In James v. Cerebras Sys. Inc., No. 4:25-cv-09361-AMO (N.D. Cal. July 7, 2026), California Magistrate Judge Robert M. Illman entered a comprehensive ESI protocol governing the preservation, review, and production of diverse data sources, including enterprise solutions, the handling of hyperlinked files, strict disclosure and validation metrics for AI and TAR workflows, including prompt-iteration logs, a 95 percent confidence level for Null Set sampling, and a maximum 3 percent elusion rate, while protecting inadvertent disclosures of privileged ESI from waiver.
Stipulated Order Regarding the Production of ESI and Hard Copy Documents
While largely based on the Northern District of California’s model ESI order, the parties negotiated several notable enhancements in this comprehensive ESI protocol that reflect the realities of modern discovery, particularly regarding AI, hyperlinked files, collaboration platforms such as Slack and Microsoft Teams, and cloud-based document management. Here are a few highlights:
Hyperlinked Files
Rather than treating hyperlinked documents as traditional email attachments, the parties agreed that hyperlinked documents are not automatically considered part of a document family for production purposes. Instead:
- A requesting party may identify up to 100 non-public, responsive, non-privileged hyperlinked documents by providing the associated hyperlinks and Bates numbers.
- The responding party must produce the version of the hyperlinked document that existed at the time of collection within 14 days or explain why it is being withheld.
- The parties acknowledged that the document existing when collected may differ from the version that existed when the hyperlink was originally sent, so, if possible, the responding party will produce the version of the document as it existed at the time the hyperlink was sent.
- The parties also agreed to meet and confer regarding disputes and reasonable requests to exceed the 100-document cap.
Artificial Intelligence
Notable provisions were contained in Appendix 4 of the order and included:
- Disclosure of the identity, version, and hosting environment of each AI model or system to be used (e.g., LLM, classifier, embedding model; on-premise, private cloud, or vendor-secured environment).
- Criteria and methodology used to identify the document population subject to AI Responsiveness Review, including any culling performed prior to AI processing (e.g., search terms, date ranges, custodians, deduplication, threading, near-duplicate suppression).
- Identification of any Document types excluded from AI analysis (e.g., images, spreadsheets, unsupported foreign-language Documents, multimedia files, poorly OCR’d Documents, or Documents deemed incompatible with LLM processing), and the method for manually or otherwise reviewing those materials.
- The methodology used to train or instruct the AI system, including any supervised learning, exemplar-based instruction, embeddings-based classification, transfer learning, prompt-engineering, or hybrid approach.
- Disclosure of all Prompts, templates, instruction sets, or parameter configurations used to guide any AI component. Any modification to a Prompt after service of the AI Review Protocol shall be disclosed to the Requesting Party in redline form within three (3) business days.
- Validation information, including the validation method used (random sampling, system-generated sampling, or external statistical method), The number of validation Documents reviewed, calculated using a 95% confidence level with a + or – 2% margin of error for Documents excluded by the AI as non-responsive, and an expectation that the elusion rate won’t exceed 3%.
Short Message Data
Electronic messages exchanged between users on communication software such as Microsoft Teams and Slack shall be produced in a searchable format that preserves the conversational relationship and presentational features of the original messages, such as emojis, images, video files, and animations. Electronic messages must not be converted to unitized files that contain less than a 24-hour period of conversation. Redactions may be applied to privileged portions of a conversation. To the extent electronic messages cannot be produced in a reasonably useable format, the Parties will meet and confer to address the identification, production, and production format of short-message data.
Databases, Structured, Aggregated or Application Data
Requirement for the Parties to meet and confer regarding the production and format and scope of relevant structured data or aggregated or threaded data source or otherwise maintained by an application (e.g., Microsoft Access, SharePoint, Oracle, Salesforce, ACT!, or any other or proprietary databases or services) in order to ensure that any information produced is reasonably usable by the Requesting Party and that its production does not impose an undue burden on the Responding Party.
Inadvertent Production of Privileged Material
Pursuant to Federal Rule of Evidence 502(d) and (e), the production of a privileged or work-product-protected document is not a waiver of privilege or protection from discovery in this case or in any other federal or state proceeding.
There are several other components to this comprehensive ESI protocol, but these are the most notable components. It’s very interesting and surprising that the parties agreed to disclosure of all prompts used for AI – not sure why they don’t consider those to be attorney work product.
So, what do you think of this comprehensive ESI protocol? Please share any comments you might have or if you’d like to know more about a particular topic.
Case opinion link courtesy of Minerva26, an Affinity partner of eDiscovery Today.
Disclaimer: The views represented herein are exclusively the views of the author, and do not necessarily represent the views held by my employer, my partners or my clients. eDiscovery Today is made available solely for educational purposes to provide general information about general eDiscovery principles and not to provide specific legal advice applicable to any particular circumstance. eDiscovery Today should not be used as a substitute for competent legal advice from a lawyer you have retained and who has agreed to represent you.