E‑Discovery and Predictive Coding Standards

E‑Discovery and Predictive Coding Standards: Best Practices for Technology‑Assisted Review under Fed. R. Civ. P. 26 and 34

Technology‑assisted review workflows and predictive coding standards in modern litigation

Meta Summary: This playbook examines the legal and technical standards governing technology‑assisted review (TAR) in electronic discovery. Structured for legal practitioners, e‑discovery professionals, corporate counsel, and court administrators, it covers foundational concepts, Federal Rules of Civil Procedure 26 and 34, TAR methodologies, validation protocols, leading case law, best practices, and strategic integration with the rule of law.

Table of Contents

• Chapter 1: Foundations – Electronic Discovery and Predictive Coding
• Chapter 2: Federal Rules of Civil Procedure – Rules 26 and 34
• Chapter 3: TAR Methodologies, Validation, and Quality Control
• Chapter 4: Leading Case Law and Judicial Endorsement of TAR
• Chapter 5: Best Practices for TAR Protocols and Strategic Integration
• Related Topics
• FAQ
• References

Chapter 1: Foundations – Electronic Discovery and Predictive Coding

⬅ Back to Table of Contents

Introduction – The Scale and Complexity of Modern ESI

Modern litigation routinely involves electronically stored information (ESI) measured in terabytes, if not petabytes. Organisations generate vast quantities of data across email systems, messaging platforms, collaborative workspaces, cloud repositories, dynamic databases, and mobile devices. Traditional linear review—where human reviewers examine every document sequentially—has become economically and logistically impossible for large‑scale discovery. Technology‑assisted review (TAR), also referred to as predictive coding or computer‑assisted review, addresses this challenge by using machine learning algorithms to rank, prioritise, or classify documents according to their likely relevance to the legal issues in dispute, thereby reducing the volume of human review required while improving consistency and recall accuracy.

E‑discovery is governed by the Federal Rules of Civil Procedure (FRCP), particularly Rules 26 and 34, which prescribe the scope of discovery, the duty to preserve, the obligation to confer, and the requirements for producing ESI. The 2006 amendments to the FRCP formally recognised ESI as a discoverable category, and subsequent amendments have refined the framework for proportionality, preservation, and privilege. Rule 26(b)(1) limits discovery to matter that is relevant to any party’s claim or defence and proportional to the needs of the case, considering among other factors the importance of the issues, the amount in controversy, the parties’ relative access to relevant information, the parties’ resources, and the importance of discovery in resolving the issues. Rule 26(g) imposes a certification requirement that every discovery response, objection, or request is consistent with the rules, and sanctions may be imposed where the certification is made without reasonable inquiry.

Key Concepts in Technology‑Assisted Review

• Technology‑Assisted Review (TAR): AI‑powered review tools that analyse patterns in human reviewer decisions and apply those patterns across larger document sets to predict relevance. TAR encompasses predictive coding, continuous active learning (CAL), and emerging generative AI (GenAI) technologies.
• Predictive Coding (TAR 1.0 / Simple Active Learning): A supervised learning approach where a seed set of documents is coded manually for relevance, and the software extrapolates those decisions to the remaining corpus. The model is fixed after training.
• Continuous Active Learning (TAR 2.0 / CAL): An iterative workflow that starts with a small seed set and continuously revises the relevance model as reviewers code additional documents. The algorithm actively selects the most useful documents for further training, achieving higher accuracy.
• Generative AI (GenAI) Review: Uses large language models to comprehend, classify, and summarise documents based on natural language prompts, rather than binary classification. Outputs include narrative explanations of relevance reasoning.
• Recall (Sensitivity): The proportion of truly relevant documents correctly identified as relevant. High recall is critical where missing a relevant document carries high risk.
• Precision: The proportion of documents identified as relevant that are actually relevant. High precision reduces the burden of reviewing non‑relevant documents.
• Validation Sampling: Statistical quality control where random samples of documents are examined to estimate recall and precision, ensuring the TAR workflow meets performance thresholds.
• Proportionality (FRCP 26(b)(1)): The principle that discovery must be proportional to the needs of the case, informing TAR design decisions including data culling, search term refinement, and validation frequency.

Why Predictive Coding Matters for Disputes Strategy

The strategic significance of predictive coding extends well beyond cost reduction. Legal teams that master TAR can gain substantial advantages: faster document review enables earlier case assessment and more informed settlement decisions; consistent coding reduces the risk of missing critical evidence; and defensible use of TAR can reduce exposure to sanctions for over‑ or under‑production.

Conversely, poorly designed TAR workflows can backfire. In a 2025‑2026 multidistrict litigation against major US airlines, the producing party’s TAR process produced over 3.5 million documents of which only about 17% were responsive. The court described the situation as a significant “glitch” and reminded litigants that AI‑powered discovery must be subject to robust quality control. This incident underscores that TAR must be implemented with appropriate validation metrics, transparent reporting, and meaningful human oversight.

Recent studies confirm that TAR, when properly implemented, can reduce e‑discovery costs by 30% to 70% compared to linear review, while achieving recall rates comparable to or exceeding those of manual review. Courts have held that it is “inappropriate to hold TAR to a higher standard than keywords or manual review,” applying the same reasonableness and proportionality framework across all discovery methodologies.

Chapter 2: Federal Rules of Civil Procedure – Rules 26 and 34

⬅ Back to Table of Contents

FRCP 26 – Duty to Disclose, Meet and Confer, and Proportionality

Rule 26 establishes the foundational framework for discovery. Rule 26(a) imposes mandatory initial disclosures. Rule 26(f) mandates a meet‑and‑confer conference at least 21 days before a scheduling conference, where parties must discuss ESI issues, including the form of production. Effective December 1, 2025, amendments to Rules 26 and 16 mandate that parties specifically address privilege and work‑product issues at the initial Rule 26(f) conference, requiring early privilege protocols to streamline discovery.

Rule 26(b)(1) defines the permissible scope of discovery: “Parties may obtain discovery regarding any nonprivileged matter that is relevant to any party’s claim or defence and proportional to the needs of the case, considering the importance of the issues at stake, the amount in controversy, the parties’ relative access to relevant information, the parties’ resources, the importance of the discovery in resolving the issues, and whether the burden or expense outweighs its likely benefit.” Proportionality is thus a central constraint that must inform all TAR design decisions.

Rule 26(g) requires every discovery request, response, or objection to be signed by an attorney, certifying that after a reasonable inquiry the disclosure is consistent with the rules and not interposed for an improper purpose. Sanctions may be imposed for Rule 26(g) violations. This certification imposes a personal professional obligation on counsel to understand and validate the methodologies used in TAR workflows.

FRCP 34 – Requests for Production of ESI

Rule 34 governs requests for production of documents, ESI, and tangible things. It explicitly provides that discovery of ESI stands on equal footing with discovery of paper documents. Key provisions include:

• Rule 34(a)(1): A party may request production of documents or ESI “in the responding party’s possession, custody, or control.” Control is interpreted broadly.
• Rule 34(b)(1)(C): A request must specify “the form or forms in which ESI is to be produced.” If no form is specified, the responding party must state the form it intends to use.
• Rule 34(b)(2)(E): The responding party must produce documents “as they are kept in the usual course of business” or must organise and label them to correspond to categories in the request. For ESI, this ordinarily means producing native files with metadata.
• Rule 34(b)(2)(E)(ii): ESI must be produced “in a reasonably usable form,” meaning the requesting party can search, sort, and use the materials without extraordinary conversion efforts.

TAR‑driven identification of relevant documents must be calibrated to the specific requests under Rule 34. Overly broad TAR training that fails to align with the precise categories of information requested can result in massive over‑production of non‑responsive documents, as highlighted by the airline antitrust litigation.

The 2025/2026 Amendments – Privilege and Work Product

• Rule 26(f) Privilege Discussions: Parties must now discuss privilege and work‑product issues during the initial discovery conference, including the need for a privilege log, a non‑waiver order, and agreements concerning inadvertent production.
• Rule 16(b) Scheduling Order Requirements: Scheduling orders must address privilege‑related topics, including whether the parties will enter a non‑waiver order under FRE 502(d) and agreements regarding privileged ESI in TAR workflows.
• Integration with TAR: Where TAR is used for privilege classification, parties must address the risk of inadvertent production, the scope of clawback provisions, and documentation necessary to demonstrate a reasonable privilege review process.

These amendments reflect that early, cooperative planning—particularly where TAR is employed—is essential to avoid costly privilege disputes and potential waiver of protections.

Chapter 3: TAR Methodologies, Validation, and Quality Control

⬅ Back to Table of Contents

TAR Methodologies – SAL, CAL, and GenAI

Predictive Coding (TAR 1.0 / Simple Active Learning – SAL): A human reviewer codes a statistically representative seed set of documents. The software builds a predictive model and then scores each remaining document. SAL does not provide for iterative retraining. Advantages include simplicity and predictability. Disadvantages include the risk that the seed set may not capture full diversity of relevance patterns. Courts have routinely approved SAL workflows, beginning with the 2012 Da Silva Moore decision.

Continuous Active Learning (TAR 2.0 / CAL): Starts with a small seed set. The software then identifies documents most likely to be responsive and presents them for human coding. As each document is coded, the model is updated in real time. CAL typically achieves higher recall with less manual review and can adapt to heterogeneous datasets. Courts have endorsed CAL, allowing flexibility in stopping rules and validation methods, as seen in the In re Insulin Pricing Litigation.

Generative AI (GenAI) Review: Uses large language models trained on written instructions rather than example coding. Outputs include both a relevance determination and a narrative explanation. Hybrid workflows combine SAL, CAL, and GenAI sequentially. Courts have not yet provided comprehensive guidance on GenAI review, but many practitioners anticipate a landmark endorsement analogous to Da Silva Moore.

Validation Sampling – Elusion, Control Set, and Null Set

• Elusion Sampling: A random sample drawn from documents not produced as responsive. Manual review estimates the number of relevant documents missed (elusion rate). Low elusion indicates high recall.
• Control Set Sampling: A random sample drawn from the entire document corpus, reviewed manually irrespective of TAR output. Yields estimates of both precision and recall.
• Null Set Sampling: Sampling from documents classified as non‑responsive. Courts have rejected null‑set‑only validation when it risks opaque and unreliable recall calculations. In the Insulin Pricing multidistrict litigation, the court required sampling from the full document population, rejecting the defendant’s proposal to validate solely on a null set.

Key performance metrics include recall (completeness), precision (efficiency), F1 score (harmonic mean), and elusion rate. Parties should negotiate performance targets during the Rule 26(f) meet‑and‑confer. Agreed recall and precision thresholds provide a benchmark and reduce disputes.

Quality Control and Human Oversight

• Quality‑Controlled Training: Coding decisions used to train the TAR model should be subject to second‑level review. Courts have endorsed training using only quality‑controlled documents.
• Periodic Sample Review: Random samples of documents coded by the algorithm should be reviewed manually to detect model drift or degradation.
• Exception Handling: Edge cases—documents unique, ambiguous, or outside training distribution—must be flagged for manual review.
• Documentation: Every step of the TAR workflow should be documented: seed set composition, software version and parameters, validation sample sizes and results, and any deviations from the protocol.

Counsel bear an affirmative duty under Rule 26(g) to ensure the TAR process is reasonable and proportional. Ignorance of TAR methodology is not an excuse; lawyers must either develop sufficient technological competence or retain qualified experts.

Chapter 4: Leading Case Law and Judicial Endorsement of TAR

⬅ Back to Table of Contents

Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012) – The TAR Precedent

Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012). The first judicial endorsement of predictive coding. Magistrate Judge Andrew J. Peck held that “computer‑assisted review is an acceptable way to search for relevant ESI” and that where a producing party wants to use TAR, “the court should permit it.” The court emphasised that TAR is no less accurate than manual review and that opposition based on unfamiliarity with the technology should not block its use. This decision provided a template for TAR protocols and remains the foundational precedent.

View Da Silva Moore v. Publicis Groupe (Casetext)

Rio Tinto PLC v. Vale S.A. (S.D.N.Y. 2015) – TAR as “Black Letter Law”

Rio Tinto PLC v. Vale S.A., 306 F.R.D. 125 (S.D.N.Y. 2015). Judge Peck declared that “it is now black letter law that where the producing party wants to utilize TAR for document review, courts will permit it.” The court approved a joint TAR protocol and clarified that TAR should not be held to a higher standard than keywords or manual review. The decision also affirmed that a producing party may use either SAL or CAL; courts will defer to the producing party’s choice under proportionality and reasonableness.

View Rio Tinto PLC v. Vale S.A. (CourtListener)

In re Insulin Pricing Litigation (D.N.J. 2025) – Flexibility and Transparency

In re Insulin Pricing Litigation, No. 23-md-3080 (BRM) (RLS) (D.N.J. 2025). The court addressed a CAL protocol dispute. It held that responding parties are best situated to evaluate their own ESI preservation and production procedures, but transparency and cooperation are important considerations. The court rejected a rigid stopping point based solely on numeric thresholds and required that validation include sampling from the full document population, not merely a null set. It allowed the defendant to train its CAL model using only quality‑controlled decisions but required meet‑and‑confer when limiting training.

View In re Insulin Pricing Litigation analysis (Exterro)

Berger v. Graf Acquisition (Del. Ch. 2024) – Delaware Endorses TAR

Berger v. Graf Acquisition, LLC, 2024 WL 4541011 (Del. Ch. Oct. 21, 2024). Vice Chancellor Will explicitly endorsed TAR, stating that “statistics clearly show that computerized searches are at least as accurate, if not more so, than manual review.” The court held that “it is not up to the requesting party to block TAR if the producing party prefers it” and permitted TAR subject to transparency requirements. This decision is significant because the Delaware Court of Chancery is a highly influential business court.

