Technology-assisted review (TAR) is only as defensible as the validation protocol behind it. A well-designed validation protocol is what transforms a statistical ranking exercise into evidence a court, regulator, or opposing party will accept. As of September 2026, the rules governing disclosure and discovery — including the UK Disclosure Pilot under PD 57AD and the US Federal Rules of Civil Procedure — have not been rewritten for generative AI, but the practical baseline for what constitutes reasonable validation is moving quickly, driven by case law, vendor case studies, and the growing use of large language models in review workflows. This guide sets out the definitive best practices for structuring, executing, and documenting a TAR validation protocol.
Start With the Direct Answer: What a Validation Protocol Must Do
Also worth reading: How Should Legal Teams Measure eDiscovery Validation Performance in 2026? · What Are the Definitive AI eDiscovery Validation Protocols for 2027? · How are AI eDiscovery and legal technology trends reshaping medical cannabis litigation in 2026?
A TAR validation protocol must answer one question with statistical and documentary evidence: did the review process find the responsive or relevant documents it was supposed to find? Best practice requires three linked components. First, a defensible training and control-set methodology, typically a statistically valid random sample (often 500 to 2,000 documents) coded by senior reviewers to establish a golden baseline. Second, elusion testing — sampling the discard pile (documents the system classified as non-responsive) to estimate recall of privileged or responsive material left behind. Third, iterative quality control during the active review, including periodic re-sampling and reconciliation of reviewer disagreements with the model's classifications.
The widely cited Da Silva Moore decision (2012) accepted TAR with a control set of roughly 1,499 documents and a recall target discussed in the range of 75 percent. Since then, many practitioners treat 75 to 80 percent recall with high precision as the practical floor for defensible production, though the right target depends on case value, the prevalence of responsive material, and the sanctioning court's tolerance. What has changed by 2026 is that courts and commentators increasingly expect parties using generative AI tools to explain not just recall statistics but how the model was prompted, configured, and tested — a higher documentation burden than classic TAR.
Why Validation Matters More Now That Generative AI Has Entered Review
Classic TAR platforms (linear support vector machines, logistic regression classifiers) produced explainable relevance rankings, and validation protocols evolved around them. Generative AI review works differently: documents are evaluated against prompts and rubrics, sometimes in a single pass with no iterative training at all. Recent case studies published through JD Supra in 2025 and 2026 describe practical testing of GenAI review against manual review baselines, and the recurring finding is that performance varies sharply with prompt quality, document type, and prevalence. A validation protocol designed for a trained classifier does not automatically transfer to a prompt-driven system.
This matters because the legal rules have not yet changed. Hogan Lovells Cadwalader's analysis of generative AI in UK disclosure makes the point directly: the Civil Procedure Rules and PD 57AD are silent on AI, but the reasonable and proportionate search standard is being reinterpreted in practice. A party cannot argue that AI review is per se defensible; it must show that the specific deployment was tested. Validation is the bridge between a novel technology and the established obligation to conduct a reasonable search. Without it, a party faces the worst of both worlds — the cost of new technology and the indefensibility of an untested process.
Building the Control Set: Sample Sizes, Coding Standards, and Timing
The control set is the foundation of every validation protocol, and sloppy construction here invalidates everything downstream. Best practice calls for a simple random sample drawn from the full review population before any training begins, with sample size justified by confidence-interval math rather than round numbers. A 95 percent confidence level with a ±3 to ±5 percent margin of error typically requires roughly 380 to 1,000 documents; larger samples of 1,500 to 2,000 reduce variance when prevalence is low. Documents must be coded by experienced subject-matter reviewers using a written protocol with defined decision rules, and disagreements between coders should be measured with an inter-rater agreement statistic (Cohen's kappa of 0.7 or higher is a common benchmark) before the set is used.
Timing matters as much as size. The control set should be sealed — coded and frozen — before model training, so that it functions as an honest yardstick rather than a set the model has effectively memorized. When generative AI tools are used without iterative training, a control sample coded by humans still serves as the benchmark against which the AI's classifications are compared. In low-prevalence populations (under 5 percent responsive), practitioners should oversample edge categories and consider stratified sampling, because a simple random sample may contain too few responsive documents to measure recall with any confidence. Document every sampling decision in the validation memo; the paper trail is part of the defensibility.
Elusion Testing and Recall Estimation: The Core Metrics
Elusion testing — sampling the discard pile after review completes — remains the single most persuasive validation artifact. A random sample of several hundred to 2,000 documents from the non-responsive population, coded blind by senior reviewers, yields an elusion rate. Multiply the elusion rate by the discard-pile size to estimate the number of missed responsive documents; the binomial confidence interval around that estimate tells you the plausible worst case. A discard pile of 500,000 documents with a 0.5 percent elusion rate plus or minus 0.3 percent at 95 percent confidence implies between roughly 1,000 and 4,000 missed documents — a number the producing party should be prepared to explain and, if necessary, remediate.
Recall and precision should be reported together, never recall alone. A protocol that achieves 90 percent recall by flooding the review queue with junk may be worse than an 80 percent recall at 70 percent precision, depending on the cost of review and the sensitivity of the matter. Best practice also distinguishes between documents with zero relevance signal and documents with near-miss scores, and samples both separately. With generative AI review, add a third test: consistency testing, in which the same documents are re-submitted to the model to check whether classifications are stable. Instability — the same document flipping from responsive to non-responsive across runs — is a red flag that prompts or model temperature settings need adjustment before the review can be trusted.
