AntitrustPro is antitrust compliance software.

Your antitrust compliance program can tell you who completed the training. It cannot tell you where your exposure actually is.

AntitrustPro scores antitrust risk across your business by an explicit method, applied the same way each time — so you can answer the board, and the Division, with something better than a binder.

Built by Linh Duong, who ran a global antitrust centre of excellence across four regions at a ~$4B medical device company, and who in December 2024 built and launched the legal AI assistant that answered antitrust compliance questions for all 10,000+ of its employees at a self-reported accuracy of over 95%.

That result is Align Technology’s, achieved by the founder before this company existed. It is not a performance claim about AntitrustPro.

The question you will be asked

The Department of Justice’s Antitrust Division frames an antitrust compliance program around three questions.

Is the program well designed?
Is it applied earnestly and in good faith?
Does it work in practice?

The DOJ Antitrust Division asks whether your program works in practice. That is the question this product is designed around — a statement of design intent against a public standard; it does not imply any endorsement by the Division.

Competition authorities are using AI-based screening tools, alongside whistleblower reward programmes, to detect and investigate cartels. Many corporate compliance programs still run on a binder and an annual module.

Who this is for

The in-house lawyer who owns antitrust at a $500M–$5B multinational with no antitrust team of their own, and who must keep a compliance program defensible and answer the board’s exposure question without a full-time staff of experts.

Role: Associate General Counsel or Senior Counsel who owns competition — or the compliance deputy who does, where no such role exists.
Company: $500M–$5B revenue, US-headquartered or US-listed, operating in more than one region.
Legal team: 5–50 lawyers, with zero or exactly one dedicated antitrust lawyer.

Above that size there is usually an internal antitrust team of several lawyers. Below it, there is usually no named antitrust owner and no budget line for it.

What AntitrustPro does

A change in the score points you at what changed in the inputs, because nothing in the method changed between runs.

To be exact about what that is and is not: we have not measured the score’s stability across re-runs, we do not claim the score is on a common calibrated scale across different business units, and we make no claim that the score responds to a compliance intervention. Those are measurement claims and we have not done the measurement.

AntitrustPro is built on a proprietary scoring model for antitrust risk, with a working prototype that has been built and tested.

A design property worth interrogating

The model has a hard stop: above a severity threshold the answer is CRITICAL and the expected-value arithmetic is bypassed entirely — so the maths cannot be talked into telling your business team that the crime is worth it. It is specified as non-negotiable, it is implemented, and it is tested.

This is a statement about how the model is built. It is not a prediction about anything in your business.

It does not replace your antitrust firm, and you should not use it that way. Keep your firm for the hard calls: what a firm gives you is answers to the questions you bring, and the risk you did not think to ask about never gets priced. A standing score covers the surface rather than the queue.

If you already run a legal AI platform, keep it: it makes your lawyers faster on the work in front of them. It does not tell you where your antitrust exposure is.

The method — and what it is not

Everything in this section is here because your antitrust counsel will ask, and because you would otherwise find it out in diligence. Each limit sits with the claim it limits.

The scoring model is specified in writing

A versioned specification with a changelog, a sign-off register closed before anything shipped, hard constraints a verifier enforces, a closed set of reason codes, and an evidence grade on every parameter — so the model states its own confidence in its own inputs.

And every parameter in the set that ships today is at the bottom grade — unverified legacy. The model’s own rules require that flagged in anything client-facing. Re-deriving them against cited sources is scheduled work, not finished work.

The specification is our trade secret. Beyond what is described here, we do not hand it over, we do not quote it, and we do not describe its internals.

An independent in-house implementation was built, and the defects were written down

We did not just write the method down and trust it. An independent in-house implementation was checked against the specification across every parameter row, and the exercise produced a defect list of two dozen findings, five of them blocking, which we wrote down rather than buried. That list is why we can tell you where the model is underdetermined instead of waiting for you to find out.

That proves the implementation does what the specification says. It proves nothing about whether the specification is right about the world. Those are different claims and we are not going to let the first one do the second one’s work.

The protocols were pre-registered

We pre-registered a validation protocol that states in advance what would prove the model wrong — written before any data was scored.

That protocol tests one thing: whether a score ranks notified EU merger cases by Commission enforcement outcome. It does not test conduct exposure, it does not cover the US, and it says nothing about compliance programmes. So we do not offer it as evidence for the conduct-risk scoring in this product.

There is also a pre-registered protocol for validating the conduct model itself — written before any outcome data exists, fixing the hypotheses, the thresholds and the analysis plan in advance, including the number that would falsify the model. It is also explicit that one of the three things you might mean by “validation” — whether using the model produces better decisions — cannot be tested retrospectively at all, so we do not claim it.

A protocol is not a result. It describes a test we have not run, because the outcome data does not exist.

What we are not claiming

We have no validated accuracy figure for AntitrustPro’s model, and we will not quote you one. No validation of the model has been run: no dataset has been scored and no performance result about it exists, from any source. There is no outcome dataset to measure against — no penalties, no dispositions — so the figure is not unflattering, it is uncomputable.

We have no customers yet and no reference to offer you. That is a real cost of being early.

We are not your privileged adviser, and you should assume the output is not privileged. AntitrustPro is a professional-use tool for lawyers and compliance professionals. It does not give legal advice and does not create an attorney-client relationship; your lawyers review and own the output. Assume the scored output is an unprivileged corporate record, and design for that: scoped inputs, minimal retention, and questions that need a privileged answer routed to your counsel.

What the lab is working on

The research exists so that the method behind the products can be judged; the products are what that method is built into.

Research

Three programmes: antitrust risk measurement, legal privilege in legal AI, and compliance by design.

Their job here is to let you judge the method behind the products below.

Products

The stage words are the lab’s own description of how finished each product is.

AntitrustPro beta: Scores antitrust risk across your business by an explicit method, applied the same way each time.
Cadenza beta: The lab’s compliance training material generator.
Graplex beta: The lab’s legal knowledge graph: a structured map of statutes, cases and regulatory material.
AgentCert coming: The lab’s programme for building and evaluating legal AI agents against an organisation’s own playbook.
LegalEval coming: Evaluation tooling for legal AI systems.

The research

There is a completed white paper on compliance by design, and two further papers in draft.

Privilege, unauthorized practice of law and governance risk in legal AI are things we have researched — in depth, and in writing — not things we hope nobody asks about.

The specification and the conduct-model protocol are trade secret; the papers are drafts. This page describes the method and states its limits; it does not distribute the work.

Who built it

Linh Duong has been on your side of the table for 15+ years: General Counsel of a private equity fund, Regional General Counsel for a ~$1B Asia Pacific region, then Vice President & Associate General Counsel at a ~$4B medical device company — supporting a $50M+ global legal budget, and selecting and managing outside counsel.

At Align Technology, Linh Duong ran a global antitrust centre of excellence — 10+ experts across four regions — and designed a global antitrust compliance program and bespoke executive antitrust training.

In December 2024, Linh Duong built and launched Align’s global legal AI assistant for all 10,000+ of its employees, answering antitrust compliance questions at a self-reported accuracy of over 95% — curating its knowledge base and testing it before release.

That result is Align Technology’s, achieved by the founder before this company existed. It is not a performance claim about AntitrustPro.

An early in-house adopter of Checkbox, Brightflag, LitLingo, Everlaw and Ivo, and a Harvey pilot, Linh Duong advised the Checkbox and Brightflag product teams on what in-house lawyers need.

Next step

The first conversation is 45 minutes and it is about methodology, not software. Bring the part of your business you worry about most.

admin@legalsciencelab.com