The TAR 1 Reference Model: What Sedona Conference's New Framework Means for GenAI Review

In March 2024, The Sedona Conference published a paper that quietly reshaped how practitioners should think about AI-assisted review. Written by Tara Emory, Jeremy Pickens, and Wilzette Louis of Redgrave Data, the TAR 1 Reference Model argues that generative AI does not require a new review framework at all. It fits inside the same defensible process litigation teams have relied on for over a decade. That claim matters more now than when the paper was published, because generative AI review tools have moved from pilot projects to daily use in litigation support departments across Canada and beyond. Understanding the reference model gives your team a shared vocabulary for evaluating any GenAI review tool, including Claira, against a standard that courts and opposing counsel already recognize.

What TAR 1 actually is

Technology-Assisted Review, or TAR, was defined over a decade ago by Maura Grossman and Gordon Cormack as a process that uses subject-matter experts to train a computerized system, then extrapolates those judgments across a document population. TAR 1 is the specific workflow within that broader category most relevant to first-pass review. It builds a predictive model, then applies that model to tag an entire population as responsive or not responsive. Historically, TAR 1 relied on discriminative machine learning algorithms such as support vector machines or logistic regression classifiers, trained on thousands of human-coded examples. The Sedona paper's central insight is that a large language model, prompted correctly, functions as a supervised classifier too. The engine changes. The vehicle, meaning the overall TAR 1 process, stays the same.

That distinction is not academic. It means the review protocols, defensibility arguments, and quality metrics your team has already built around TAR 1 do not need to be reinvented for generative AI. They need to be applied consistently to a new kind of engine.

The five steps of the reference model

The paper lays out five steps that apply regardless of which predictive algorithm sits underneath: scope, label control set, iterate model, classify, and validate. Scope means assembling the document population and having attorneys define the boundaries of responsiveness before any tagging begins. Label control set means pulling a random sample, reviewing it by hand, and holding it aside as an independent benchmark the model never trains on. Iterate model is where the real work happens: selecting additional training examples, encoding that information into the model or the prompt, applying the updated version to the control set, and evaluating whether precision and recall have improved enough to justify another round. Classify applies the finished model to the full, untagged population. Validate, which the paper notes is optional but underused, samples the classified output a final time to confirm the extrapolation held.

What changes between discriminative TAR 1 and GenAI TAR 1 sits entirely inside the iterate model step. Discriminative systems learn from labeled examples fed directly into an algorithm. GenAI systems learn from a prompt that a human writes, tests, and refines in natural language. The Sedona authors are careful to note that the prompt writer should be a different person than whoever reviewed the control set, so the control set stays an unbiased check on the model rather than a source of leaked answers. That separation of duties is worth building into your own review protocols, whichever engine you use.

Where GenAI changes the tradeoffs

The paper is candid about what remains unproven. Discriminative TAR 1 produces finely graded scores, which lets a review team choose a cutoff point that balances recall and precision with some precision. GenAI TAR 1 often classifies at coarser gradations, which can force a choice between a cutoff with very high recall and low precision, or the reverse, with little room in between. GenAI also currently costs more per document reviewed and runs more slowly than discriminative models, though the authors expect both gaps to narrow. What GenAI offers in return is flexibility: the same pass that tags a document as responsive can also flag privilege risk, extract key facts, or summarize content, tasks that used to require separate workflows entirely.

For litigation support teams, this is the practical question the reference model helps answer: not whether GenAI review is defensible in principle, but whether a given tool implements the TAR 1 steps rigorously enough to make it defensible in practice. A tool that skips control sets, blends training data with production population, or cannot show its scope of responsiveness is not offering GenAI TAR 1. It is offering an unsupervised shortcut wearing TAR 1's language.

Where this leaves practitioners

Claira was built around the assumption that the reference model would become the standard, not a novelty. Every scope decision in a matter starts with Case Context, where your team defines the background, issues, and responsiveness criteria a matter turns on, the same work the Sedona paper describes as the scope step. That context carries through every subsequent pass, so the definition of responsiveness stays fixed the way the reference model requires, rather than drifting as reviewers move through the population. When it comes time to apply that definition across the full document set, whether through objective coding or a bulk pass tagging documents for responsiveness, the underlying steps map directly onto classify and, where a team chooses to run one, validate.

We have written before about how Sedona Canada's principles are reshaping expectations for what a defensible AI review process looks like north of the border, and this reference model gives that expectation a concrete shape. Proportionality and cooperation are easier to demonstrate when your workflow can point to a named, well-documented step for every stage of the process, from initial scoping through final validation.

The larger point of the Sedona paper is reassuring rather than alarming. Generative AI does not ask litigation support teams to abandon the frameworks that have made TAR defensible for years. It asks them to hold GenAI tools to those same standards, and to be skeptical of any tool that cannot show its work at each of the five steps. If your team is evaluating what a rigorous GenAI TAR 1 workflow should look like inside Nuix Discover, book a walkthrough with our team and we will walk through how Claira maps to each stage of the reference model, control sets included.