diff --git a/internal/pipeline/candidate_fusion.go b/internal/pipeline/candidate_fusion.go
index d754896..aa8aa55 100644
--- a/internal/pipeline/candidate_fusion.go
+++ b/internal/pipeline/candidate_fusion.go
@@ -16,10 +16,12 @@ func AddCandidateEvidence(candidate *Candidate, evidence LevelEvidence) {
 			evLevel = "unknown"
 		}
 		if evLevel == levelName && ev.Provenance == evidence.Provenance {
+			RefreshCandidateEvidenceState(candidate)
 			return
 		}
 	}
 	candidate.Evidence = append(candidate.Evidence, evidence)
+	RefreshCandidateEvidenceState(candidate)
 }
 
 func MergeCandidateEvidence(dst *Candidate, src Candidate) {
@@ -32,18 +34,8 @@ func MergeCandidateEvidence(dst *Candidate, src Candidate) {
 }
 
 func CandidateEvidenceLevelCount(candidate Candidate) int {
-	if len(candidate.Evidence) == 0 {
-		return 0
-	}
-	levels := map[string]struct{}{}
-	for _, ev := range candidate.Evidence {
-		name := ev.Level.Name
-		if name == "" {
-			name = "unknown"
-		}
-		levels[name] = struct{}{}
-	}
-	return len(levels)
+	state := CandidateEvidenceStateFor(candidate)
+	return state.DetectionLevelCount
 }
 
 func FuseCandidates(primary []Candidate, derived []Candidate) []Candidate {
@@ -74,6 +66,9 @@ func FuseCandidates(primary []Candidate, derived []Candidate) []Candidate {
 		}
 		out = append(out, cand)
 	}
+	for i := range out {
+		RefreshCandidateEvidenceState(&out[i])
+	}
 	return out
 }
 
@@ -113,6 +108,9 @@ func candidateSpanHz(candidate Candidate) float64 {
 
 func candidateBinHz(candidate Candidate) float64 {
 	for _, ev := range candidate.Evidence {
+		if IsPresentationLevel(ev.Level) || !IsDetectionLevel(ev.Level) {
+			continue
+		}
 		if ev.Level.BinHz > 0 {
 			return ev.Level.BinHz
 		}
diff --git a/internal/pipeline/candidate_fusion_test.go b/internal/pipeline/candidate_fusion_test.go
index ae33cfd..2ffd179 100644
--- a/internal/pipeline/candidate_fusion_test.go
+++ b/internal/pipeline/candidate_fusion_test.go
@@ -30,6 +30,9 @@ func TestFuseCandidatesDedup(t *testing.T) {
 	if got := CandidateEvidenceLevelCount(fused[0]); got != 2 {
 		t.Fatalf("expected 2 evidence levels after fuse, got %d", got)
 	}
+	if fused[0].EvidenceState == nil || !fused[0].EvidenceState.Fused || !fused[0].EvidenceState.MultiLevelConfirmed {
+		t.Fatalf("expected fused multi-level evidence state, got %+v", fused[0].EvidenceState)
+	}
 }
 
