Form Outline of a Detailed Plan for Predictive Coding

Outline of 12-Step Plan for Predictive Coding Review

1. Basic Numerics of the Project

a. Number and type of documents to be reviewed

b. Time to complete review

c. Software to be used for review

(1) Active Machine Learning features

(A) General description

(B) Document ranking system (ie- Kroll ranks documents by percentage probability, .01% – 99.9%)

(2) Vendor expert assistance to be provided

d. Budget Range (supported by separate document with detailed estimates and projections)

2. Basic Goals of the Project, including analysis of impact of Proportionality Doctrine and Document Ranking. Here are some possible examples:

a. High recall and production of responsive documents within budget proportionality constraints and time limits.

b. Top 25% probable relevant, and all probable (50%+) highly relevant is a metric goal proportional and reasonable in this particular case for this kind of ESI. (Note – these numbers are often used in high-end, large-scale projects where there is a premium on quality.)

c. All probable relevant and highly relevant within a specified range or set of ranges.

d. Zero Errors in document review screening for attorney client privileged communications.

e. Evaluation of large production received by client.

f. Time sensitive preparations for specific hearings, mediation, depositions, or 3rd party subpoenas.

g. Private internal corporate investigations as part of quality control, business information, compliance and dispute avoidance..

h. Compliance with government requests for information, state criminal investigations and private civil litigation.

3. General Cooperation Strategy

a. Disclosures planned

(1) Transparent

(2) Translucent

(3) Brick Wall

b. Treatment of Irrelevant Documents

c. Relevancy Discussions

d. Sedona Principle Six

4. Team Members for Project

Penrose_triangle_Expertisea. Predictive Coding Chief. Experienced searcher in charge of the Predictive Coding aspects of the document review

1. Experienced ESI Searcher

2. Same person in charge of non-PC aspects, if not, explain

3. Authority and Responsibilities

4. List qualifications and experience

b. Subject Matter Experts (SME)

(1) Senior SME

A. Final Decision Maker – usually partner in charge of case

B. Determines what is relevant or responsive

(i) Based on experience with the type of case at issue

(ii) Predicts how judge will rule on relevance and production issues

C. Formulates specific rules when faced with particular document types

D. Controls communications with requesting parties senior counsel (usually)

E. List qualifications and experience

(2) Junior SME(s)

A. Lead Document Review expert(s)

B. Usually Sr. Associate working directly with partner in charge

C. Seeks input from final decision maker on grey area documents (Undetermined Category)

D. Responsible for Relevancy Rule articulations and communications

E. List qualifications and experience

(3) Amount of estimated time in budget for the work by Sr and Jr SMEs.

A. Assurances of adequate time commitments, availability

B. Reference time estimates in budget

C. Time should exclude training

(4) Response times guaranties to questions, requests from Predictive Coding Chief

c. Vendor Personnel

(1) Anticipated roles

(2) List qualifications and experience

d. Power Users of particular software and predictive coding features to be used

(1) Law Firm and Vendor

(2) List qualifications and experience

e. Outside Consultants or other experts

(1) Anticipated roles

(2) List qualifications and experience

f. Contract Lawyers

(1) Price list for reviewers and reviewer management

A. $50-$75 per hr is typical, but rates vary.

B. Competing bids requested? Why or why not.

(2) Conflict check procedures

(3) Licensed attorneys only or paralegals also

(4) Size of team planned

A. Rationale for more than 5 contract reviewers

B. “Less is More” plan

(5) Contract Reviewer Selection criteria

g. Plan to properly train and supervise contract lawyers

5. One or Two-Pass Review

a. Two pass is standard, with first pass selecting relevance and privilege using Predictive Coding, and second pass by reviewers with eyes-on review to confirm relevance prediction and code for confidentiality, and create priv log.

b. If one pass proposed (aka Quick Peek), has client approved risks of inadvertent disclosures after written notice of these risks?

6. Clawback and Confidentiality agreements and orders

a. Rule 502(d) Order

b. Confidentiality Agreement: Confidential, AEO, Redactions

c. Privilege and Logging

(1) Contract lawyers

(2) Automated prep

7. Categories for Review Coding and Training

a. Irrelevant – this should be a training category

b. Relevant – this should be a training category

(1) Relevance Manual for contract lawyers (see form)

(2) Email family relevance rules

A. Parents automatically relevant is child (attachment) relevant

B. Attachments automatically relevant if email is?

C. All attachments automatically relevant if one attachment is?

c. Highly Relevant – this should be a training category

d. Undetermined – temporary until final adjudication

e. No or Very Few Sub-Issues of Relevant, usually just Highly Relevant

f. Privilege – this should be a training category

g. Confidential

(1) AEO

(2) Redaction Required

(3) Redaction Completed

i. Second Pass Completed

8. Search Methods to find documents for training and production

a. ID persons responsible and qualifications

CULLING.2-Filters.3-lakes-ProductionLb. Methods to cull-out documents before Predictive Coding training begins to avoid selection of inappropriate documents for training and to improve efficiency

(1) Eg – any non-text document; overly long documents

(2) Plan to review by alternate methods

(3) ID general methods for this first stage culling; both legal and technical

c. ID general methods for Predictive Coding, ie – Machine selected only, or multimodal

d. Describe machine selection methods.

