Sentinel

Government estimates fraud and error across the public sector at £55bn–£81bn a year, excluding tax and welfare. The Public Sector Fraud Authority reported £7.53bn saved in the last financial year and credited the result explicitly to AI and advanced data-matching — but that work concentrates on grants and payments, not on the award stage where procurement fraud originates.

The classic red flags of bid-rigging are all detectable in data that is now public: single-bidder tenders, repeated runner-up patterns, suppliers sharing directors or registered addresses, contracts split just below thresholds, and award values clustering suspiciously close to estimates. Nobody in UK government is systematically looking for them across all buyers at once.

The Procurement Act 2023 improved the raw material substantially. From 1 January 2026 contracting authorities must publish supplier performance for contracts over £5m, and from 1 April 2026 suppliers awarded below-threshold contracts must register on the Central Digital Platform and supply their Companies House number.

Identifier coverage is contested, and the disagreement is itself the finding. Two independent measurements against the live feed in August 2026 disagreed sharply — roughly 65% of supplier records carrying no scheme when counting company numbers on supplier entries alone, against roughly 8% when counting the Procurement Act’s own organisation number across all parties. Both are defensible. The identifier exists; it is not where most consumers look for it. That organisation number is the key worth building on — it exists for councils, NHS trusts and unincorporated bodies with no company number, it survives group restructuring, and the debarment list is keyed on it. What nothing publishes is the mapping between it and Companies House.

One constraint eliminates most of the academic literature. Individual bid prices are never published in UK procurement data — only aggregate statistics, and bidder counts appear on roughly 30% of awards. That rules out the entire statistical screen family: coefficient of variation, relative distance, normalised distance and the missing-bids test. Sentinel is designed around what exists rather than what the papers assume.

The Open Contracting Partnership’s one-year analysis found 57 of 71 integrity indicators are now computable, up from 35 under the old regime — and that single-bid tenders cost around 7% more. It also found institutional variation nobody is acting on: Bristol awards 95% of below-threshold contracts non-competitively; Leeds under 30%. Transparency International identified 135 COVID contracts worth £15.3bn carrying three or more corruption red flags.

One structural obstacle explains why this has not been done. Companies House publishes free bulk files for companies and PSCs, but there is no free bulk officers file — the directorship graph can only be assembled company-by-company through a rate-limited API across 4.93 million active companies. That single asymmetry is why UK director-network analysis is dominated by commercial products rather than public ones.

Sentinel ingests the full public procurement record — Find a Tender, Contracts Finder and the Central Digital Platform — and resolves suppliers to Companies House entities. Because roughly two-thirds of records carry no usable identifier, that resolution is the system’s hardest engineering problem rather than a given: identifier where present, probabilistic name and address matching where not, with match confidence published on every link and low-confidence links never silently merged. It then builds a national graph connecting suppliers to their directors, persons with significant control, registered addresses and corporate parents.

Against that graph it runs an indicator suite derived from established international practice — the EU’s Digital Whistleblower red-flag set, the World Bank’s procurement analytics, and Ukraine’s ProZorro/Dozorro monitoring model — scoring every award and every buyer for risk rather than accusing anyone of wrongdoing.

Output is a triage queue for investigators, not a verdict. Each flag carries its evidence, its statistical basis and a plain-English explanation, so a commercial officer can dismiss it in thirty seconds or escalate it with a full audit trail.