View Berger v. Graf Acquisition (Delaware Courts)

In re Domestic Airline Travel Antitrust Litigation (D.D.C. 2025) – TAR Failure and Sanctions

In re Domestic Airline Travel Antitrust Litigation, No. 1:15-mc-1404 (D.D.C. 2025). A cautionary example. The producing party’s TAR process produced over 3.5 million documents, of which only about 17% were responsive. The court described it as a “glitch,” noting that the parties had agreed to a 75% minimum recall and reasonable precision, but validation metrics were not shared until the last minute. The plaintiffs were forced to request a six‑month extension to review irrelevant documents. The case underscores that TAR workflows must be continuously monitored, validated, and recalibrated, with timely disclosure of validation results.

View In re Domestic Airline Travel Antitrust Litigation analysis (Logikcull)

Chapter 5: Best Practices for TAR Protocols and Strategic Integration

⬅ Back to Table of Contents

Negotiating the TAR Protocol under Rule 26(f)

The Rule 26(f) meet‑and‑confer is critical. Parties should address:

• Data Sources and Custodians: Identify ESI sources (email, Slack, Teams, SharePoint, etc.) and custodians. Proportionality guides these decisions.
• TAR Methodology Selection: Agree on SAL, CAL, GenAI, or hybrid. The producing party’s choice should generally be followed unless unreasonable.
• Training and Validation Metrics: Specify sampling methods, target recall and precision rates, and validation frequency. Full‑population validation is generally required.
• Stopping Rules: Define how the parties determine when TAR has identified substantially all relevant documents. Rigid numeric thresholds alone are disfavoured.
• Disclosure and Transparency: Specify what information the producing party will disclose about its methodology—seed set composition, training documents, model parameters, validation results.
• Privilege Review and Clawback: Under the 2025 amendments, address privilege and work‑product issues, including TAR for privilege classification and clawback mechanisms under FRE 502(d).

Documenting TAR Workflows for Defensibility

• Preserving audit trails for every document: unique identifier, predicted relevance score, human vs. algorithmic classification, reviewer identity.
• Version control for models and parameters: record software version, algorithm parameters, and any retraining iterations with rationale.
• Formal validation reports describing sampling methodology, sample sizes, confidence intervals, recall, precision, and elusion.
• Declarations or certifications from counsel or experts confirming compliance with FRCP 26(g) and reasonableness of the TAR process.
• Change logs for any modifications to the TAR protocol, including reasons and effects on recall/precision.

Strategic Integration – Early Case Assessment, Settlement, and Cross‑Border

Early Case Assessment (ECA): Data mapping, proportionality analysis, pilot TAR sampling, and technology selection. ECA reduces discovery costs by 30–50% and avoids last‑minute crises.

Leveraging TAR Results: Analytics identify key witnesses and hot documents, assess case strength, support Daubert challenges, and prepare trial exhibits. TAR databases serve as the foundation for deposition and trial preparation.

Cross‑Border Considerations: GDPR, LGPD, PIPEDA, and data residency requirements may restrict data transfer. TAR processes may need to be deployed locally. Blocking statutes in some countries prohibit disclosure of ESI in response to foreign discovery orders. Courts in England, Ireland, and Australia have endorsed TAR; the UK’s ILTA Active Learning Best Practice Guide provides detailed guidance for TAR under Practice Direction 57AD.

The Future – Generative AI and Emerging Standards

Generative AI is moving from novelty to regular use. Many practitioners believe a definitive judicial opinion on GenAI discovery—“Judge Peck’s GenAI moment”—is approaching. Courts have already ordered production of user prompts and AI output logs in major copyright litigation. Recent Canadian and US decisions have sanctioned lawyers who submitted AI‑generated fabricated legal citations, underscoring the risks of relying on GenAI without human verification. Courts are also beginning to use AI for document review and drafting, which will influence the standards expected of parties.

Legal teams should monitor these developments and update TAR protocols accordingly. Those who embrace GenAI thoughtfully, with robust quality control and transparency, will be well‑positioned to achieve efficiency gains while maintaining defensibility.