Comparing Validation Approaches Across Review Models
Validation protocols differ materially depending on whether you run classic TAR, continuous active learning (CAL), or generative AI review. The table below summarizes the practical differences.
| Feature | Classic TAR (SPL/CAL) | Generative AI (LLM) Review | Hybrid AI + Human QC |
|---|---|---|---|
| Control set requirement | Mandatory, pre-training, 500–2,000 docs | Required as benchmark, even without training | Mandatory, pre-deployment |
| Elusion testing | Standard, post-production | Essential; model may miss semantically odd documents | Standard, plus model-vs-human diff analysis |
| Recall target | 75–85% typical | No settled norm; test against manual baseline | 80–90% with documented rationale |
| Primary risk | Training drift, seed-set bias | Prompt instability, hallucinated rationales | Integration errors, over-reliance on AI |
| Documentation burden | Moderate | High: prompts, versions, configs, test results | High: both tracks plus reconciliation logs |
| Cost profile | Sampling + senior review time | Testing cycles + prompt engineering | Highest, but strongest defensibility |
Common Mistakes That Undermine Validation Defensibility
The most frequent failure is coding the control set after or during training, which lets the model shape the yardstick. The second is using convenience samples — documents pulled from custodians known to be hot — instead of random samples, which inflates apparent recall. Third is declaring the review complete based on a subjective feeling of diminishing returns rather than a measured stability metric; best practice with CAL is to stop only when a full new training round changes classifications on a negligible share of documents and projected recall gain falls below a pre-set threshold, such as 1 percent.
Other recurring errors include skipping elusion testing entirely (a sure way to invite a spoliation-style challenge), reporting recall without confidence intervals, and failing to version-control the model and prompts. With generative AI tools, parties have been caught out when a vendor silently updates the underlying model mid-matter, invalidating earlier test results. Lock the model version, re-run validation after any change, and record dates. Finally, do not conflate privilege review validation with responsiveness validation — privilege elusion requires a stricter protocol, often 100 percent second-pass review of near-miss documents plus sampling of the discard pile, because the consequences of a miss are waive-by-injection risk under FRE 502.
When to Act: Sequencing Validation in the Matter Lifecycle
Validation is not a post-production formality; it is planned at the outset. During the meet-and-confer or Rule 31.5 / PD 57AD disclosure discussions, decide whether you will disclose the use of AI-assisted review, agree or contest search-term overlays, and set the negotiation posture on validation transparency. US courts have generally held that a producing party need not disclose its methodology absent a showing of deficiency, but volunteering a high-level protocol description early is increasingly common and can preempt motion practice. In the UK, the disclosure obligations under PD 57AD require a reasonable and proportionate search, and the AD's assessment is easier when the methodology — including validation — is documented at the Disclosure Review Document stage.
Operationally, build the control set within the first one to two weeks of processing, run interim validation checkpoints at roughly 25, 50, and 75 percent completion of active review, and complete elusion testing before certifying production. Budget two to four weeks for validation activities in a mid-sized matter (100,000 to 1 million documents). If you have already produced without validation, the second-best time to act is now: retrospective elusion sampling and a supplemental memo can often cure or at least mitigate the deficiency before it becomes a motion to compel further discovery.
Cost, Budget, and What Drives the Price of a Defensible Protocol
Validation costs are dominated by senior reviewer time. At blended senior-review rates of roughly $150 to $400 per hour (or £100 to £300 in the UK), a control set of 1,000 documents coded at four to six documents per hour runs $8,000 to $25,000 per coding pass, and elusion testing of 1,500 documents adds a similar amount. Add project management, statistical analysis, and validation memo drafting, and a full validation program for a matter of 500,000 documents typically runs $40,000 to $120,000 — a small fraction of total review spend, which can exceed $1 million on matters of that size. Skimping on validation to save 5 percent of the budget while exposing the matter to a motion to compel re-review is a poor trade.
Generative AI review changes the economics in both directions. Per-document AI review costs have fallen to fractions of a cent to a few cents per document on some platforms, so a full second AI pass as a quality-control layer is often cheaper than sampling alone. But the cost of prompt engineering, test cycles, and expert testimony if challenged is real and should be budgeted. The rational allocation: spend generously on validation and modestly on redundant AI passes, not the reverse.
The Bottom Line: Documentation Is the New Defensibility
The through-line across the 2025–2026 commentary — from JD Supra's eDiscovery case-law tracking to Hogan Lovells' UK disclosure analysis — is that judges are not yet imposing AI-specific rules, but they are asking harder questions, and parties with a documented validation protocol answer them. A defensible protocol in 2026 means: a pre-training random control set with recorded inter-rater agreement, locked model versions and prompts for AI tools, interim quality checkpoints, elusion testing with confidence intervals, a written validation memo, and a rational, stated recall target tied to proportionality. Courts have accepted TAR for over a decade on exactly this kind of showing; generative AI will earn similar acceptance through the same discipline, not through vendor marketing. Treat validation as a first-class workstream with its own budget, timeline, and named owner, and the technology choice matters far less than the proof behind it.