diff --git a/PLAN.md b/PLAN.md
index 8e2b1ff..9008adc 100644
--- a/PLAN.md
+++ b/PLAN.md
@@ -126,7 +126,7 @@ Wichtig:
   - track
   - present
   - record
-- Gemeinsame Arbitration-/Budget-Sicht für refinement/record/decode vorbereiten
+- Gemeinsame Arbitration-/Budget-Sicht für refinement/record/decode zentralisieren (Admission + Queue/Hold + Debug-Surface)
 
 ### F. Dokumentierte Betriebsprofile
 - initiale Profile definieren, z. B.:
diff --git a/README.md b/README.md
index 57b2567..a9bbeb1 100644
--- a/README.md
+++ b/README.md
@@ -84,7 +84,7 @@ Edit `config.yaml` (autosave goes to `config.autosave.yaml`).
 - `resources.max_refinement_jobs` — processing budget hint
 - `resources.max_recording_streams` — recording/streaming budget hint
 - `resources.max_decode_jobs` — decode budget hint
-- `resources.decision_hold_ms` — hold time for queue slots before churn
+- `resources.decision_hold_ms` — baseline hold time for queue slots before churn (arbitration scales per profile/strategy)
 - `profiles[]` — named operating profiles/intent metadata
 
 In phase 1, the engine stays backward compatible, but the config model now reflects the intended separation between:
@@ -95,6 +95,8 @@ In phase 1, the engine stays backward compatible, but the config model now refle
 - presentation
 - operator goals / future autonomous intent
 
+Refinement plans now rank candidates, while a shared arbitration step admits refinement/record/decode work based on budgets and hold policy.
+
 The long-term target is that you describe *what the system should do* (for example broad-span monitoring intent, preferred signal families, auto-record/decode priorities), while the engine decides *how* to allocate surveillance, refinement and decoding budgets.
 
 **CFAR modes:** `OFF`, `CA`, `OS`, `GOSCA`, `CASO`
@@ -159,7 +161,7 @@ go build -tags sdrplay ./cmd/sdrd
 - `GET /api/gpu`
 - `GET /api/pipeline/policy`
 - `GET /api/pipeline/recommendations`
-- `GET /api/refinement` → latest refinement plan/windows snapshot (includes `window_stats`, `queue_stats`, `decision_summary`, `decision_items`, `arbitration`, levels, request/context/budgets/work_items)
+- `GET /api/refinement` → latest refinement plan/windows snapshot (includes `window_stats`, levels, request/context/work_items, plus `arbitration` with budgets/hold policy/refinement admission/queue/decision summary)
 
 ### Signals / Events
 - `GET /api/signals` → current live signals
diff --git a/web/app.js b/web/app.js
index 2a2f6c9..2f5c649 100644
--- a/web/app.js
+++ b/web/app.js
@@ -757,7 +757,7 @@ async function loadRefinement() {
     if (!res.ok) return;
     refinementInfo = await res.json();
     decisionIndex = new Map();
-    const items = Array.isArray(refinementInfo.decision_items) ? refinementInfo.decision_items : [];
+    const items = Array.isArray(refinementInfo.arbitration?.decision_items) ? refinementInfo.arbitration.decision_items : [];
     items.forEach(item => {
       if (item && item.id != null) decisionIndex.set(String(item.id), item);
     });
@@ -864,13 +864,13 @@ function updateHeroMetrics() {
   healthFps.textContent = `${renderFps.toFixed(0)} fps`;
   if (healthRefinePlan) {
     const plan = refinementInfo.plan || {};
-    const decisionSummary = refinementInfo.decision_summary || {};
+    const decisionSummary = refinementInfo.arbitration?.decision_summary || {};
     const recOn = decisionSummary.record_enabled ?? 0;
     const decOn = decisionSummary.decode_enabled ?? 0;
     const reasonCounts = decisionSummary.reasons || {};
     const topReason = Object.entries(reasonCounts).sort((a, b) => b[1] - a[1])[0];
     const reasonText = topReason ? `· ${topReason[0]}` : '';
-    const queueStats = refinementInfo.queue_stats || {};
+    const queueStats = refinementInfo.arbitration?.queue || {};
     const queueText = (queueStats.record_queued || queueStats.decode_queued)
       ? `· q ${queueStats.record_queued || 0}/${queueStats.decode_queued || 0}`
       : '';