diff --git a/adventure_expansion_spec.md b/adventure_expansion_spec.md
index e51079b..6191211 100644
--- a/adventure_expansion_spec.md
+++ b/adventure_expansion_spec.md
@@ -135,6 +135,23 @@ going forward, same as every other trophy.
 
 ## 2. LLM-authored dispatches (`headline` / `lede`)
 
+**SHIPPED 2026-07-17: Pete `eeeac08`, gogobee `22b7949`. All tests green both
+sides, gofmt-clean, screenshot-verified (realistic + max-length prose both sit
+cleanly on the card), NOT deployed.** Built as written below (additive
+`headline`/`lede`, prose-guard, template fallback), with one architecture
+decision the spec did not surface:
+
+> **This section contradicts `pete_adventure_news_voice.md`**, the older
+> foundational doc, which says *Pete* owns the voice and gogobee is "compute,
+> not ghostwriter" — the flow there is Pete builds a voiced prompt and calls a
+> generic gogobee inference endpoint. That needs a **Pete→gogobee route**, which
+> `roster.go:23-25` forbids ("no route back into the game box's network"). The
+> network constraint kills the voice-doc design, so §2's gogobee-authors-and-
+> pushes model is the only one that fits one-way delivery. Confirmed with the
+> owner before building. Cost: the warm-reporter voice now lives in gogobee's
+> prompt (`pete_dispatch_voice.go`), softening "gogobee never sees Pete" — a
+> deliberate, owner-approved trade, not an oversight.
+
 The template-rendered dispatches are all identical and read as boilerplate.
 gogobee's LLM writes the prose instead; Pete's templates stop being the renderer
 and **become the safety net**.