Gridnorth
Messaging for Denied, Degraded, Intermittent and Limited (DDIL) networks
Gridnorth is built to work when the network does not: it runs without a connection, shrinks messages to fit narrow links and queues them until a link opens. Message sizes below are measured. Radio performance has not yet been tested.
Why the network cannot be assumed
Peer and near-peer adversaries contest the electromagnetic spectrum. When satellite communications (SATCOM) are jammed or unavailable, units fall back to High Frequency (HF) radio and limited line-of-sight links, where one message can take many seconds to send. A tool that needs a cloud service stops working when the link to that service does.
Working without a connection
Gridnorth is designed to run locally, with no external service. In the laboratory evaluation the model ran locally on a single NVIDIA T4 graphics processing unit (GPU). Timing on end-user device (EUD) hardware has not yet been measured. Messages are held on the device until a link opens.
Making messages small
General-purpose compression does little for short messages. Schema-informed Efficient XML Interchange (EXI) does a great deal, because sender and receiver already share the schema and only the values need to travel. EXI vs gzip explains why.
| Encoding | Bytes |
|---|---|
| Plain XML text | 328 |
| gzip | 202 |
| EXI without the schema | 190 |
| Schema-informed EXI | 19 |
What size means on the air
Time on air grows with message size. The lower bound is the number of bits divided by the link's bit rate. Modem overhead, synchronization, error correction and protocol headers all add to it. We have not yet transmitted over a radio link, so we publish no airtime figures.
Priority and store-and-forward
A priority queue sends urgent messages first when a link returns, and store-and-forward holds the rest so that nothing is lost while a link is down. This is designed behaviour. We have not yet published measurements of it.
What we have not done
- Transmitted over HF, UHF or SATCOM. Airtime above is arithmetic, not measurement.
- Measured extraction time on EUD hardware. The 7.7 s median was taken on a data-centre GPU.
- Tested with APP-11 schemas, where the size reduction will differ.
Last reviewed September 29, 2026