The problem

01 / 02

The best enterprise data is the data you cannot export.

Sensitive axis. Never leaves the enclosure

Regulation, contracts, and risk keep it inside the firewall.

So the models that need it most never see it.

Where the value sits

LongitudinalNot public

Years of the same patients, the same machines, the same processes, recorded in one place.

UnstructuredNot scraped

Clinician notes, handover text and scanned artefacts that no public corpus contains.

ConsequentialNot synthetic

Records attached to a decision that was actually taken, and the outcome that followed.

None of this exists outside the institution that produced it.

What holds it in place

RegulationCannot leave

Data protection law and jurisdiction rules decide where records may physically sit.

ContractsNo transfer

Processing agreements with the data owner forbid onward transfer to a third party.

RiskNo copies

Breach liability and audit exposure make every copy a new place to be compromised.

None of these are solved by moving the data. They are solved by not moving it.

02The flow

Enterprise data

Training on Mbrane

NodeHosts data, runs blind evaluations, tracks contribution.IN PROGRESS

Co-owned model IP

TrainManages training across contributors.IN PROGRESS

Inference on Mbrane

AttestDLT layer and model versioning.PLANNED

Inference royalties

InferServes co-owned models and pays royalties.PLANNED

01 / 04 · Node

Keep scrolling

03Mbrane

Data quality evaluation oraclesTo create high-quality training tokens.

Distributed learning on enterprise dataNo data movement.

A lightweight DLT layerTo create co-owned model IP and to track contributors.

Evaluation · scored in place

Token quality0.94Scored

Coverage0.87Scored

Duplication0.02Scored

Leakage0.00Blocked

Records exported0

Every token is scored where it sits. The scores travel; the records do not.

Distributed training · clients

ClientRndsAccΔContrib.Share

CLIENT-0114/2091.4%+1.1031%

CLIENT-0214/2092.4%+0.9021%

CLIENT-0314/2089.1%-0.2018%

CLIENT-0414/2088.1%-0.8015%

CLIENT-0513/2088.1%+0.209%

CLIENT-0612/2084.2%-2.106%

Leave-one-out contribution score, recomputed on a held-out set. It decides royalty share.

DLT ledger · records committed

Records committed42

Last hour17

Inferences metered4

Payouts settled2

13 Julpeak 42026 Jul

Every round, gate result and payout is committed as an event. Contributors appear as IDs.

Node & Train, running now

04The turn

01 / 02

Data stays.

Raw records, identifiers and contracts never reach the surface.

Weights travel.

Only gradients, weights and a signed contribution record leave.

Stays inside

Raw records

Patient identifiers

Contracts

Derived tables

Nothing in this group is copied, pooled or shipped.

Travels out

Gradients

Model weights

Signed contribution record

Only these cross the surface, and only outward.

The surface is semi-permeable in one direction only.

05Close