Comprehensive ESI Protocol Includes Hyperlinked Files, Generative AI and More: eDiscovery Case Law

A Federal Order Just Turned AI Prompt Logs Into a Discovery Obligation

On July 7, 2026, Magistrate Judge Robert M. Illman in the Northern District of California approved a stipulated ESI protocol in James v. Cerebras Systems Inc. (No. 4:25-cv-09361-AMO) that goes well beyond the district's standard model order. Built on the N.D. Cal. template but negotiated further, the protocol sets a 100-document cap on hyperlinked files a requesting party can flag for production, and locks in validation math for any AI or TAR review: a 95 percent confidence level with a 2 percent margin of error for documents the AI excludes as non-responsive, and a 3 percent ceiling on the elusion rate. It also requires disclosure of every prompt, template, and parameter set used to run the review, with any later change to a prompt redlined and turned over.

What does the order actually require of AI review tools?

Vendors must disclose model identity, version, hosting environment, training methodology, every prompt used, and statistically validated exclusion rates before responsiveness calls hold up.

Appendix 4 of the order is effectively a technical audit checklist: identity and hosting of the model (on-premise, private cloud, or vendor-secured), the culling steps applied before AI ever sees a document (search terms, date ranges, deduplication, threading), which document types were excluded from AI analysis and how those were reviewed instead, and the full methodology — supervised learning, exemplar-based instruction, embeddings, prompt engineering, or a hybrid. None of that is optional under this protocol; it's baked into the stipulated order itself.

Why does the hyperlinked-files carve-out matter?

Hyperlinked documents are no longer treated as automatic attachments — they're a separate, capped, negotiated production category with their own dispute process.

Under the order, hyperlinked files aren't part of a document family by default. A requesting party can identify up to 100 non-public, responsive, non-privileged hyperlinks by Bates number, and the responding party has to produce the collected-time version or explain the withholding — with the parties acknowledging that the version at collection may not match what existed when the link was actually sent. Anything beyond the 100-document cap goes to a meet-and-confer.

What should buyers ask their own e-discovery vendor now?

Ask whether the platform can produce prompt-change redlines, model-hosting disclosures, and Null Set statistics on demand, not after a dispute.

The 95/2/3 figures in this order aren't an industry standard — they were negotiated by these two parties for this case. A vendor quoting "95% accuracy" in a pitch deck should be asked: accuracy against what defined population, measured by whose sampling method, and can the tool actually generate a Null Set validation report and a prompt-revision audit trail without custom engineering? A platform that logs every AI interaction against its source documents by default, rather than reconstructing it after a magistrate judge orders it, is the one that survives protocols like this one becoming boilerplate.

Frequently asked questions

Is this order binding outside the Cerebras case?

No. It's a stipulated order between these parties, but it closely follows the N.D. Cal. model ESI order, which other litigants and judges in the district commonly reference.

Does the protocol ban AI review tools that can't meet these numbers?

The order doesn't ban tools — it conditions AI-assisted responsiveness review on disclosure and validation, meaning a tool that can't produce the required statistics or prompt logs can't be used as agreed here.

What happens to hyperlinked documents flagged by mistake as privileged?

The protocol includes standard inadvertent-disclosure protection, so producing a privileged hyperlinked document under this process does not waive privilege.

Source: eDiscovery Today, "Comprehensive ESI Protocol Includes Hyperlinked Files, Generative AI and More: eDiscovery Case Law," by Doug Austin, July 9, 2026.