Related Topics

⬅ Back to Table of Contents

The following topics expand this playbook into related areas of e‑discovery, technology law, and litigation practice:

• Algorithmic Accountability in Judicial Decisions: How courts are reviewing AI‑assisted sentencing and risk assessments under due process and equal protection standards.
• Data Privacy and Cross‑Border Discovery: GDPR, CCPA, Schrems II, and other privacy regimes affecting e‑discovery in multinational litigation.
• Sanctions for E‑Discovery Misconduct: Standards under FRCP 37(e) (loss of ESI) and Rule 26(g) (certification violations), including recent case law.
• Legal Ethics and Technology Competence: The obligation of lawyers to understand and supervise e‑discovery technologies under state ethics rules.
• Forensic Data Collection and Chain of Custody: Best practices for preserving ESI in a forensically sound manner, including metadata preservation and hashing.
• Proportionality in Discovery: Applying FRCP 26(b)(1) proportionality factors to e‑discovery, including cost‑shifting and staged discovery.
• Non‑Waiver Orders and FRE 502(d): Crafting effective non‑waiver orders to protect privilege in the event of inadvertent production, particularly when using TAR for privilege review.
• ILTA Active Learning Best Practice Guide: Practitioner‑led framework for responsible and effective use of TAR in disclosure, endorsed by senior UK judiciary.

FAQ

⬅ Back to Table of Contents

What is the difference between TAR 1.0 and TAR 2.0 (CAL)?

TAR 1.0 (simple active learning) uses a static seed set to train a model applied to the corpus without further retraining. TAR 2.0 (continuous active learning) starts with a small seed set and continuously refines the model as reviewers code additional documents, with the algorithm actively selecting the most useful documents for training. CAL generally achieves higher recall with less manual effort.

What level of transparency is required when using TAR?

Courts require sufficient transparency for the opposing party to evaluate the adequacy of production. This typically includes disclosure of the TAR methodology, number and composition of training documents, validation sample sizes and results, recall and precision estimates, and any changes made during the process. Proprietary trade secrets can be protected by confidentiality orders or attorney’s‑eyes‑only designations. Full‑population validation is generally required, not merely null‑set validation.

Can I use TAR for privilege review?

Yes, but risks are higher. If a privileged document is misclassified and produced, privilege may be waived unless a clawback agreement (FRE 502(d)) is in place. Best practices include human‑in‑the‑loop validation for documents flagged as possibly privileged, quality control reviews of the privilege‑negative set, and a robust non‑waiver order negotiated before using TAR for privilege review.

What happens if the requesting party refuses to agree to a TAR protocol?

The producing party may still use TAR as long as the workflow is reasonable and proportional. As held in Berger v. Graf Acquisition, “it is not up to the requesting party to block TAR if the producing party prefers it.” The producing party can file a motion seeking approval or proceed and defend its reasonableness if challenged. Cooperation is strongly encouraged; unilateral implementation without disclosure may lead to disputes.

How often should validation be performed?

Validation should be performed periodically throughout the TAR process. For large, dynamic datasets, continuous validation is appropriate. For smaller, static datasets, validation at the beginning and end may suffice. At a minimum, validate before production begins, after any material changes to the model, and at regular intervals as agreed in the protocol.

Do generative AI review tools require judicial approval?

No explicit judicial approval is required if the production complies with FRCP 26 and 34. However, because GenAI review is newer, practitioners may seek guidance or approval in high‑stakes matters. Many believe a landmark GenAI endorsement is approaching, analogous to Da Silva Moore.

References

⬅ Back to Table of Contents

Verified source links (embedded – all confirmed live):

• Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012) – Casetext
• Rio Tinto PLC v. Vale S.A., 306 F.R.D. 125 (S.D.N.Y. 2015) – CourtListener
• Berger v. Graf Acquisition, LLC, 2024 WL 4541011 (Del. Ch. 2024) – Delaware Courts
• In re Insulin Pricing Litigation (D.N.J. 2025) – Exterro case analysis
• In re Domestic Airline Travel Antitrust Litigation (D.D.C. 2025) – Logikcull analysis
• 2025 Amendments to FRCP 26 and 16 – eDiscovery Watch (GT Law)
• Federal Rule of Civil Procedure 26 – Cornell LII
• Federal Rule of Civil Procedure 34 – Cornell LII
• The Sedona Conference – Publications on E‑Discovery and TAR
• ILTA Active Learning Best Practice Guide (endorsed by Master of the Rolls, 2025)
• Dinsmore & Shohl – 98.2% recall in contract review (Everlaw case study)
• How Predictive Coding Makes E‑Discovery More Efficient – Thomson Reuters
• What Is Technology‑Assisted Review – TrustArray
• The Evolving Landscape of AI in Defensive Document Discovery – Quinn Emanuel (2026)
• What We Learned from 2025 – JD Supra (2026)
• Defensibility Considerations Around AI – Barbri (2026)