A plain-language glossary for TAR, sampling, and defensibility.

The share of documents in a population that are actually responsive.

Why it matters: Low richness changes sample sizes, expectations, and review cost.

A sample-based estimate of what responsive material may remain in the unreviewed set.

Why it matters: It informs stopping decisions but never proves a perfect review.

A fixed reference sample used to compare workflow behavior over time.

Why it matters: Helpful in some workflows, but weak design can mislead.

The share of truly responsive documents that were found by the workflow.

Why it matters: Often the headline metric in defensibility discussions.

The share of documents marked responsive that were actually responsive.

Why it matters: It affects cost and reviewer burden more than completeness.

The uncertainty introduced when a conclusion is based on a sample instead of the full population.

Why it matters: It is the reason confidence intervals exist.

A document review process that uses machine learning to predict which documents are relevant, rather than relying solely on linear human review.

Why it matters: It is the umbrella term for the workflows the rest of this glossary describes.

An earlier name for TAR, referring to using a trained model to code documents as responsive or not.

Why it matters: Courts and protocols often still use this term interchangeably with TAR.

A TAR approach where the model is continuously retrained on reviewers' coding decisions, constantly reprioritizing the highest-value documents to review next.

Why it matters: It is the dominant modern workflow and changes how you decide when review is complete.

A one-shot TAR workflow: train a model on a fixed sample, apply it to the full set, then validate the result.

Why it matters: It produces a clean statistical story but assumes the collection is stable.

A continuous TAR workflow (usually CAL) where training and review happen together until a stopping point is reached.

Why it matters: It finds responsive documents faster but shifts the defensibility question to the stopping decision.

The proportion of a population that is responsive, often used interchangeably with richness.

Why it matters: It drives every sample-size and recall-uncertainty calculation.

The number or proportion of responsive documents recovered by a step in the workflow.

Why it matters: It helps you judge whether a culling or prioritization move actually captured the responsive material.

The harmonic mean of precision and recall, expressed as a single number between 0 and 1.

Why it matters: It is a convenient summary when you need to balance completeness against purity.

The proportion of documents in the discarded set that are actually responsive, estimated from a sample.

Why it matters: It quantifies what a culling decision left behind and is central to defensibility.

How far down a ranked list you must review to reach a target recall level.

Why it matters: It connects a recall goal to the actual review volume and cost it implies.

A range around a sample-based estimate that is likely to contain the true value, at a stated confidence level.

Why it matters: A recall or elusion estimate without an interval is only half a number.

The half-width of a confidence interval; how far the estimate may sit from the true value.

Why it matters: It is the precision you are buying with a given sample size.

The number of documents drawn for a statistical estimate, set by the target precision, confidence level, and how rare the trait is.

Why it matters: There is no universal number; it follows from the assumptions you state up front.

The set of documents predicted non-responsive and set aside, also called the discard set.

Why it matters: It is the population an elusion test samples from to check what was missed.

The initial set of coded documents used to begin training a TAR model.

Why it matters: Its composition shapes early model behavior and is sometimes negotiated between parties.

The documents whose human coding decisions are used to teach the model what is responsive.

Why it matters: How it is built and documented is a frequent point of scrutiny.

The point in a continuous learning workflow where more training stops meaningfully changing the model's rankings.

Why it matters: It is a common, though not the only, signal that review can wind down.

The pre-defined rules that determine when a TAR review is considered complete.

Why it matters: Defensibility often turns on whether these were set in advance and met.

A random sample reviewed to estimate the quality of a completed review, such as its recall.

Why it matters: It is the evidence behind a claim that the review was reasonable.

The proportion of coding decisions changed when documents are re-reviewed during quality control.

Why it matters: A high overturn rate signals inconsistent coding that can undermine results.