 func TestFuseCandidatesSingleVsMultiResolution(t *testing.T) {
diff --git a/internal/pipeline/evidence.go b/internal/pipeline/evidence.go
new file mode 100644
index 0000000..f502002
--- /dev/null
+++ b/internal/pipeline/evidence.go
@@ -0,0 +1,173 @@
+package pipeline
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// CandidateEvidenceState summarizes fused evidence semantics for a candidate.
+type CandidateEvidenceState struct {
+	TotalLevelEntries       int      `json:"total_level_entries"`
+	LevelCount              int      `json:"level_count"`
+	DetectionLevelCount     int      `json:"detection_level_count"`
+	PrimaryLevelCount       int      `json:"primary_level_count,omitempty"`
+	DerivedLevelCount       int      `json:"derived_level_count,omitempty"`
+	PresentationLevelCount  int      `json:"presentation_level_count,omitempty"`
+	Levels                  []string `json:"levels,omitempty"`
+	Provenance              []string `json:"provenance,omitempty"`
+	Fused                   bool     `json:"fused,omitempty"`
+	DerivedOnly             bool     `json:"derived_only,omitempty"`
+	MultiLevelConfirmed     bool     `json:"multi_level_confirmed,omitempty"`
+	MultiLevelConfirmedHint string   `json:"multi_level_confirmed_hint,omitempty"`
+}
+
+// EvidenceScoreDetails explains how evidence influenced refinement scoring.
+type EvidenceScoreDetails struct {
+	RawScore            float64 `json:"raw_score"`
+	Weight              float64 `json:"weight"`
+	WeightedScore       float64 `json:"weighted_score"`
+	DetectionLevels     int     `json:"detection_levels"`
+	PrimaryLevels       int     `json:"primary_levels,omitempty"`
+	DerivedLevels       int     `json:"derived_levels,omitempty"`
+	ProvenanceCount     int     `json:"provenance_count,omitempty"`
+	DerivedOnly         bool    `json:"derived_only,omitempty"`
+	MultiLevelConfirmed bool    `json:"multi_level_confirmed,omitempty"`
+	MultiLevelBonus     float64 `json:"multi_level_bonus,omitempty"`
+	ProvenanceBonus     float64 `json:"provenance_bonus,omitempty"`
+	DerivedPenalty      float64 `json:"derived_penalty,omitempty"`
+	StrategyBias        float64 `json:"strategy_bias,omitempty"`
+}
+
+// IsPresentationLevel reports whether a level is intended only for presentation.
+func IsPresentationLevel(level AnalysisLevel) bool {
+	role := strings.ToLower(strings.TrimSpace(level.Role))
+	truth := strings.ToLower(strings.TrimSpace(level.Truth))
+	name := strings.ToLower(strings.TrimSpace(level.Name))
+	if strings.Contains(role, "presentation") || strings.Contains(truth, "presentation") {
+		return true
+	}
+	return strings.Contains(name, "presentation") || strings.Contains(name, "display")
+}
+
+// IsDetectionLevel reports whether a level is intended for detection/analysis.
+func IsDetectionLevel(level AnalysisLevel) bool {
+	if IsPresentationLevel(level) {
+		return false
+	}
+	role := strings.ToLower(strings.TrimSpace(level.Role))
+	truth := strings.ToLower(strings.TrimSpace(level.Truth))
+	name := strings.ToLower(strings.TrimSpace(level.Name))
+	if strings.Contains(truth, "surveillance") {
+		return true
+	}
+	if role == "surveillance" || strings.HasPrefix(role, "surveillance-") {
+		return true
+	}
+	return strings.Contains(name, "surveillance")
+}
+
+func isPrimarySurveillanceLevel(level AnalysisLevel) bool {
+	role := strings.ToLower(strings.TrimSpace(level.Role))
+	name := strings.ToLower(strings.TrimSpace(level.Name))
+	return role == "surveillance" || name == "surveillance"
