by: Rachel Fernandez, Director of Sales Engineering
Once your initial data scoping is complete and you have a starting population, the next question is whether your matter is a good candidate for a TAR workflow and if so, which one. Choosing the wrong workflow for your data set is one of the more avoidable sources of cost overrun in eDiscovery. This article covers what to evaluate before you commit, with practical guidance on how those decisions play out in both Relativity and Everlaw.
TAR Foundation
There are multiple types of TAR workflows, but the two most common are TAR 1.0 and TAR 2.0. TAR 2.0 is also often referred to as Continuous Active Learning (CAL).
TAR 1.0 uses a fixed training phase where a subject matter expert trains the model by coding a statistically significant seed set. These coding decisions are then applied to the rest of the unreviewed population. The machine learning algorithm ranks documents across the full data set and focuses its training on the documents that can best refine its predictions. This workflow is less about linear review and more about expediency in prediction. TAR 1.0 is still used today, though less often for machine learning purposes and more commonly as part of artificial intelligence tool workflows.
TAR 2.0, or Continuous Active Learning, works differently. The model ingests coding decisions from the review team continuously, allowing it to adapt and re-rank documents in real time. This methodology prioritizes the most likely relevant documents for review first, surfacing them to the queue on an ongoing basis as the model learns. CAL is more adaptable than TAR 1.0 and is generally the better fit for most modern litigation matters where large document populations need to be addressed efficiently.
Understanding which methodology fits your matter starts with an honest look at your data.
Evaluating Your Matter for TAR
There are several factors that contribute to a successful TAR project. Each should be weighed carefully against the others to ensure you are selecting the workflow that best fits your data set and the needs of your matter.
- Volume. TAR workflows are best suited for larger document populations. There is no fixed number that guarantees a good outcome, but generally the larger the set, the greater the efficiency gains. On smaller matters, the setup and management overhead may not justify the workflow, and a linear or search-term-driven review may be the more practical choice.
- Prevalence. Prevalence, sometimes called richness, refers to the percentage of relevant documents within your total population. This factor has a significant effect on how well a TAR workflow will perform. Low prevalence matters present a challenge because the model has less signal to learn from early in the process. Conversely, very high prevalence sets can sometimes diminish the culling gains that make TAR worthwhile in the first place. The most reliable way to establish your prevalence is to run a statistical sample early in the process. This gives you a baseline understanding of what your data contains and helps inform which workflow is most appropriate before you commit.
- Document Quality. TAR workflows depend on clean, extractable text. Evaluate the composition of your data set before selecting a workflow. Emails, PDFs, and Word documents typically perform well within a TAR model. Spreadsheets, images, and non-standard file types with poor or out-of-context extracted text will often need a separate review track outside of the TAR model. Identifying these populations during scoping allows you to plan for them rather than discover them mid-review.
- Consistency of Coding. TAR workflows perform best when they are built on a clear and well-documented review protocol. Establishing a precise definition of relevance before the first document is coded helps both the review team and the model make consistent distinctions across the document population. Inconsistent coding decisions are one of the most common sources of model degradation during a CAL review and are far easier to prevent than to correct once the review is underway. Investing time in reviewer training and written guidelines before the project begins pays off throughout the lifecycle of the matter.
Platform Considerations
All platforms address TAR workflows differently and understanding how your tool manages and reports on your model is an important part of selecting the right environment for your matter.
- Relativity. Relativity’s Active Learning tool is one of the most widely used CAL implementations in the industry. It is built around a project-based structure where you define your review population, configure your coding layout, and allow the model to continuously surface and re-rank documents as reviewers code. Relativity produces exportable reporting on model performance throughout the review, and the coverage dashboard gives you real-time visibility into elusion rates and estimated remaining relevant documents. The platform supports multiple simultaneous Active Learning projects, which is particularly useful for complex matters where you are coding across several legal issues at once. This depth of configuration makes Relativity well suited for large, complex matters where defensibility requirements are rigorous and documentation needs to be thorough.
- Everlaw. Everlaw approaches predictive coding through a round-based training structure that provides natural checkpoints between model updates. Precision and recall metrics are surfaced visually between rounds, which makes it easier to communicate progress to supervising counsel or client teams who may be less familiar with TAR methodology. The interface is more accessible for teams with less TAR experience, and Everlaw’s collaboration features allow outside counsel and in-house reviewers to work within the same environment effectively. It is less configurable than Relativity but provides the core metrics and reporting necessary for a defensible workflow. Everlaw tends to be a strong fit for mid-sized matters where accessibility, transparency, and collaboration between review teams are priorities.
When selecting between the two platforms, the decision is less about which is superior and more about which fits your matter, your team, and your client’s expectations. Both are capable of producing defensible TAR workflows when managed thoughtfully and with adequate planning at the outset.
Platform Considerations
A successful CAL workflow is built on the decisions and foundations made well before the first document is ever coded. Custodian selection, culling discipline, a clear relevance definition, and an honest assessment of your data composition all feed directly into how well your model performs and how defensible your process will be at the end of the matter.
Partnering with experienced TAR consultants can make this process significantly more manageable. The Sandline team has deep experience in defensible TAR workflows and can help develop processes tailored to your litigation matters.