A Flag Is Not a Verdict: Building Cartel Detection That Can Survive Contact with Reality

Cartel detection is often framed as an AI problem. It is not. It is a systems-design problem involving statistics, procurement law, data governance, and operational discipline.

A model can rank suspicious tenders. It cannot establish an infringement. Treating those as the same thing is how an otherwise useful analytics programme becomes legally fragile, operationally distrusted, and unfair to the companies it is supposed to assess.

Public procurement is an unusually important place to get this right. The OECD notes that its member countries spend roughly 13% of GDP through public procurement, while bid rigging can raise prices or reduce quality in the goods and services the public receives.1 The European Commission identifies several familiar cartel patterns: agreeing bids in advance, refraining from bidding, allocating customers or territories, and rotating winners while preserving the appearance of competition.2

My research draft proposes a machine-readable framework for turning those patterns into a disciplined screening system for Estonian public procurement. Its central premise is deliberately modest:

A screen produces a lead. A human authority evaluates evidence and decides.

That distinction is not bureaucratic caution. It is the architecture.

Rotation is a sequence problem, not an entropy problem

One of the more interesting findings in the draft concerns winner rotation. A naive approach might calculate the distribution of awards across firms and conclude that a market with evenly distributed wins is competitive. That is wrong.

A perfectly rotating cartel can produce an almost uniform winner distribution: A wins, then B, then C, then A again. Marginal entropy is high, the result looks diverse, but the next winner is highly predictable. The signal is therefore conditional entropy in the ordered winner sequence, tested against a permutation null, not the marginal entropy of win shares.

That detail matters because poor statistical specifications create false confidence. It also matters because procurement markets are not stationary laboratory systems. Firms enter, exit, merge, rebrand, and win one-off contracts. Pooling all occasional winners into a generic “other” state can manufacture a neat-looking two-state alternation and a false cartel signal. The proposed approach treats entrant wins as sequence breaks: no artificial transition is created across a gap. Less data survives, but the remaining evidence means what the system says it means.

The lesson is broader than procurement: when the process is temporal, do not reduce it to a static distribution and hope the important structure survives.

Price screens are useful, but price alone is not enough

Bid distributions still matter. Cover bidding can compress dispersion, and incomplete cartels can leave detectable signatures in subgroups of bidders even when genuine competitors are present. Research on incomplete bid-rigging cartels shows why subgroup-level screening can outperform treating every tender as one undifferentiated bid set.3

Yet cartels do not always submit cover bids. Sometimes they simply stay away from one another’s territories, customer groups, or lots. That calls for participation analysis: which firms were plausibly capable of bidding, active in comparable work elsewhere, but repeatedly absent from each other’s target tenders?

Even that is only corroborative. In a thin market, apparent avoidance can reflect capacity limits, geography, framework contracts, specialisation, or ordinary oligopolistic self-selection. A credible system must model those alternatives explicitly and must be able to remain silent when the data cannot distinguish them.

This is the opposite of dashboard theatre. A red score with no model of legitimate explanations is not intelligence.

Keep deterministic rules, learned scoring, and generative AI in separate lanes

The framework uses three different technical roles, and they must not be blurred.

Layer	What it does	What it must not do
Deterministic screens	Calculate pinned statistics, apply documented thresholds, and identify reproducible graph or metadata patterns.	Change a flag after the fact or infer legal liability.
Learned scorers	Rank an alert queue and estimate calibrated risk from several signals.	Create, erase, or override deterministic hard flags.
Generative assistance	Summarise documents, prepare a reviewer dossier, and point to source records.	Decide an infringement, open a case, or communicate an adverse decision.

This separation is not an argument against AI. It is a way to use it without losing the audit trail. Every lead should carry the data-snapshot identifier, screen version, threshold version, and, where a model assisted, its model version and prompt hash. The outcome of human review should be recorded beside the system’s output, never overwritten by it.

A system should be able to answer a hard question years later: why did this alert exist on that date, using that data? If it cannot, it is not ready for enforcement-adjacent work.

Estonia has an opportunity to build this as public infrastructure

The practical advantage is not a magic model. It is the ability to join the right evidence under the right controls: procurement events, company and ownership relationships, documents, and, inside the protected boundary, submission and digital-signature metadata.

Estonia’s interoperable digital-government foundations matter here. X-Road is designed for unified and secure data exchange among organisations, which makes it a relevant building block for controlled, auditable exchange rather than unrestricted data replication.4 The resulting design should separate open procurement data from restricted registry data and internal operator data, minimise personal data, use pseudonymised egress where analysis leaves the protected environment, and treat submitted documents as untrusted input.

The point is not to create a surveillance machine for suppliers. It is to make secrecy more expensive for coordinated manipulation while preserving due process for legitimate businesses.

Start with a pilot that can admit failure

The next step is not national deployment. It is a retrospective pilot.

Run the screens across historical procurement data. Review the highest-ranked constellations with domain experts. Include a structured sample of unflagged cases so the programme can learn where it is blind. Measure precision at the actual review capacity, not headline accuracy on an artificially balanced dataset. Publish the methodology, keep gaming-sensitive thresholds confidential, and recalibrate when markets change.

The OECD’s guidance is consistent with this direction: digital cartel screens need quality procurement data, trained staff, and effective cooperation between procurement and competition authorities.1 The European Commission likewise distinguishes plausible indications from evidence and emphasises balanced assessment.2

That is the standard worth building to. Detect patterns early. Preserve provenance. Escalate carefully. Let people, not models, carry responsibility for the decision.

References