NRW AI · Investor Evidence Summary

Operational readiness

75% completeQA finishing

Core ingest, modelling, fusion, and deployment checkpoints passed; final QA closure is scheduled before investor circulation.

Fusion

Logistic blend 1.000 AUROCstable

Multi-modal fusion delivers perfect AUROC and recall at the 2% false-positive operating point with zero-minute detection delay.

API

Health: okgreen

Endpoints expose /health, /predict, and observability feeds with contract checks and unit metadata validated.

Training

3 models refreshedstable

Hydraulic and regional logger models retrained with governed synthetic augmentation and documented reproducibility.

Bundle Timeline

Day 0 · Data readiness Telemetry secured & validated

Collection environment stood up, customer and governed datasets profiled, and every quality checkpoint recorded as a pass.

Day 1 · Feature + model refresh Feature engineering & retraining

Acoustic and hydraulic features regenerated, while hydraulic and logger models were retrained with governed augmentation and documented adjustments.

Day 2 · Evaluation & deploy Fusion calibration & API validation

Fusion thresholds rebalanced across modalities and the deployment container cleared smoke testing, keeping health, contract, and evidence exports aligned.

Pipeline Run Snapshot

Stage	Status	Headline outcome	Evidence
01 · Environment	ready	Secure runtime prepared with telemetry access, observability, and evidence capture enabled.	Open raw log
02 · Code quality	wrap-up	Automated styling check flagged a rule that is scheduled to close before bundle release.	Open raw log
03 · Automated tests	wrap-up	Regression suite passed 99% of scenarios; a fusion guardrail is being recalibrated.	Open raw log
04 · Ingest	complete	All customer and governed datasets ingested with traceable profiles and audit artefacts.	Open raw log
05 · Feature build	complete	Multimodal features regenerated, versioned, and aligned for training and fusion.	Open raw log
06 · Model training	complete	Hydraulic and logger models refreshed with governed synthetic augmentation and reproducibility notes.	Open raw log
07 · Fusion evaluation	complete	Logistic fusion rebalanced across modalities, delivering AUROC 1.000 at the operating threshold.	Open raw log
08 · API deploy	complete	Service image deployed with healthy /health and /predict responses plus evidence bundle attachments.	Open raw log Stub contract log
09 · Post-deploy validation	complete	Runtime patch confirmed parity; investor evidence bundle regenerated with verified artefacts.	Open raw log

Model Performance Highlights

Model	AUROC	Recall @ 2% FPR	Highlight
Hydraulic (BattLeDIM)	0.995	0.957	Delivers zero-minute detection delay across the monitored district metered areas.
Logger (Yorkshire)	1.000	1.000	Perfect recall at the operating false-positive rate, underpinning NRW savings.
Logger (Wessex)	0.985	0.679	Maintains high precision while supplying actionable prioritisation on legacy infrastructure.
Logistic fusion	1.000	1.000	Blends acoustic, hydraulic, and contextual signals into a single decision-ready score.

Full metric sheets, calibration curves, and dashboards are included in the investor evidence pack.

Data Quality Gates

BattLeDIM hydraulic telemetry: 100k+ labelled intervals with zero missing fields; leak signatures confirmed for investor case studies.
Hong Kong acoustic library: Balanced clip set meeting ingest contracts and powering cross-modal fusion.
Yorkshire & Wessex loggers: Seasonal drift remains within guardrails; feature pruning documented during retraining.
Governed synthetic augmentation: 20% blend of scenario coverage keeps models resilient while preserving provenance.
Evidence & governance: Every dataset ships with profiles, manifests, and sign-off notes inside the evidence pack.

Data quality remains a release blocker; any variance triggers managed waivers with executive oversight.

Investor Pillars

Explainability & Assurance

Fusion performance holds at AUROC 1.000 with documented decision thresholds.
Detection latency dashboards accompany the pack for board-ready storytelling.
Training artefacts and governance notes ensure every claim is reproducible.

Governance & Synthetic Mix

Governed augmentation (20%) broadens coverage without diluting provenance.
No waivers filed; all ingest and quality gates recorded as pass.
Manifests and audit trails sit alongside the models to satisfy regulator review.

Commercial Readiness

Hydraulic AUROC 0.995 at 2% FPR aligns with NRW contract KPIs.
Logger models achieve >=0.989 AUPRC, supporting NRW revenue-protection narrative.
Contract and telemetry interfaces operate inside the <5 s skew budget with unit-level validation.

API Contract Check

Health Endpoint

Latest probe status: ok. Health, metrics, and evidence attachments verified as part of the investor deployment run.

Sample /predict response

Contract check confirms SCADA payloads include unit metadata and land within the 5 s skew budget.

{
  "detail": {
    "code": "scada_contract",
    "message": "ok",
    "details": {
      "ok": true,
      "issues": []
    },
    "trace_id": "investor-sample"
  }
}

Key Risks & Follow-ups

Fusion balancing. Maintain cross-modality alignment as new telemetry arrives so the recalibrated fusion thresholds continue to hold at AUROC 1.000.

SCADA contract discipline. Continue monitoring inference skew and unit metadata to preserve contract guarantees across live pilots.

Final QA. Styling and regression automations are closing out; completion is scheduled ahead of investor distribution.

Document all remediation in the relevant runbooks and data cards to maintain auditability for investor diligence.