+}
+
+func isDerivedSurveillanceLevel(level AnalysisLevel) bool {
+	role := strings.ToLower(strings.TrimSpace(level.Role))
+	name := strings.ToLower(strings.TrimSpace(level.Name))
+	if strings.HasPrefix(role, "surveillance-") && role != "surveillance" {
+		return true
+	}
+	if strings.HasPrefix(name, "surveillance-") && name != "surveillance" {
+		return true
+	}
+	return strings.Contains(role, "lowres") || strings.Contains(name, "lowres") || strings.Contains(name, "derived")
+}
+
+func evidenceLevelKey(level AnalysisLevel) string {
+	if level.Name != "" {
+		return level.Name
+	}
+	if level.SampleRate > 0 && level.FFTSize > 0 {
+		return fmt.Sprintf("sr%d-fft%d", level.SampleRate, level.FFTSize)
+	}
+	return "unknown"
+}
+
+// CandidateEvidenceStateFor builds a fused evidence state from a candidate.
+func CandidateEvidenceStateFor(candidate Candidate) CandidateEvidenceState {
+	state := CandidateEvidenceState{}
+	if len(candidate.Evidence) == 0 {
+		return state
+	}
+	levelSet := map[string]struct{}{}
+	provenanceSet := map[string]struct{}{}
+	detectionLevels := map[string]struct{}{}
+	primaryLevels := map[string]struct{}{}
+	derivedLevels := map[string]struct{}{}
+	presentationLevels := map[string]struct{}{}
+	for _, ev := range candidate.Evidence {
+		levelKey := evidenceLevelKey(ev.Level)
+		levelSet[levelKey] = struct{}{}
+		if ev.Provenance != "" {
+			provenanceSet[ev.Provenance] = struct{}{}
+		}
+		if IsPresentationLevel(ev.Level) {
+			presentationLevels[levelKey] = struct{}{}
+			continue
+		}
+		if IsDetectionLevel(ev.Level) {
+			detectionLevels[levelKey] = struct{}{}
+			if isPrimarySurveillanceLevel(ev.Level) {
+				primaryLevels[levelKey] = struct{}{}
+			} else if isDerivedSurveillanceLevel(ev.Level) {
+				derivedLevels[levelKey] = struct{}{}
+			}
+		}
+	}
+	state.TotalLevelEntries = len(candidate.Evidence)
+	state.LevelCount = len(levelSet)
+	state.DetectionLevelCount = len(detectionLevels)
+	state.PrimaryLevelCount = len(primaryLevels)
+	state.DerivedLevelCount = len(derivedLevels)
+	state.PresentationLevelCount = len(presentationLevels)
+	state.Levels = sortedKeys(levelSet)
+	state.Provenance = sortedKeys(provenanceSet)
+	state.Fused = state.LevelCount > 1 || len(state.Provenance) > 1
+	state.DerivedOnly = state.DerivedLevelCount > 0 && state.PrimaryLevelCount == 0 && state.DetectionLevelCount == state.DerivedLevelCount
+	state.MultiLevelConfirmed = state.DetectionLevelCount >= 2
+	if state.MultiLevelConfirmed {
+		if state.PrimaryLevelCount > 0 && state.DerivedLevelCount > 0 {
+			state.MultiLevelConfirmedHint = "primary+derived"
+		} else {
+			state.MultiLevelConfirmedHint = "multi-detection"
+		}
+	}
+	return state
+}
+
+// RefreshCandidateEvidenceState updates the candidate's cached evidence summary.
+func RefreshCandidateEvidenceState(candidate *Candidate) {
+	if candidate == nil {
+		return
+	}
+	state := CandidateEvidenceStateFor(*candidate)
+	if state.TotalLevelEntries == 0 {
+		candidate.EvidenceState = nil
+		return
+	}
+	candidate.EvidenceState = &state
+}
+
+func sortedKeys(src map[string]struct{}) []string {
+	if len(src) == 0 {
+		return nil
+	}
+	out := make([]string, 0, len(src))
+	for k := range src {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}
diff --git a/internal/pipeline/scheduler.go b/internal/pipeline/scheduler.go
index cb563f4..2824c6c 100644
--- a/internal/pipeline/scheduler.go
+++ b/internal/pipeline/scheduler.go
@@ -20,11 +20,12 @@ type RefinementScoreModel struct {
 }
 
