TAR Protocol Boilerplate

DLC-M04-003 · Meet-and-Confer Ready

TAR Protocol —

Version

Date

This protocol describes the technology-assisted review methodology to be used for the responsive-set determination in the above-captioned matter. It is intended to be delivered to opposing counsel as part of the meet-and-confer under FRCP 26(f) or its state analog.

1. Workflow Selection

The producing party will use the following workflow:

TAR 1 · Predictive Coding. Reviewers code a control set drawn randomly from the corpus. The model is trained on the control-set coding and applied to the balance of the corpus once. A stable cutoff is established using precision-recall tradeoff.

TAR 2 · Continuous Active Learning. Reviewers code batches surfaced by the model, which retrains after each batch. Continues until an elusion-sample stopping criterion is met.

Hybrid. TAR 2 as the primary workflow with a defined subset (e.g., a specific custodian or date range) reviewed under TAR 1 or linear review.

Generative AI-Assisted first pass. Applied only to the subset defined below; subject to the exclusion rules in DLC-M04-002 and validated per §7.

2. Corpus & Scope

Corpus size at TAR start
Custodians in scope
Date range
Source types
Document types excluded from TAR

3. Tool & Model

Review platform
Model / algorithm
Model version
Vendor

4. Control Set (TAR 1) or Seed Set (TAR 2)

Sampling method	Simple random sample from the full corpus, drawn without replacement.
Control-set size (n)
Confidence & margin
Reviewer(s) coding control
Coding decision	Binary Responsive / Not Responsive. Privileged coded downstream.
Adjudication of disagreements
Delivery to opposing counsel	Aggregate metrics only (control-set richness). Individual document codes not shared.

5. Training & Iteration

TAR 1 · Predictive Coding:

Model trains on the coded control set.
Predictions applied to the balance of the corpus.
Cutoff established using precision-recall analysis; target recall , target precision .
Above-cutoff documents proceed to human review; below-cutoff excluded from responsive set (subject to elusion sample per §6).

TAR 2 · Continuous Active Learning:

Reviewers code documents surfaced by the model in batches of per session.
Model retrains after each batch.
Reviewer decisions are treated as authoritative training signal.
Coding continues until stopping criterion in §6 is met.

6. Validation & Stopping

Elusion sampling will be conducted:

TAR 1: after model prediction on the full corpus. Sample the excluded set.
TAR 2: after each of the last three batches when new-relevant rate falls below %.

Sample size: per DLC-M04-001, calculated for 95% confidence and ±3% margin of error on the elusion rate. Recorded in an addendum with signature and date.

Stopping criterion: upper bound of the 95% confidence interval for elusion rate is ≤ % of the excluded set.

7. Generative AI Component (if selected in §1)

Where GenAI is used in the workflow:

Applied only to the document subset defined in §2 and after application of exclusion rules in DLC-M04-002.
Every model output is spot-checked by a human reviewer before it influences a responsive-set coding call.
Model, prompt template, and any temperature/top-p parameters logged per document.
Model outputs are treated as recommendations, not classifications.
False-negative sampling of the excluded set is conducted at 95% / ±2%.

8. Cooperation & Transparency

The producing party will:

Provide the receiving party with the aggregate richness estimate from the control set within 10 business days of the control-set completion.
Provide the receiving party with the final elusion-sample results within 5 business days of the stopping decision.
Meet and confer if the receiving party identifies specific document types or custodians warranting additional review.
Not share individual document coding, work-product analyses, or internal QC results outside of properly-noticed motion practice.

9. Reviewer Quality Control

Reviewer training	Matter-specific responsiveness training; documented in-take call for every reviewer.
Second-pass rate	Minimum % of coded docs undergo second-attorney review.
Consistency check cadence	Every reviewer's inter-annotator agreement checked weekly.
Rate of exclusion for low agreement	Reviewers below % agreement retrained; below % removed.

10. Documentation

The following will be retained for the life of the matter plus statutory retention:

Control-set / seed-set document IDs and codes.
Model version, training parameters, and prediction outputs.
Every elusion-sample round with sample IDs, codes, and computed rates.
Every QC round with reviewer identities and outcomes.
Reviewer training materials (matter-specific and general).
Prompt templates and system prompts, if GenAI used.
Sign-offs from client counsel at each phase transition.

11. Definitions

Corpus: the set of documents subject to responsiveness determination.
Control set: a random sample of the corpus, coded by senior counsel, used to train the TAR 1 model.
Seed set: a starting set of coded documents provided to the TAR 2 model.
Recall: the fraction of truly responsive documents that the workflow surfaces.
Precision: the fraction of documents surfaced as responsive that actually are.
Elusion: the fraction of documents in the excluded set that are, on human review, responsive.
Family: a parent document (typically an email) and all of its attachments, grouped for review.

12. Signature

By signing below, the producing party attests to the accuracy of this protocol and its intent to apply it in the described matter.

Signature — Attorney of Record

Date

Printed Name

Firm / State Bar #

On use This boilerplate should be tailored to the specific case, jurisdiction, and any protective order in place. Consider case-specific stipulations from analogous matters (see DLC-M08 Case Law reference for TAR-related orders). Do not file without a partner review.