POPULATION IDENTIFICATION

AI Business Classification & Validation Lab

Model Engine:

Illustrative Experimental Research Output

Statistical Guardrail: Business classification is a population identification aid, not financial attribution. Being identified as an AI-relevant enterprise does not imply that 100% of the firm’s turnover or GVA is AI economic output. Diversified firms require product-level disaggregation. Human review is mandatory for official register curation.

Business Description Input

Benchmark Corpus: 60 UK Firms

Select from Curated UK Benchmark Samples:

Company Name:

SIC Code:

Activity & Capability Description:

Feature Explainability & Term Weights

"foundation model"+3.1

"large language model"+3.1

"deep learning"+2.85

"ai safety"+2.3

Model Classification Output

AI-RELEVANT ENTERPRISE

AI Relevance Probability100.0%

Confidence Score1.00

Predicted ONS Table 3 Taxonomy Categories:

AI Platforms, Foundation Models & Data InfraCybersecurity, AI Safety & Governance

Dedicated AI Enterprise

Core business activity centers primarily on proprietary AI engineering and models.

Human Review & Adjudication (Mandatory Rule 9)

Status: unreviewed

In accordance with statistical governance rules, model outputs propose candidate classifications but must never silently become ground truth.

Reviewer Justification / Audit Notes:

Benchmark Corpus Evaluation & Generalisation Audit

Curated experimental benchmark dataset — not official statistics and not a representative sample of UK businesses (DS05, N=60).

Engine: tfidf_logistic

Evaluation Protocol & Small-Sample Limitation (N=60)

The reported in-sample accuracy evaluates keyword and feature alignment on the 60 curated research profiles. Cross-validation evaluates stability across held-out test splits. These metrics do not demonstrate generalisation across the 5.6 million diverse registered UK enterprises.

1. In-Sample Fit (All 60 Samples)Calibration Fit

Accuracy98.3%

Precision100.0%

Recall97.5%

F1-Score98.7%

2. Stratified 5-Fold Cross-ValidationHeld-Out Splits

Mean Acc98.3%

Precision100.0%

Recall97.5%

F1-Score98.7%

In-Sample Confusion Matrix (Binary AI Relevance)

True Positives (TP):39Correctly identified AI firms

False Positives (FP):0Hard negatives misclassified

False Negatives (FN):1AI firms missed

True Negatives (TN):20Correctly identified non-AI

Session Review Decisions Audit Trail (0)

No human review actions recorded yet in this session.