 type RefinementScoreDetails struct {
-	SNRScore       float64 `json:"snr_score"`
-	BandwidthScore float64 `json:"bandwidth_score"`
-	PeakScore      float64 `json:"peak_score"`
-	PolicyBoost    float64 `json:"policy_boost"`
-	EvidenceScore  float64 `json:"evidence_score"`
+	SNRScore       float64               `json:"snr_score"`
+	BandwidthScore float64               `json:"bandwidth_score"`
+	PeakScore      float64               `json:"peak_score"`
+	PolicyBoost    float64               `json:"policy_boost"`
+	EvidenceScore  float64               `json:"evidence_score"`
+	EvidenceDetail *EvidenceScoreDetails `json:"evidence_detail,omitempty"`
 }
 
 type RefinementScore struct {
@@ -114,35 +115,41 @@ func BuildRefinementPlan(candidates []Candidate, policy Policy) RefinementPlan {
 	scored := make([]ScheduledCandidate, 0, len(candidates))
 	workItems := make([]RefinementWorkItem, 0, len(candidates))
 	for _, c := range candidates {
-		if !candidateInMonitor(policy, c) {
+		candidate := c
+		RefreshCandidateEvidenceState(&candidate)
+		if !candidateInMonitor(policy, candidate) {
 			plan.DroppedByMonitor++
 			workItems = append(workItems, RefinementWorkItem{
-				Candidate: c,
+				Candidate: candidate,
 				Status:    RefinementStatusDropped,
 				Reason:    RefinementReasonMonitorGate,
 			})
 			continue
 		}
-		if c.SNRDb < policy.MinCandidateSNRDb {
+		if candidate.SNRDb < policy.MinCandidateSNRDb {
 			plan.DroppedBySNR++
 			workItems = append(workItems, RefinementWorkItem{
-				Candidate: c,
+				Candidate: candidate,
 				Status:    RefinementStatusDropped,
 				Reason:    RefinementReasonBelowSNR,
 			})
 			continue
 		}
-		snrScore := c.SNRDb * scoreModel.SNRWeight
+		snrScore := candidate.SNRDb * scoreModel.SNRWeight
 		bwScore := 0.0
 		peakScore := 0.0
-		policyBoost := CandidatePriorityBoost(policy, c.Hint)
-		if c.BandwidthHz > 0 {
-			bwScore = minFloat64(c.BandwidthHz/25000.0, 6) * scoreModel.BandwidthWeight
+		policyBoost := CandidatePriorityBoost(policy, candidate.Hint)
+		if candidate.BandwidthHz > 0 {
+			bwScore = minFloat64(candidate.BandwidthHz/25000.0, 6) * scoreModel.BandwidthWeight
 		}
-		if c.PeakDb > 0 {
-			peakScore = (c.PeakDb / 20.0) * scoreModel.PeakWeight
+		if candidate.PeakDb > 0 {
+			peakScore = (candidate.PeakDb / 20.0) * scoreModel.PeakWeight
 		}
-		evidenceScore := candidateEvidenceScore(c) * scoreModel.EvidenceWeight
+		rawEvidenceScore, evidenceDetail := candidateEvidenceScore(candidate, strategy)
+		evidenceDetail.Weight = scoreModel.EvidenceWeight
+		evidenceDetail.RawScore = rawEvidenceScore
+		evidenceDetail.WeightedScore = rawEvidenceScore * scoreModel.EvidenceWeight
+		evidenceScore := evidenceDetail.WeightedScore
 		priority := snrScore + bwScore + peakScore + policyBoost
 		priority += evidenceScore
 		score := &RefinementScore{
@@ -153,17 +160,18 @@ func BuildRefinementPlan(candidates []Candidate, policy Policy) RefinementPlan {
 				PeakScore:      peakScore,
 				PolicyBoost:    policyBoost,
 				EvidenceScore:  evidenceScore,
+				EvidenceDetail: &evidenceDetail,
 			},
 			Weights: &scoreModel,
 		}
 		scored = append(scored, ScheduledCandidate{
-			Candidate: c,
+			Candidate: candidate,
 			Priority:  priority,
 			Score:     score,
 			Breakdown: &score.Breakdown,
 		})
 		workItems = append(workItems, RefinementWorkItem{
-			Candidate: c,
+			Candidate: candidate,
 			Priority:  priority,
 			Score:     score,
 			Breakdown: &score.Breakdown,
@@ -250,12 +258,67 @@ func applyStrategyWeights(strategy string, model RefinementScoreModel) Refinemen
 	return model
 }
 
