diff --git a/internal/router/bandit/bandit.go b/internal/router/bandit/bandit.go
index a2ee6188f..75e88cec5 100644
--- a/internal/router/bandit/bandit.go
+++ b/internal/router/bandit/bandit.go
@@ -109,8 +109,13 @@ func (b *Router) Route(ctx context.Context, req router.Request) (router.Decision
 	argmaxModel := dec.Model
 	pick := candidates[0]
 	b.annotate(&dec, pick.model, pick.provider, b.propensity(clusterIDs, candidates, pick.model), pick.sample)
-	if pick.model != argmaxModel && md.EffectiveKnobsHash != 0 {
-		md.EffectiveKnobsHash = mixModel(md.EffectiveKnobsHash, pick.model)
+	if pick.model != argmaxModel {
+		if md.EffectiveKnobsHash != 0 {
+			md.EffectiveKnobsHash = mixModel(md.EffectiveKnobsHash, pick.model)
+		}
+		// The scorer's runner-up was computed against the argmax and may now
+		// equal the served peer; recompute it against the served model.
+		repairBandPair(&md.PairedModel, &md.PairedProvider, &md.PairedScore, md.CandidateScores, candidates, pick.model)
 	}
 	return dec, nil
 }
@@ -185,6 +190,40 @@ func (b *Router) annotate(dec *router.Decision, model, provider string, propensi
 	}
 }
 
+// repairBandPair recomputes the band pair's runner-up against the served
+// model. Picks the highest-scoring servable peer other than served (ties
+// broken by name); clears the pair if none remains. Providers come from the
+// live candidate list so peers resolved only via Decision.Provider stay eligible.
+func repairBandPair(outModel, outProvider *string, outScore *float32, scores map[string]float32, cands []candidate, served string) {
+	providers := make(map[string]string, len(cands))
+	for _, c := range cands {
+		providers[c.model] = c.provider
+	}
+	models := make([]string, 0, len(scores))
+	for m := range scores {
+		models = append(models, m)
+	}
+	sort.Strings(models)
+
+	bestModel, bestProvider := "", ""
+	var bestScore float32
+	for _, m := range models {
+		if m == served {
+			continue
+		}
+		sc := scores[m]
+		if bestModel != "" && sc <= bestScore {
+			continue
+		}
+		provider, ok := providers[m]
+		if !ok || provider == "" {
+			continue
+		}
+		bestModel, bestProvider, bestScore = m, provider, sc
+	}
+	*outModel, *outProvider, *outScore = bestModel, bestProvider, bestScore
+}
+
 func mixModel(h uint64, model string) uint64 {
 	f := fnv.New64a()
 	var b [8]byte
diff --git a/internal/router/bandit/bandit_test.go b/internal/router/bandit/bandit_test.go
index 4eeda5b79..03b82086e 100644
--- a/internal/router/bandit/bandit_test.go
+++ b/internal/router/bandit/bandit_test.go
@@ -131,8 +131,12 @@ func TestRoute_PropensityReflectsPosteriorOverlap(t *testing.T) {
 }
 
 func TestRoute_KeepsArgmaxWhenSameModel(t *testing.T) {
+	// Keeping the argmax must leave the scorer's PairedModel untouched.
 	scores := map[string]float32{"claude-haiku-4-5": 0.9, "claude-sonnet-4-6": 0.5}
 	inner := clusterDecision(scores, []int{0}, "claude-haiku-4-5", "anthropic")
+	inner.Metadata.PairedModel = "claude-sonnet-4-6"
+	inner.Metadata.PairedProvider = "anthropic"
+	inner.Metadata.PairedScore = 0.5
 	post, err := LoadPosterior("testdata/ts_posterior.json")
 	if err != nil {
 		t.Fatal(err)
@@ -149,6 +153,53 @@ func TestRoute_KeepsArgmaxWhenSameModel(t *testing.T) {
 	if dec.Metadata.EffectiveKnobsHash != 99 {
 		t.Fatal("keeping argmax must preserve cache key")
 	}
+	if dec.Metadata.PairedModel != "claude-sonnet-4-6" {
+		t.Fatalf("argmax keep must preserve scorer runner-up sonnet, got %q", dec.Metadata.PairedModel)
+	}
+	if dec.Metadata.PairedScore != 0.5 {
+		t.Fatalf("argmax keep must preserve scorer paired score, got %v", dec.Metadata.PairedScore)
+	}
+}
+
+func TestRoute_RepairsBandPairWhenServingRunnerUp(t *testing.T) {
+	// Serving the former runner-up must recompute PairedModel to the prior argmax.
+	scores := map[string]float32{
+		"claude-sonnet-4-6": 0.90,
+		"claude-haiku-4-5":  0.85,
+	}
+	inner := clusterDecision(scores, []int{0}, "claude-sonnet-4-6", "anthropic")
+	inner.Metadata.PairedModel = "claude-haiku-4-5"
+	inner.Metadata.PairedProvider = "anthropic"
+	inner.Metadata.PairedScore = 0.85
+	post := &Posterior{cells: map[int]map[string]Arm{
+		0: {
+			"claude-haiku-4-5":  {Mean: 0.9, Variance: 0},
+			"claude-sonnet-4-6": {Mean: 0.1, Variance: 0},
+		},
+	}}
+	b := New(&fakeInner{dec: inner}, post)
+	b.norm = func() float64 { return 0 }
+	b.trials = 0
+
+	dec, err := b.Route(context.Background(), router.Request{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if dec.Model != "claude-haiku-4-5" {
+		t.Fatalf("expected Thompson pick haiku (runner-up), got %q", dec.Model)
+	}
+	if dec.Metadata.PairedModel == dec.Model {
+		t.Fatalf("band pair collapsed: Model and PairedModel are both %q", dec.Model)
+	}
+	if dec.Metadata.PairedModel != "claude-sonnet-4-6" {
+		t.Fatalf("expected runner-up recomputed to sonnet, got %q", dec.Metadata.PairedModel)
+	}
+	if dec.Metadata.PairedProvider != "anthropic" {
+		t.Fatalf("expected paired provider anthropic, got %q", dec.Metadata.PairedProvider)
+	}
+	if dec.Metadata.PairedScore != 0.90 {
+		t.Fatalf("expected paired score 0.90, got %v", dec.Metadata.PairedScore)
+	}
 }
 
 func TestRoute_UsesPerRequestProviderBinding(t *testing.T) {