Add GTM Hub real-world regression case

docs/experiments.md

@@ -2259,3 +2259,32 @@ the basis of this prototype. Further work should first analyze whether the
 failure comes from the schema/prompt representation, the model's sparse-output
 reliability, or the boundary between subject grouping and semantic synthesis.
 It should not proceed through repeated prompt tuning against these nine cases.
+
+## EXP-0027 — GTM-Hub real-world protocol regression case
+
+Status: Accepted as a qualitative regression case
+
+Date: 2026-09-09
+
+The private `samples/real_live/gtm_hub_2026-09-07/` case preserves an existing
+78-minute German GTM-Hub discussion, its exact protocol input transcript, a
+human colleague's concise reference, and two existing Meeting Assistant
+protocols (with anonymous speaker labels and with five confirmed mappings).
+It is an evaluation dataset, not a prompt/model experiment; no inference was
+run to create it.
+
+The evaluation establishes product and architecture learning without changing
+the current architecture:
+
+- the detailed/working protocol remains the primary product artifact; a short
+  distribution protocol is a later separate renderer, not a replacement;
+- diarization has particular value for long, multi-speaker,
+  disagreement-heavy meetings because it improves discourse attribution;
+- diarization does not solve semantic commitment extraction: speaker
+  attribution and normative interpretation are separate problems;
+- responsibility and action extraction require stricter evidence than ordinary
+  discussion summarization, especially when a named speaker is involved.
+
+Evidence and explicit qualitative regression expectations are recorded in
+`samples/real_live/gtm_hub_2026-09-07/evaluation.md`. The case does not define
+a synthetic gold protocol or a benchmark metric.