-func candidateEvidenceScore(candidate Candidate) float64 {
-	levels := CandidateEvidenceLevelCount(candidate)
-	if levels <= 1 {
-		return 0
+func candidateEvidenceScore(candidate Candidate, strategy string) (float64, EvidenceScoreDetails) {
+	state := CandidateEvidenceStateFor(candidate)
+	details := EvidenceScoreDetails{
+		DetectionLevels:     state.DetectionLevelCount,
+		PrimaryLevels:       state.PrimaryLevelCount,
+		DerivedLevels:       state.DerivedLevelCount,
+		ProvenanceCount:     len(state.Provenance),
+		DerivedOnly:         state.DerivedOnly,
+		MultiLevelConfirmed: state.MultiLevelConfirmed,
 	}
-	return float64(levels - 1)
+	score := 0.0
+	if state.MultiLevelConfirmed && state.DetectionLevelCount > 1 {
+		bonus := 0.85 * float64(state.DetectionLevelCount-1)
+		score += bonus
+		details.MultiLevelBonus = bonus
+	}
+	if len(state.Provenance) > 1 {
+		bonus := 0.15 * float64(len(state.Provenance)-1)
+		score += bonus
+		details.ProvenanceBonus = bonus
+	}
+	if state.DerivedOnly {
+		penalty := 0.35
+		score -= penalty
+		details.DerivedPenalty = -penalty
+	}
+	switch strings.ToLower(strings.TrimSpace(strategy)) {
+	case "multi-resolution", "multi", "multi-res", "multi_res":
+		if state.DerivedOnly {
+			bias := 0.2
+			score += bias
+			details.StrategyBias = bias
+		} else if state.MultiLevelConfirmed {
+			bias := 0.1
+			score += bias
+			details.StrategyBias = bias
+		}
+	case "digital-hunting":
+		if state.DerivedOnly {
+			bias := -0.15
+			score += bias
+			details.StrategyBias = bias
+		} else if state.MultiLevelConfirmed {
+			bias := 0.05
+			score += bias
+			details.StrategyBias = bias
+		}
+	case "archive-oriented":
+		if state.DerivedOnly {
+			bias := -0.1
+			score += bias
+			details.StrategyBias = bias
+		}
+	case "single-resolution":
+		if state.MultiLevelConfirmed {
+			bias := 0.05
+			score += bias
+			details.StrategyBias = bias
+		}
+	}
+	return score, details
 }
 
