Part 3b: clarify the MHS reliability corpus

reports/part-3b-reliability-corpus-addendum-2026-07-19.md

@@ -0,0 +1,32 @@
+# Part 3b reliability-corpus addendum
+
+**Date:** 2026-07-19<br>
+**Status:** binding prospective clarification, approved before any MHS outcome run<br>
+**Applies to:** [`part-3b-dataset-selection-audit.md`](part-3b-dataset-selection-audit.md), Primary metric 4<br>
+**Repository base:** `af34ebdd6b11a1f42eb86584c96613fc3835df76`
+
+## Reliability-corpus disambiguation
+
+Primary metric 4—“the current CONFIDENT-only, leave-one-out reliability
+definition over the full converted corpus for eligible annotators”—is read as
+follows:
+
+> The converted reliability corpus contains only eligible cohort annotators.
+> Consistent with Eligibility rule 4, annotators with neutral, no-opinion, or
+> null ideology are excluded entirely: their judgments enter neither the
+> reliability corpus, the CONFIDENT-cell determination, nor any leave-one-out
+> computation.
+
+This is the interpretation implemented in `adapters/mhs.py` and independently
+audited in PR #15. It is recorded prospectively, before any real-MHS outcome is
+computed.
+
+## Effect on the protocol
+
+This clarification changes no eligibility floor, cohort definition, primary-item
+filter, metric, threshold, bootstrap rule, evidence tier, permitted claim, or
+non-claim. It resolves only which already-excluded annotators can contribute to
+the reliability corpus.
+
+Phase 2 remains separately gated. This addendum authorizes neither a real-data
+run nor publication of an outcome.