(1) Random – should be used sparingly, and never as sole method

(2) Uncertainty – documents that machine is currently unsure of ranking, usually in 40%-60% range

(3) High Probability – documents as yet un-coded that machine considers likely relevant

(4) All or some of the above in combination

Multimodal Search Pyramide. Describe other human based multimodal methods

(1) Expert manual

(2) Parametric Boolean Keyword

(3) Similarity and Near Duplication

(4) Concept Search (passive machine learning, such as latent semantic indexing)

(5) Various Ranking methods based on probability strata selected by expert in charge

f. Describe whether a Continuous Active Learning (CAL) process for review will be used, or two-stage process (train, then review), and if later, rationale

9. Describe Quality Control procedures, including, where applicable, any features built into the software, to accomplish following QC goals. Zero Error Numerics.

quality_trianglea. Three areas of focus to maximize quality of predictive coding

(1) Quality of the AI trainers work to select documents for instruction in the active machine learning process

(2) Quality of the SME work to properly classify documents, especially Highly Relevant and grey area documents, in accord with true probative value and court opinions

(3) Quality of the software algorithms that apply the training input to create a mathematical model that accurately separates the document cloud into probability polar groupings

b. Supervise all reviewers, including contract reviewers who usually do the bulk of the document review work.

(1) ID persons responsible

(2) ID general methods

c. Avoid incorrect conceptions and understanding of relevance and responsiveness, iw – what are you searching for and what will you produce.

(1) Target matches legal obligations

(2) Relevance scope dialogues with requesting party

(3) 26(f) conferences and 16(b) hearings

(4) Motion practice with Court for early resolution of disputes

(5) ID persons responsible

d. Minimize human errors in document coding

(1) Mistakes in relevance rule applications to particular documents

(2) Physical mistakes in clicking wrong code buttons

(3) Inconsistencies in coding of same or similar documents

(4) Inconsistencies in coding of same or similar document types

(5) ID persons responsible

(6) Use AI to double check human work. Zero Error Numerics

e. Facilitate horizontal and vertical communications in team

(1) ID persons responsible

(2) ID general methods

f. Corrections for Concept Drift inherent in any large review project where understanding of relevance changes over time

(1) ID persons responsible

(2) ID general methods

g. Detection of inconsistencies between predictive document ranking and coding (part of AI correction)

(1) ID persons responsible

(2) ID general methods

h. Avoid incomplete, inadequate selection of documents for training

(1) ID persons responsible

(2) ID general methods

i. Avoid premature termination of training

(1) ID persons responsible

(2) ID general methods

j. Avoid omission of any Highly Relevant documents, or new types of strong relevant documents

(1) ID persons responsible

(2) ID general methods

(3) accept on zero error

k. Avoid inadvertent production of privileged documents

(1) List of attorneys names and email domains

(2) Active multimodal search supplement to predictive coding

(3) Dual pass review

(4) ID persons responsible

(5) ID general methods

l. Avoid inadvertent production of confidential documents without proper labeling and redactions

(1) ID persons responsible

(2) ID general methods

m. Avoid incomplete, inaccurate privilege logs

(1) ID persons responsible

(2) ID general methods

n. Avoid errors in final media production to requesting party

(1) ID persons responsible

(2) ID general methods

UpSide_down_champagne_glass10. Decision to Stop Training for Predictive Coding

a. ID persons responsible

b. Criteria to make the decision

(1) Probability distribution

(2) Separation of documents into two poles

(3) Ideal of upside down champagne glass visualization

(4) Few new relevant documents found in last rounds of training

(5) Few new strong relevant types found

(6) No new Highly Relevant documents found

11. Quality Assurance Procedures to Validate Reasonability of Decision to Stop

ei-Recall_smalla. Random Sample Tests to validate the decision

(1) ei-Recall method used, if not, describe

(2) accept on zero error for any Highly Relevant found in elusion test, or new strong relevant type.

(3) Recall and Precision goals

b. Judgmental sampling

c. Zero Error Numerics; consultants and owners agents to approve

12. Procedures to Document the Work Performed and Reasonability of Efforts

a. Clear identification of efforts on the review platform itself with screen shots before project closure

b. Memorandums to file or opposing counsel

(1) Basic metrics for possible disclosure

(2) Detail for internal use only and possible testimony

c. Availability of expert testimony if court challenges arise

________________

_______

5 Responses to Form Outline of a Detailed Plan for Predictive Coding

  1. […] for AI-Enhanced search and review is in the form of a detailed 1,500 word outline. It looks like this Plan is commonly used in the future to obtain client and insurer approval of […]

    Like

  2. […] to plan for it. This is critical to the success of the project. Here is detailed outline of a Form Plan for a Predictive Coding Project that I use as a complete […]

    Like

  3. […] Form Plan of a Predictive Coding Project. Detailed Outline for project planning purposes. […]

    Like

  4. […] it.) For background on how to plan for a complex predictive coding document review project, see Form Plan of a Predictive Coding Project. The plan consists of detailed Outline for the project. To understand the 3.0 method, you also […]

    Like

  5. […] it.) For background on how to plan for a complex predictive coding document review project, see Form Plan of a Predictive Coding Project. The plan consists of detailed Outline for the project. To understand the 3.0 method, you also […]

    Like

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: