The Keyed Ensemble (EA-LO-KEYED-ENSEMBLE-01) is a Lagrange Observatory! white paper by Nobel Glas on the text watermark Anthropic announced in August 2026 — a version of Google DeepMind's SynthID-Text, adopted for EU AI Act transparency and applied worldwide. Its central formal object is the certification mismatch: non-distortion is certified per sequence while training corpora are ensembles, and the vendor's own source paper reports the ensemble movement — inter-response diversity falls under Self-BLEU — while every per-response quality instrument reads flat. The paper distinguishes throughout between properties of published SynthID-Text and properties of Claude's undisclosed instantiation; separates the measured Self-BLEU fact from its proposed joint-entropy interpretation; and types the asymmetry of the key as issuer control of the verification primitive — the public can receive a verdict but cannot independently reproduce the keyed test — with differential collapse exposure held as a hypothesis contingent on practice. Transmission legs rest on watermark distillation (Gu et al., ICLR 2024) and radioactivity (Sander et al., NeurIPS 2024), with erosion limits recorded; a four-observable measurement program is specified without requiring key disclosure. Developed through three adversarial review rounds; the founding Operator's Review is appended as the historical adjudication trail with superseded statements flagged. Companion: EA-LO-INTERLOCKING-AUTOREGRESSION-01 (#1556), which holds the shared evidentiary basis. Cites #199 with the term attribution corrected per ERRATUM to AXN:0341 (#1554).