Panel estimates never enter a score

In-session one-vendor panel under D154(b) as amended by D155, run on 2026-10-17 by a Claude Code orchestrating session (Claude Opus 5.5) on dataset 9.4.0, after D167. Three panelists, each stance a separate set of contexts: institutionalist on claude-opus-5-5, bottom_up on claude-opus-5-5, wealth_sceptic on claude-sonnet-5-5, all Anthropic, each through Claude Code subagents. Each panelist answered in batches of ten countries, one fresh context per batch and per round, and the indicator audit in one context per stance, so no context held two stances and none saw another panelist's text except the anonymised round-1 summary the round-2 prompt carries. Two rounds plus the indicator audit, at maxCoverage 1: all nine dimensions for all 125 countries, 1125 country-dimension cells, not the 0.5 ceiling of the 9.0.1 run. Prompts are byte-identical to a gateway run's, built by pnpm bench delphi --in-session with the current prompt version; the system rules were given as a file each panelist read first. Every batch reported it read every prompt in full and wrote every answer by hand. A first attempt the same day stopped at the account's usage limit before any batch finished; its partial answers were discarded and every prompt was regenerated and answered from scratch. No call in this run failed and none was rerun into it. One vendor means the spread measures stance, not vendor, and shared model-family blind spots are hidden by a narrow IQR, which is a floor on real uncertainty (D155). A multi-vendor or gateway run on the same dataset supersedes this one. This is an interpretation layer: its estimates reach only delphiScore and delphiIqr, and blendedScore only where a dimension has no observed indicator; they never enter score.

No cell has an interquartile range above 25 points.