Build TAR Validation Report

Skill: Convert TAR/predictive-coding metrics into a defensibility validation report

Region: United States Category: Legal / eDiscovery Does: Takes the outputs and metrics of a Technology-Assisted Review (TAR / predictive coding) effort and assembles a validation report — seed/training detail, recall and precision estimates, elusion testing, and a defensibility statement — for disclosure to opposing counsel or the court and to satisfy ESI-protocol obligations. Authority: EDRM Technology-Assisted Review · The Sedona Conference TAR Case Law Primer / TAR perspectives · Da Silva Moore v. Publicis Groupe (SDNY 2012) · Rio Tinto plc v. Vale S.A. (SDNY 2016)

Courts accept TAR where the process is reasonable and validated, not where a particular tool is used — Rio Tinto framed the question as the reasonableness and transparency of the process. This report documents that the review met a defensible recall target via statistically sound sampling. Metric targets and disclosure scope are case- and protocol-specific (this is a Medium-confidence skill) — confirm against the ESI protocol and current case law before relying on it.

When this applies

• Closing out a TAR 1.0 (SAL/SPL) or TAR 2.0 (continuous active learning) review to demonstrate completeness before certifying substantial completion.
• Responding to an ESI-protocol requirement or an opposing-party/court request to validate the TAR process.

Input data required

Report structure

1. Scope & methodology — population, TAR method, tool, workflow narrative
2. Prevalence (richness) — estimated responsive rate with confidence level & margin of error
3. Training — seed selection, rounds, stabilization criteria
4. Cutoff determination — how the responsiveness threshold was set
5. Validation results:
     - Recall estimate (+ confidence interval)
     - Precision
     - Elusion test — responsive rate in the null/discard set (key completeness check)
     - Control-set performance
6. QC — reviewer-consistency / disagreement resolution
7. Defensibility statement — why the process was reasonable & proportional; what is disclosed
8. Appendices — sampling math, statistical assumptions

Build rules

• Recall is the headline metric (what share of all responsive documents were found); report it with a confidence interval, not a bare point estimate, and state the sampling design (confidence level, margin of error, sample size) that supports it.
• Run an elusion test on the discard/null set — the responsive rate among documents not produced for review is the most persuasive completeness evidence; a low elusion rate supports defensibility, a high one means the cutoff is too aggressive.
• Use statistically valid sampling (random control/validation samples sized to the target confidence/margin); document the math so the numbers are reproducible.
• Describe the workflow honestly — TAR 1.0 vs 2.0, seed selection, training rounds, stabilization — and how the cutoff was chosen; transparency is what Rio Tinto/Da Silva Moore reward.
• Scope the disclosure to the protocol — parties often disclose recall and methodology but not work-product seed coding; state what is and isn't disclosed and why.
• Don't overclaim: present estimates with their uncertainty; note limitations (e.g. low prevalence widening intervals).

Worked example (results summary)

Population: 480,000 docs (text-bearing). Method: TAR 2.0 (CAL), Tool vX.
Prevalence: 4.2% responsive (95% confidence, ±0.6%).
Training: judgmental seeds + CAL; 11 batches; stabilized when batch yield < 2 responsive/200.
Cutoff: rank score ≥ 0.55 → 27,400 docs produced for review.
Validation (95% confidence):
   Recall   = 82% (CI 78–86%)
   Precision= 61%
   Elusion test on the 452,600 discarded docs: responsive rate 0.7% (CI 0.4–1.0%)
Conclusion: recall target (≥80%) met; low elusion rate supports completeness. Methodology and
   recall disclosed to opposing counsel per ESI Protocol §5; seed coding withheld as work product.

Validation checklist

• TAR method (1.0 vs 2.0), tool/version, and workflow narrative documented
• Prevalence/richness estimated with confidence level and margin of error
• Recall reported with a confidence interval and the supporting sample design; precision reported
• Elusion test run on the null/discard set; responsive rate reported with CI
• Sampling is statistically valid and the math is reproducible (documented assumptions)
• Cutoff/threshold rationale stated; training/stabilization described
• Recall target (per protocol/agreement) met; limitations noted; no overclaiming
• Disclosure scope matches the ESI protocol; work-product boundaries stated

Last updated: 2026-05-31 — TAR validation metrics, recall targets, and disclosure obligations are case- and protocol-specific; confirm the statistical approach and what must be disclosed against the case ESI protocol, current FRCP, and TAR case law (Da Silva Moore, Rio Tinto) and Sedona Conference guidance before relying on the report.