RasembliResearch Inc.FR

Gridnorth

How Gridnorth works

A language model proposes field values. Every other stage, from verification to rendering to encoding, is deterministic code that can be tested line by line.

Probabilistic where it must be, deterministic elsewhere

Language models understand informal input well and guarantee structure poorly. Military messaging needs both. Gridnorth confines the model to one task, proposing field values, and assigns every other task to code.

Message path
voice / text / legacy message
  -> constrained extraction     (model proposes field values)
  -> confidence and abstention  (uncertain fields go to the operator)
  -> operator confirmation
  -> deterministic rendering    (MTF-XML, COIE, NIEM) and validation
  -> schema-informed EXI encoding
  -> priority queue and store-and-forward -> radio

  1. Input

    Speech is transcribed on the device. Text and chat are used as written. Legacy character-oriented messages are parsed directly.

  2. Constrained extraction

    The model receives the definition of the target message: field names, allowed values and formats. Its output is constrained at every step to what that definition permits, so the record is always structurally valid. The accuracy of the values is a separate question and is measured separately.

  3. Confidence and abstention

    For every field, Gridnorth records the model's confidence among the values the schema allowed. A field below the threshold, or one that fails a format rule (for example, a grid reference that is not a valid Military Grid Reference System (MGRS) string), is presented to the operator before release.

  4. Deterministic rendering

    The confirmed record is written as MTF-XML, as a character-oriented message and as NIEM. Each output is validated against its schema. A message that fails validation is not released.

  5. Encoding and delivery

    Schema-informed EXI encodes the XML without loss. Sender and receiver share the published schema, so the encoded message carries values and not markup. A priority queue and store-and-forward handle intermittent links.

Where the uncertainty sits

Encoding is lossless and rendering is deterministic, so information can only be lost at extraction. Testing concentrates there, and the operator's confirmation step sits there. The evaluation reports extraction accuracy separately from structure, and states how many errors the confirmation step currently catches.

Last reviewed September 29, 2026