Implementing a robust technology assisted review tar guide protocol is critical for mitigating organizational risk, ensuring regulatory compliance, and streamlining defensible data discovery.
Technology-Assisted Review (TAR 2.0) vs Linear Review: Defensible AI Document Review
Introduction & Industry Evolution
Document review has historically represented the single largest cost component of eDiscovery, often consuming over 60% of total litigation budgets. Manual “linear review”—where teams of contract attorneys inspect documents one by one in date or custodial order—is slow, expensive, and subject to high human inconsistency and fatigue.
Technology-Assisted Review (TAR), commonly referred to as predictive coding, utilizes machine learning classification algorithms to analyze document text vectors and prioritize coding decisions based on attorney feedback. The evolution from TAR 1.0 (Simple Passive Learning using static upfront seed sets) to TAR 2.0 (Continuous Active Learning / CAL) has fundamentally improved review efficiency, precision, and courtroom defensibility.
This technical guide provides a detailed comparison between Linear Review, TAR 1.0, and TAR 2.0 (CAL), outlines the mathematical formulas used for recall/precision validation, details elusion rate sampling methodologies, and summarizes landmark judicial precedents approving TAR globally.
Architectural Comparison: Linear Review vs. TAR 1.0 vs. TAR 2.0 (CAL)
| Evaluation Dimension | Manual Linear Review | TAR 1.0 (Simple Passive Learning) | TAR 2.0 (Continuous Active Learning / CAL) |
|---|---|---|---|
| **Learning Engine** | None (Human manual evaluation) | Upfront training on static “Seed Sets” | Continuous real-time model retraining |
| **Training Requirement** | N/A | Requires expert senior attorney seed set coding | No control seed set required; learns from live review |
| **Workflow Adaptability** | Low | Cannot adapt if review scope or terms change | Dynamically updates as new documents or topics emerge |
| **Review Volume** | 100% of collected data set | Evaluates full set; stops at static rank cutoff | Focuses review on top-ranked relevant documents |
| **Defensibility** | High subjective error rate (~60% consistency) | Vulnerable to seed-set challenge by opposing counsel | Highly defensible with statistical elusion sampling |
Technical Deep Dive: TAR 2.0 Continuous Active Learning (CAL)
In a TAR 2.0 Continuous Active Learning (CAL) workflow, the machine learning classification algorithm (typically Support Vector Machines or Logistic Regression over term-frequency vector spaces) continuously re-evaluates the unreviewed document population after every batch of human attorney coding decisions.
[ Unreviewed Document Pool ]
||
\/ (Ranks highest probability relevant documents)
[ Active Review Queue (Human Attorney Review) ]
||
\/ (Coding Decision: Relevant / Non-Relevant)
[ Real-Time Model Re-Training Engine ] ==> Updates Text Vector Weights ==> Loops back to QueueMathematical Validation Formulas: Recall, Precision, & Elusion Rate
To establish legal defensibility in court proceedings, legal teams must statistically prove that the TAR process achieved an acceptable target Recall rate (typically 75% to 85% or higher).
1. Recall Formula
Recall measures the proportion of truly relevant documents in the population that were successfully identified by the TAR process:
$$\text{Recall} = \frac{\text{True Positives (Relevant Documents Found)}}{\text{True Positives} + \text{False Negatives (Relevant Documents Missed)}}$$
2. Precision Formula
Precision measures the proportion of identified documents that are actually relevant:
$$\text{Precision} = \frac{\text{True Positives}}{\text{True Positives} + \text{False Positives}}$$
3. F1-Score (Harmonic Mean)
The F1-score balances precision and recall into a single scalar performance metric:
$$F_1 = 2 \times \frac{\text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}}$$
4. Elusion Testing & Sampling
Elusion testing validates the unreviewed document collection (the documents deemed non-relevant by the CAL model). A statistically valid random sample (typically at a 95% confidence level with a $\pm 1\%$ to $\pm 2\%$ margin of error) is drawn from the discarded pool and manually reviewed to calculate the Elusion Rate.
