Field manual|Data & More|Classification engine

How the Privacy & Security Model classifies data

Ten years of taxonomy, compressed into a local 8B model that runs inside a harness.
10 yrs
taxonomy development
9M
synthetic labelled examples
(See sample on Hugging Face)
8B
parameters, Mistral-based
Local
runtime, on local infrastructure
Section 01|Training

Building the model

The taxonomy is used to generate a large synthetic training set of already-classified examples. That set trains a compact Mistral-based model to recognise classified data.

SOURCETaxonomy10 years of rulesand data typesgenerateTRAINING SETSynthetic data9M labelled examplesgenerated from taxonomytrainMODELPrivacy & Security8B parametersMistral-based
main path
trained artefact
Section 02|Inference

Classifying documents

INPUTDocumentsall in scopeSTEP APre-filterstatistical modeltoken compositionpassSTEP BPrivacy & Security Model HarnessMODEL · LOCALP&S Model 8BMistral-basedoutput ⊂ permittedclassification types + levelsOUTPUTClassifiedpermitted labelskipSKIPPEDNot sent
main path
skipped branch
classified result
model inside harness
Step A · pre-filter
Only likely candidates reach the model
A statistical model reads each document's token composition and context. Documents with a relevant profile pass; the rest are never sent to the model.
Step B · harness
The model runs inside guardrails
The Privacy & Security Model Harness restricts which classification types and levels the model may return. The model runs locally, with no external AI service in the loop.
Section 03|Unique IP

Four things only Data & More has

IP 01
10-year classification taxonomy
A classification framework built and refined over ten years. Everything downstream is generated from it.
IP 02
9 million synthetic labelled examples
A training set generated from our own taxonomy. The data and the method behind it belong to Data & More.
IP 03
Privacy & Security Model Harness
Our own control layer restricting which classification types and levels the model can return.
IP 04
Statistical pre-filter
Our token-composition model deciding which documents are worth sending to the model at all.
Summary|Principles
| EXPERIENCE
Built on ten years of classification work.
| EFFICIENCY
A small model, fed only the documents that matter.
| CONTROL
Outputs limited to approved classification types.
| LOCAL
A Mistral-based model running locally.
Don't try this at home
Data & More ApS | Privacy & Security Model | Field manual