 func minFloat64(a, b float64) float64 {
diff --git a/internal/pipeline/scheduler_test.go b/internal/pipeline/scheduler_test.go
index e6fa4f4..38834f8 100644
--- a/internal/pipeline/scheduler_test.go
+++ b/internal/pipeline/scheduler_test.go
@@ -173,6 +173,60 @@ func TestScheduleCandidatesEvidenceBoost(t *testing.T) {
 	if plan.Ranked[0].Breakdown == nil || plan.Ranked[0].Breakdown.EvidenceScore <= 0 {
 		t.Fatalf("expected evidence score to be populated, got %+v", plan.Ranked[0].Breakdown)
 	}
+	if plan.Ranked[0].Breakdown.EvidenceDetail == nil || !plan.Ranked[0].Breakdown.EvidenceDetail.MultiLevelConfirmed {
+		t.Fatalf("expected evidence detail for multi-level candidate, got %+v", plan.Ranked[0].Breakdown)
+	}
+}
+
+func TestScheduleCandidatesDerivedOnlyPenalty(t *testing.T) {
+	policy := Policy{MaxRefinementJobs: 2, MinCandidateSNRDb: 0}
+	primary := Candidate{
+		ID:          1,
+		SNRDb:       10,
+		BandwidthHz: 12000,
+		Evidence: []LevelEvidence{
+			{Level: AnalysisLevel{Name: "surveillance", Role: "surveillance", Truth: "surveillance"}},
+		},
+	}
+	derived := Candidate{
+		ID:          2,
+		SNRDb:       10,
+		BandwidthHz: 12000,
+		Evidence: []LevelEvidence{
+			{Level: AnalysisLevel{Name: "surveillance-lowres", Role: "surveillance-lowres", Truth: "surveillance"}},
+		},
+	}
+	plan := BuildRefinementPlan([]Candidate{derived, primary}, policy)
+	if len(plan.Ranked) != 2 {
+		t.Fatalf("expected ranked candidates, got %d", len(plan.Ranked))
+	}
+	if plan.Ranked[0].Candidate.ID != primary.ID {
+		t.Fatalf("expected primary evidence to outrank derived-only, got %+v", plan.Ranked[0])
+	}
+}
+
+func TestScheduleCandidatesDerivedOnlyStrategyBias(t *testing.T) {
+	cand := Candidate{
+		ID:          1,
+		SNRDb:       9,
+		BandwidthHz: 12000,
+		Evidence: []LevelEvidence{
+			{Level: AnalysisLevel{Name: "surveillance-lowres", Role: "surveillance-lowres", Truth: "surveillance"}},
+		},
+	}
+	singlePlan := BuildRefinementPlan([]Candidate{cand}, Policy{MinCandidateSNRDb: 0})
+	multiPlan := BuildRefinementPlan([]Candidate{cand}, Policy{MinCandidateSNRDb: 0, SurveillanceStrategy: "multi-resolution"})
+	if len(singlePlan.Ranked) == 0 || len(multiPlan.Ranked) == 0 {
+		t.Fatalf("expected ranked candidates in both plans")
+	}
+	singleScore := singlePlan.Ranked[0].Breakdown.EvidenceScore
+	multiScore := multiPlan.Ranked[0].Breakdown.EvidenceScore
+	if multiScore <= singleScore {
+		t.Fatalf("expected multi-resolution strategy to improve derived-only evidence score, got %.3f vs %.3f", multiScore, singleScore)
+	}
+	if multiPlan.Ranked[0].Breakdown.EvidenceDetail == nil || multiPlan.Ranked[0].Breakdown.EvidenceDetail.StrategyBias <= 0 {
+		t.Fatalf("expected strategy bias detail for multi-resolution, got %+v", multiPlan.Ranked[0].Breakdown.EvidenceDetail)
+	}
 }
 
 func TestBuildRefinementPlanPriorityStats(t *testing.T) {
diff --git a/internal/pipeline/types.go b/internal/pipeline/types.go
index f453edb..70889e6 100644
--- a/internal/pipeline/types.go
+++ b/internal/pipeline/types.go
@@ -8,17 +8,18 @@ import (
 // Candidate is the coarse output of the surveillance detector.
 // It intentionally stays lightweight and cheap to produce.
 type Candidate struct {
-	ID          int64           `json:"id"`
-	CenterHz    float64         `json:"center_hz"`
-	BandwidthHz float64         `json:"bandwidth_hz"`
-	PeakDb      float64         `json:"peak_db"`
-	SNRDb       float64         `json:"snr_db"`
-	FirstBin    int             `json:"first_bin"`
-	LastBin     int             `json:"last_bin"`
-	NoiseDb     float64         `json:"noise_db,omitempty"`
-	Source      string          `json:"source,omitempty"`
-	Hint        string          `json:"hint,omitempty"`
-	Evidence    []LevelEvidence `json:"evidence,omitempty"`
+	ID            int64                   `json:"id"`
+	CenterHz      float64                 `json:"center_hz"`
+	BandwidthHz   float64                 `json:"bandwidth_hz"`
+	PeakDb        float64                 `json:"peak_db"`
+	SNRDb         float64                 `json:"snr_db"`
+	FirstBin      int                     `json:"first_bin"`
+	LastBin       int                     `json:"last_bin"`
+	NoiseDb       float64                 `json:"noise_db,omitempty"`
+	Source        string                  `json:"source,omitempty"`
+	Hint          string                  `json:"hint,omitempty"`
+	Evidence      []LevelEvidence         `json:"evidence,omitempty"`
+	EvidenceState *CandidateEvidenceState `json:"evidence_state,omitempty"`
 }
 
 // LevelEvidence captures which analysis level produced a candidate.