$$\text{Elusion Rate} = \frac{\text{Relevant Documents Found in Sample}}{\text{Total Sample Size}}$$
Defending TAR 2.0 in Court: Overcoming Opposing Counsel Objections
Despite clear judicial endorsement, opposing counsel frequently raise objections regarding predictive coding defensibility. Legal teams can proactively dismantle these challenges by establishing rigorous documentation protocols:
1. Protocol Transparency vs. Work-Product Protection: While courts do not generally mandate sharing non-relevant training seed documents (as affirmed in Rio Tinto), providing full transparency regarding the mathematical validation methodology, target recall goals, and elusion sampling metrics satisfies court inquiry without compromising attorney work product.
2. Defending High-Volume Data Additions (Rolling Ingests): Unlike TAR 1.0 (which required restarting seed set training if new data was ingested), TAR 2.0 CAL dynamically incorporates new custodian mailboxes into the active vector space, automatically prioritizing newly ingested relevant documents based on existing model weights.
3. Validating Machine Learning vs. Manual Reviewer Variance: Documenting that human reviewer inconsistency in traditional linear review averages 30% to 40% due to subjective interpretation, whereas CAL machine learning classification maintains 100% deterministic mathematical consistency over vector feature spaces.
4. Seed Set Validation vs. Continuous Training: Establishing that TAR 2.0 completely eliminates the need for expensive upfront control sets or seed set negotiations between opposing parties, allowing review to begin immediately while the machine continuously learns from active attorney coding decisions.
Judicial Precedents & Global Acceptance
[!IMPORTANT]
**Legal Fact-Checking Gate (`VERIFY BEFORE PUBLICATION`):** Judicial acceptance of TAR varies by court jurisdiction and case facts. Always verify docket numbers and judicial standing orders prior to executing predictive coding stipulations.
1. United States
- Da Silva Moore v. Publicis Groupe (858 F. Supp. 2d 137 (S.D.N.Y. 2012)): Landmark opinion by Magistrate Judge Andrew Peck establishing that TAR is an acceptable and defensible method for searching ESI in federal litigation.
- Rio Tinto PLC v. Vale S.A. (306 F.R.D. 125 (S.D.N.Y. 2015)): Confirmed the superiority of TAR 2.0 CAL workflows and noted that transparency in training sets is not always required.
2. United Kingdom
- Pyrrho Investments Ltd v. MWB Property Ltd ([2016] EWHC 256 (Ch)): High Court decision explicitly approving predictive coding for eDisclosure under CPR Part 31.
3. Canada
- Air Canada v. WestJet Airlines Ltd (2006): Early guidance establishing the use of automated search technologies under Sedona Canada proportionality principles.
Recommended Internal Content Links
- Parent Hub: Document Review Master Hub
- Related Deep Dive: Generative AI & LLMs in eDiscovery
- Related Deep Dive: Early Case Assessment Strategy
- Contextual Solutions Page: Managed Document Review & TAR Services
Frequently Asked Questions (FAQ)
What is Continuous Active Learning (CAL) in TAR 2.0?
Continuous Active Learning (CAL) is a TAR 2.0 machine learning methodology where the predictive classification algorithm continuously updates document relevance scores in real-time as attorneys code documents, continuously feeding the highest-probability relevant documents to the review queue.
How do courts evaluate the defensibility of predictive coding?
Courts evaluate TAR defensibility by examining the qualifications of the operators, the transparency of the workflow, the statistical validation metrics (such as recall and precision), and the results of elusion sampling over the unreviewed data pool.
What sample size is required to validate TAR accuracy?
Sample sizes are calculated using statistical sampling formulas based on total population size, target confidence level (typically 95%), and margin of error (typically +/- 1% to 2.5%). Standard validation sample sizes range from 1,500 to 2,500 documents.
For additional technical frameworks and legal standards, reference official guidance at NIST Computer Security Resource Center and EDRM Official Frameworks.
