diff --git a/AGENTS.md b/AGENTS.md
index 7a8681e..ab56ca5 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -2575,12 +2575,27 @@ shared `filterAdminHidden` drops muted authors across every list; the flagged
 notice has a Mute action. **Alerts landed:** `useNetworkAlerts` + `/alerts`
 (`pages/NetworkAlertsPage.tsx`) list accounts you follow that trusted people
 reported/muted, with Ignore/Unfollow/Mute/Report; Settings → Trust links to it
-with a count (ignored set in `skatespots:trust:alerts-ignored`). Still open:
-running our own GrapeRank-style scorer in `cvm-server` — the full confidence
-model is provider-side today. Check
+with a count (ignored set in `skatespots:trust:alerts-ignored`). Deferred: our
+own GrapeRank-style scorer — see the dated note below. Check
 `TRUST_SCORE.md` before touching `useTrustScore` / `useRelatrScore` /
 `trustFilter`.
 
+**Own GrapeRank-style scorer — DEFERRED (2026-10-09).** Do NOT build the
+`cvm-server` per-observer scorer yet. The full confidence model is provider-side
+and already consumed via NIP-85, and the client fallback now has a local
+confidence gate (`useViewerGraph`, `FOF_MIN_RATERS`). Why hold: personalized
+kind-30382 needs a **per-observer signing key** (server key custody or per-user
+delegation — a trust/key-management burden, not a checkbox); an 8-hop traversal is
+heavy without a graph store; plus a provider identity, the publish/delete
+lifecycle, deploy to the external host and monitoring. With ~zero users the ROI is
+negative. Cheaper path if a controlled score is ever wanted: a
+**house-perspective** scorer (one key, one observer, computed on a schedule,
+publishes kind-30382) — Settings → Trust's provider choice consumes it with no
+client change. **Revisit when:** there are users, no trustworthy provider exists,
+or there are skate-specific signals (vouches, trade attestations) worth baking
+into a score only we can compute. The client is already provider-ready
+(`useTrustAssertions` + kind-10040 choice), so this decision costs nothing today.
+
 **Keep the door open (cheap):** admin keys and relay sets must stay *data you can
 swap*, not constants. Hardcoded `ADMIN_PUBKEYS` is fine now — converting it later
 is a small refactor, not a trap.
diff --git a/TRUST_SCORE.md b/TRUST_SCORE.md
index 1dd3937..701e164 100644
--- a/TRUST_SCORE.md
+++ b/TRUST_SCORE.md
@@ -91,9 +91,11 @@ flagged notice's reporter breakdown, and Settings → Trust.
   `deriveLocalScore` returns `verified` only at `FOF_MIN_RATERS` (2) or more. A
   lone endorsement no longer reads as trust.
 - Provider assertions are unaffected: the full GrapeRank confidence model lives
-  in the provider and we consume it via NIP-85. Running our own scorer (planned
-  for `cvm-server`, publishing kind-30382) remains the open item — it is the only
-  way to get the real confidence gate and a server-published score.
+  in the provider and we consume it via NIP-85. Running our own scorer is
+  **deferred** — see the dated decision in `AGENTS.md` ("Own GrapeRank-style
+  scorer — DEFERRED"): the per-observer signing key and 8-hop compute are the
+  blockers, and the client is already provider-ready. A house-perspective scorer
+  (one key, one observer) is the cheaper path if ever wanted.