Click Print to PDF below. In the print dialog, set paper size to A4, margins to None, and enable Background graphics.

Frontier AI Discovery — Innovate UK · Appendix

CarbonRoute

Foundation Model for Carbon Attribution in UK Construction Logistics

Page 1 of 2

System & Solution Architecture

System Architecture — Data Pipeline

📥

Data Ingestion

Tradesman Here platform

›Job postings + postcodes
›Hire confirmations
›Days on site × workers
›Round-trip distance (km)

🗄️

Dataset Construction

Phase 1 output

›Postcode-pair records
›BEIS emission factors
›CITB locality benchmarks
›Scope 3 tagging (Sec. 3)

🧠

Model Architecture

Foundation model design

›Attribution head design
›Spatial embedding layer
›Trade × region encoders
›Uncertainty quantification

✅

Validation & Output

Feasibility deliverables

›BEIS benchmark checks
›Expert panel review
›White paper + spec doc
›Phase 2 roadmap

Model Architecture — Foundation Model Design

Input Layer

•Postcode-pair embedding (OSRM)
•Trade category encoding (37 types)
•Contract type + duration
•Temporal features (season, year)

Core Model

•Transformer-based architecture
•Spatial attention mechanism
•Multi-head: distance / CO₂ / locality
•Uncertainty estimation (Monte Carlo)

Output Layer

•kg CO₂ per hire (point estimate)
•Confidence interval (±%)
•Locality tier classification
•Scope 3 category attribution

💡

Feasibility scope: This study determines whether a foundation model approach can outperform simple emission-factor lookup for carbon attribution across 37 UK trade categories and variable journey types — informing the Phase 2 full model build.

Data Sources & Methodology Chain

Proprietary

Tradesman Here

Live hire data: postcode pairs, trade type, days on site, round-trip km

Government

BEIS 2023

UK official emission factors: 0.171 kg CO₂/km (van/car average)

Industry

CITB / ONS

National commute baseline: 61 km one-way average, workforce survey data

Spatial

OSRM / OS

Postcode-level routing for Phase 2 true distance calculation

Project Plan — 6-Month Timeline

Workstream

Month 1

Month 2

Month 3

Month 4

Month 5

Month 6

WS1 · Dataset Construction

Validated dataset ≥500 records

WS2 · Model Architecture

Prototype model + spec doc

WS3 · Methodology Validation

Peer-reviewed white paper

WS4 · Feasibility Report

Final report + Phase 2 roadmap

Key Milestones

Kickoff

Dataset locked

Model v1

Expert review

White paper

Final report

CarbonRoute · Frontier AI Discovery Application · Tradesman Here Ltd

April 2026 · CONFIDENTIAL

Frontier AI Discovery — Innovate UK · Appendix

CarbonRoute

Foundation Model for Carbon Attribution in UK Construction Logistics

Page 2 of 2

Validation & Benchmark Matrix

Validation Framework

What We Validate	Benchmark / Reference	Method	Success Criterion	Workstream
CO₂ attribution accuracy per hire	BEIS 2023 emission factors (0.171 kg/km)	Compare model output vs. manual BEIS calculation across 500 held-out hire records	< 10% mean absolute error vs. BEIS-calculated ground truth	WS1 Dataset + WS2 Model
Locality classification accuracy	ONS / CITB commute distance bands (local / regional / national)	Classification report (precision, recall, F1) against labelled postcode pairs	> 90% F1 score on locality tier classification	WS2 Model Architecture
Scope 3 category alignment	GHG Protocol — Category 9 (downstream transport) definition	Expert panel review: 2 sustainability consultants + 1 RICS-accredited assessor	Unanimous confirmation of Category 9 compliance from expert panel	WS3 Methodology
Model vs. lookup table comparison	Simple emission-factor lookup (baseline model)	A/B test: foundation model predictions vs. postcode-area lookup on same dataset	Foundation model ≥ 15% lower MAE than lookup baseline across trade types	WS2 Model Architecture
Dataset representativeness	CITB UK Workforce Survey 2023 trade distribution	Chi-squared test: compare trade category distribution in dataset vs. CITB census	p > 0.05 (no significant deviation from national distribution)	WS1 Dataset Construction
Uncertainty quantification reliability	Calibration curve — predicted confidence vs. empirical accuracy	Reliability diagram + Expected Calibration Error (ECE) on test set	ECE < 0.05 (well-calibrated confidence intervals)	WS2 Model Architecture
Methodology reproducibility	ISO 14064-1 GHG accounting principles	Independent replication by external researcher using white paper only	External replication within ±5% of published figures	WS3 Methodology

Workstream → Deliverable Mapping

WS1 · Month 1–2

Dataset Construction

Validated dataset of ≥ 500 postcode-pair hire records with CO₂, locality, trade-type labels

KPIs

≥ 500 records · ≥ 10 trade types · p>0.05 CITB alignment

WS2 · Month 2–4

Model Architecture

Prototype transformer model + architecture specification document with layer design rationale

KPIs

< 10% MAE · ≥ 15% improvement vs. lookup · ECE < 0.05

WS3 · Month 3–5

Methodology

Peer-reviewed methodology white paper submitted to CIBSE / Innovate UK journal

KPIs

Expert panel sign-off · GHG Protocol aligned · Reproducible

WS4 · Month 5–6

Feasibility Report

Phase 2 roadmap with cost model, compute requirements, dataset scale plan, and go/no-go recommendation

KPIs

Submitted on time · Accepted by Innovate UK · Phase 2 bid

Key Risks & Mitigations

Risk	Likelihood	Impact	Mitigation
Insufficient hire volume for statistically valid dataset	Low	High	Supplement with synthetic records generated from CITB/ONS distributions if live data < 500 records; clearly labelled as synthetic in methodology
Foundation model shows no improvement over lookup baseline	Medium	Medium	Feasibility study design anticipates this outcome — negative result is a valid finding that informs Phase 2 architecture decision (hybrid approach)
Expert panel unavailability causing methodology delays	Low	Low	Two sustainability consultants pre-identified (RICS-accredited); written review process allows async participation

CarbonRoute · Frontier AI Discovery Application · Tradesman Here Ltd

April 2026 · CONFIDENTIAL