TAR Workflow Summary

A practical workflow map for users who need to understand what happens from collection through validation.

011. Define scope and review goals
Identify custodians, sources, date ranges, claims/defenses, privilege concerns, and production obligations before choosing the review method.
022. Process, deduplicate, and normalize documents
Ingest ESI, remove duplicates where appropriate, extract text and metadata, identify families, and preserve defensible processing logs.
033. Create coding guidance
Write relevance, issue, privilege, and confidentiality guidance so reviewers train the model consistently.
044. Train with expert review
Use seed sets, judgmental samples, random samples, or continuous active learning queues; track who coded what and why.
055. Monitor model behavior
Watch richness, overturns, hot documents, unstable issue areas, privilege risks, and whether new batches are still finding relevant material.
066. Validate and document results
Use elusion testing, statistical sampling, QC review, and documented thresholds to support reasonable stopping decisions.
077. Produce, withhold, or escalate
Apply privilege review, redactions, confidentiality designations, family handling, production specs, and exception workflows before production.

Approach	Best used when	Watch-outs
TAR 1.0 / Simple Passive Learning	You can build a stable training/control set before ranking the collection.	Seed-set bias and slower iteration can be a problem.
Continuous Active Learning (CAL)	Review teams want the model to keep learning as reviewers code documents.	Requires disciplined coding guidance and monitoring.
Technology-assisted prioritization	You need to find likely relevant or hot documents early.	Prioritization is not the same as validated culling unless measured.
GenAI-assisted review	You need summarization, issue spotting, clustering, or review acceleration.	Validate outputs; do not assume AI explanations are correct.