Doe v. Meta: When the Selection Records Are the Dispute

On the evening of July 13, 2026, twenty-six current and former Meta employees filed suit as Doe 1–26 in the Northern District of California (Case No. 4:26-cv-07122, assigned to Judge William H. Orrick as 3:26-cv-07122-WHO). Contrary to early coverage, it is not a class action: it is a multi-plaintiff individual action under FRCP 20, with class-waiver arbitration agreements in force; the merits go to individual arbitration, and the court is asked only to preserve the status quo in the meantime.

What is verified — and what is alleged

The layoff itself is documented: announced in April 2026, notices from May 20, roughly 8,000 positions (about 10% of a workforce of 78,865 at end-2025), alongside a separate reassignment of about 7,000 people into the AI division. The plaintiffs' account of how selections were made is, throughout, a set of untested allegations: a "constellation of internal artificial-intelligence systems" — the Metamate assistant, "second-brain" agents, keystroke and activity monitoring, AI-token-usage dashboards, algorithm-assisted performance ranking and calibration — formalized in April 2026 as a "Model Capability Initiative"; and a claim that all twenty-six plaintiffs had taken protected medical or family leave or received disability accommodations in the prior twenty-four months, with leave allegedly weighing as a negative factor. Meta denies this comprehensively: workforce decisions, it says, were made by people, not AI.

The records at issue

Strip away the labels and the factual dispute is a provenance question: for each contested selection, what did each system output, and what role did that output play in the human decision? Answering it requires records that bind model outputs, their versions and inputs, and the human sign-offs that followed — with cross-references holding the chain together across systems, which is what VAP's XREF mechanism exists for. But we owe readers the same honesty we demand of others: VAP v1.2's acknowledged gaps include the absence of a standardized vocabulary for decision-authority allocation — a machine-verifiable way to record "the model scored; the human decided" versus "the model decided; the human ratified." Doe v. Meta is a live demonstration of why that gap matters: both sides' positions are, in effect, competing decision-authority claims, and today there is no standard record format in which such claims could have been contemporaneously fixed.

What VAP would not have done

VAP takes no position on whether any selection was lawful or discriminatory, and it would not have prevented any termination. Records exist only where systems were instrumented to emit them at decision time; selections that left no record are Tier 1 and cannot be reconstructed by any provenance mechanism.

Per the normative non-guarantee clause, conformance “…does not warrant the correctness, fairness, or safety of the underlying AI decisions — only the integrity, completeness (at anchor granularity), and attributability of their records. VAP generates evidence; competent authorities and courts evaluate it.”
VAP v1.2 §1.6 (English text is normative)

Verified against: the 71-page complaint (filed 2026-07-13, N.D. Cal.); Reuters (Daniel Wiessner), Courthouse News, CNBC, and KRON4 reporting of 2026-07-13/14; Meta's public denial via spokesperson. All plaintiff claims remain untested.