In re Lyft, Inc. Passenger Sexual Assault Litigation, No. 26-md-03171-RFL (LJC) (N.D. Cal. Sept. 20, 2026), full opinion (PDF)
Lyft and the plaintiffs in the passenger sexual assault MDL had agreed on many of the TAR provisions in their ESI protocol. The protocol promised validation of end-to-end recall. What the parties could not settle was which documents the validation sample had to cover. Lyft argued that only the TAR process needed validation, while the plaintiffs read the protocol to reach the documents that search terms had culled before TAR review ever began. A Northern District of California magistrate judge has now read the protocol the plaintiffs' way, ordering Lyft to add a sample of keyword-excluded documents to its validation sample. The court also ordered the sample produced early, well ahead of the substantial completion deadline. It denied the plaintiffs' proposed wording insofar as that wording would have pushed validation to the end of the production process.
What happened
The ESI protocol was entered in July 2026, before Lyft had decided whether to use TAR. It stated that the search terms would be validated after review, by random sampling of the entire collection against which the terms were run, to establish end-to-end recall. Methodology was left for the parties to negotiate.
The plaintiffs had not objected to Lyft's use of search terms to cull the collection before TAR review. In their joint dispute letter, the parties disagreed over which documents validation had to cover. In the plaintiffs' reading, the protocol required end-to-end validation, reaching the documents the search terms excluded as well as the documents TAR reviewed. Lyft called that kind of validation nonstandard. It described the protocol's language on end-to-end recall as a placeholder. In the alternative, Lyft asked that the plaintiffs pay the associated costs.
The court's analysis
The court started from the protocol's text. Room had been left for the parties to confer on methodology, the court acknowledged. In the court's view, the populations to be validated and the purpose of validation were nonetheless clear. Validation, the court reasoned, was meant to test Lyft's whole search, classification, and review methodology rather than TAR alone.
The parties had agreed that a satisfactory search term hit rate would fall between 20 and 36 percent. The court drew a second ground, proportionality, from that range. At least 64 percent of the custodial records, the court noted, would therefore never be subject to TAR. Validating the search term cull was proportionate, the court concluded, because a substantial portion of the collected records would never reach TAR review. The plaintiffs sought the sample, the court observed, to ensure that the rate of false negatives was reasonable.
Lyft's burden objection failed, in the court's view, because the added sample was modest. It would run to about 1,537 documents by the plaintiffs' count. Lyft did not dispute that figure. The plaintiffs were also willing to consider a smaller sampling design, the court noted. The court expected those documents to be quick to review because most would be non-responsive. It denied Lyft's request to shift the costs to the plaintiffs.
The plaintiffs had proposed drawing the sample from the entire population of documents subject to keyword filtering and TAR. That formulation, as the court read it, could let the sample wait until document discovery closed. A search term problem is cheaper to correct early than late, the plaintiffs themselves had argued, an argument the court accepted. Lyft must therefore produce the sample reasonably soon after it completes collection and production under the final search terms, well before the substantial completion deadline.
Why it matters
A producing party whose protocol speaks of end-to-end recall should expect the search term cull to be validated. Once a dispute reaches the court, language a party later calls a placeholder can be read as a commitment. A producing party that wants TAR-only validation needs to say so in the protocol itself.
When search terms exclude most of a collection, the hit rate the parties have already agreed to is the requesting party's strongest proportionality argument. A sample drawn only from the TAR set measures the minority of the documents. The requesting party can make that case from the figures the parties have already put on paper.
A producing party resisting an expanded sample needs numbers of its own. Put a review cost estimate in the record, or offer a smaller sampling design. Cost shifting is unlikely where the protocol already contemplates the validation.
The validation sample should also have a fixed trigger. A sample defined by reference to the entire reviewed population invites the argument that nothing can be drawn until production ends. By then a defective term list is at its most expensive to fix. Tie the sample to a milestone, such as the completion of collection and production under the final terms. Set a date for conferring on the validation schedule as well.
The full opinion is available as a PDF.
