Read Me First: TAR Primer
Read Me First: TAR Primer
A high-level primer for understanding what Technology-Assisted Review is, how the workflow moves from goals to validation, and how prediction scores and metrics should be interpreted before relying on TAR in a real matter.
What TAR is, in plain English
What it is
- Technology-Assisted Review is a defensible review workflow that uses software predictions trained by human coding decisions.
- The goal is not to replace legal judgment; the goal is to prioritize, classify, validate, and document review decisions at scale.
- A TAR result is only as useful as the protocol, training signal, quality control, validation sample, and stopping rationale behind it.
Topics covered
- Computer-assisted review, continuous active learning, prediction scores, relevance calls, validation, proportionality, defensibility
Start with the process, not the product. TAR works when the review team can explain the goal, the coding rules, what the model learned, how the outputs were tested, and why the final stopping point was reasonable for the matter.
TAR Workflow at a Glance
- 01Set the goal
Decide whether TAR will prioritize likely relevant material, reduce review volume, support quality control, or help reach a defensible stopping point.
- 02Set the protocol
Write coding rules for relevance, issues, privilege, families, confidentiality, and reviewer escalation before training signals begin to shape the model.
- 03Train with human judgment
Use expert review, reviewer coding, seed examples, or continuous active learning so the system has examples of what matters.
- 04Predict and prioritize
Let the system score or rank documents so the team can review the most likely useful material earlier and monitor what the model is uncertain about.
- 05Review and quality check
Review high-value queues, resolve conflicts, sample uncertain or low-score areas, and track coding overturns that may signal unclear guidance.
- 06Validate the result
Use sampling, recall, precision, elusion, confidence intervals, and documented QC to test whether the workflow is meeting its goal.
- 07Document and move forward
Record the rationale, metrics, assumptions, limits, and next review phase, such as privilege review, redaction, production, or additional training.
How Prediction Score Bands Guide Review
| Prediction score | What it usually means | Useful review action |
|---|---|---|
80-100Prediction score range | The model sees strong similarity to documents reviewers have coded as relevant or important. | Review early, check for hot documents, privilege, key issues, and consistent coding. |
40-79Prediction score range | The model is less certain or the record has mixed signals. | Use targeted review, conflict checks, and additional training to clarify whether the middle is noise or a real issue area. |
0-39Prediction score range | The model predicts lower likelihood of relevance based on current training. | Do not ignore automatically; validate with sampling and elusion testing before deciding this material can be left behind. |
Metrics That Make TAR Defensible
Tests whether the workflow is capturing enough of the relevant population.
Shows how much review effort is being spent on material that is actually useful.
Estimates what relevant material may remain after a proposed stopping point.
Sets expectations for how hard it will be to find relevant material and how large samples may need to be.
Balances precision and recall when comparing model behavior, but should not replace legal judgment.
Shows uncertainty around sampling-based measurements so the team does not overstate precision.
| Metric | Formula | Purpose |
|---|---|---|
| Recall | relevant found / total relevant | Tests whether the workflow is capturing enough of the relevant population. |
| Precision | true relevant reviewed / predicted relevant | Shows how much review effort is being spent on material that is actually useful. |
| Elusion | relevant sampled from left-behind set / sampled left-behind set | Estimates what relevant material may remain after a proposed stopping point. |
| Richness | relevant documents / total population | Sets expectations for how hard it will be to find relevant material and how large samples may need to be. |
| F1 Score | 2 * precision * recall / (precision + recall) | Balances precision and recall when comparing model behavior, but should not replace legal judgment. |
| Confidence Interval | estimate +/- margin of error | Shows uncertainty around sampling-based measurements so the team does not overstate precision. |
Thresholds
| Metric | Minimum | Target | Excellent |
|---|---|---|---|
| Recall | 70 | 80 | 90 |
| Precision | 40 | 60 | 80 |
| Confidence Level | 90 | 95 | 99 |
What to Document Before You Stop
| Question | Why it matters |
|---|---|
| What population did TAR cover? | Scope, exclusions, deduplication, family handling, and date/custodian choices affect what the model could find. |
| Who coded training and QC documents? | Reviewer skill and consistency shape the signal the model learns from. |
| What score cutoff or stopping rule was used? | A cutoff is a decision point, not proof by itself; it needs validation and proportionality support. |
| What did the validation sample show? | Elusion, recall, confidence, and error review are the evidence that the workflow was reasonable. |
| What risks still need human review? | Privilege, confidentiality, issue nuance, redactions, and court/client obligations still need lawyer oversight. |
Where to go next
What it is
- Use the next section, Start Here: Additional Overview Resources, for the primary EDRM, FJC, and Grossman/Cormack source materials behind this primer.
- Use Validation, Metrics & Sampling when you need the math behind recall, precision, elusion, and confidence intervals.
- Use Practical Checklists & Protocol Prompts when you are preparing for a Rule 26(f) conference, vendor planning meeting, or internal TAR kickoff.
Topics covered
- Overview resources, validation, sampling, protocol planning, AI governance, case law, vendor documentation
The rest of this library is the primer's backup material: primary sources first, then legal authority, validation concepts, protocol prompts, and specific tool documentation.