Round 7 — Claude review of v8

Six adopted requests: separate the archived finite panel from latent-propensity inference; define “cheap” (evidentiary economy); clarify that exact-gate eligibility depends on the full episode-value composition, not the sample mean; attach Monte Carlo intervals and the add-one convention; reconcile the Phase 4–5 request/response/partial/ledger counts; use “five-phase” consistently.