fix(ucca): guidance-intent erkennt direkt benannte Guidance-Dokumente

queryWantsGuidance verfehlte rein dokument-namige Fragen ("Welche Kriterien nennt WP248 ...", "Was sagt GL 07/2020 ..."): guidanceIntentSignals enthielt zwar Herausgeber (edpb/dsk/enisa) und Verben (empfiehlt/laut), aber keine Working-Paper-/Guideline-Identifier. Dadurch loeste der Authority-Lift nicht aus -> binding_law (bzw. im homogenen Korpus sogar off-domain MaschVO/CRA) verdraengte die Guidance aus den Top-K. Fix: WP2xx / GL 0x / "working paper" als Guidance-Signal ergaenzt. Generisch ueber alle WP-/GL-Dokumente, KEINE doc-spezifische Regel (Query->Intent, nicht Query->konkretes Dokument). Validierung (homogener Build-Korpus, bge-m3 + Qdrant Cosine): - 10 Hard Cases: 8/10 -> 10/10 (WP248/WP260 zurueck in Top-8) - ComplianceBench-100: 0/100 Norm-Fragen veraendert (Freeze-Regression gruen), 18/18 Guidance-Intent-Fragen verbessert (binding -> korrekte Guidance-Klasse) - Hybrid == Dense (Keyword-RRF war NICHT die Ursache, der Lift-Gate war es) Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-06-29 14:44:53 +02:00
17 changed files with 136 additions and 631 deletions
@@ -1,73 +0,0 @@
 package iace
 // P3: pin accepted proposer decisions into the GT gate.
 //
 // When a human accepts a proposal from the offline proposer (a dedup
 // supersession, a foreign-framing gate, a vocab→tag mapping, a coverage hazard),
 // they record an AcceptedPin. A pin is a tiny, machine-scoped invariant — "this
 // pattern MUST (or must NOT) fire for this machine" — that a test re-checks on
 // every run. This is what makes the library's growth COMPOUND into the gate
 // instead of silently eroding it: a future change that re-introduces a dropped
 // duplicate, un-gates a foreign pattern, or removes a coverage hazard breaks the
 // pin and fails CI.
 //
 // A single boolean covers all four proposal types:
 //   - dedup supersession accepted → DropPattern MustFire=false
 //   - foreign-framing gate accepted → foreign pattern MustFire=false
 //   - vocab→tag / coverage hazard accepted → the enabled pattern MustFire=true
 // AcceptedPin is one regression invariant for an accepted proposal.
 type AcceptedPin struct {
 	Pattern      string `json:"pattern"`
 	MustFire     bool   `json:"must_fire"`
 	Reason       string `json:"reason"`
 	FromProposal string `json:"from_proposal,omitempty"`
 }
 // PinSet is the accepted-pin registry for one machine (testdata/accepted_pins_*.json).
 type PinSet struct {
 	Machine string        `json:"machine"`
 	Pins    []AcceptedPin `json:"pins"`
 }
 // PinResult is the verdict for one pin against an engine run.
 type PinResult struct {
 	Pin    AcceptedPin
 	OK     bool
 	Detail string
 }
 // VerifyPins checks every pin against the set of pattern IDs the engine actually
 // fired for the machine. A pin holds iff the pattern's presence equals MustFire.
 func VerifyPins(pins []AcceptedPin, firedPatternIDs []string) []PinResult {
 	fired := make(map[string]bool, len(firedPatternIDs))
 	for _, id := range firedPatternIDs {
 		fired[id] = true
 	}
 	out := make([]PinResult, 0, len(pins))
 	for _, p := range pins {
 		got := fired[p.Pattern]
 		ok := got == p.MustFire
 		detail := "ok"
 		if !ok {
 			if p.MustFire {
 				detail = "expected to fire but did NOT — coverage/mapping regressed"
 			} else {
 				detail = "expected to be suppressed but FIRED — gate/supersession regressed"
 			}
 		}
 		out = append(out, PinResult{Pin: p, OK: ok, Detail: detail})
 	}
 	return out
 }
 // GenerateDedupPin turns an accepted (verdict=duplicate) dedup candidate into the
 // pin that protects the supersession: the dropped pattern must no longer fire.
 func GenerateDedupPin(c DedupCandidate) AcceptedPin {
 	return AcceptedPin{
 		Pattern:      c.DropPattern,
 		MustFire:     false,
 		Reason:       "accepted duplicate of " + c.KeepPattern + " (" + c.Category + ")",
 		FromProposal: "dedup " + c.DropPattern + " -> " + c.KeepPattern,
 	}
 }
@@ -1,63 +0,0 @@
 package iace
 import (
 	"encoding/json"
 	"os"
 	"path/filepath"
 	"testing"
 )
 func TestVerifyPins(t *testing.T) {
 	pins := []AcceptedPin{
 		{Pattern: "HPa", MustFire: true},
 		{Pattern: "HPb", MustFire: false},
 	}
 	res := VerifyPins(pins, []string{"HPa", "HPb"})
 	if !res[0].OK {
 		t.Errorf("HPa must_fire=true and it fired -> should be OK")
 	}
 	if res[1].OK {
 		t.Errorf("HPb must_fire=false but it fired -> should be VIOLATED")
 	}
 	res2 := VerifyPins(pins, []string{})
 	if res2[0].OK || !res2[1].OK {
 		t.Errorf("expected HPa violated + HPb ok, got %+v", res2)
 	}
 }
 func TestGenerateDedupPin(t *testing.T) {
 	pin := GenerateDedupPin(DedupCandidate{KeepPattern: "HP144", DropPattern: "HP013", Category: "electrical_hazard"})
 	if pin.Pattern != "HP013" || pin.MustFire {
 		t.Fatalf("want pin {HP013, must_fire=false}, got %+v", pin)
 	}
 }
 // TestWarewashing_AcceptedPins re-checks every accepted P1 supersession against the
 // live warewashing engine output. A future change that un-suppresses HP013/016/018
 // or drops HP2201/HP144 breaks a pin here — the gate compounds, not erodes.
 func TestWarewashing_AcceptedPins(t *testing.T) {
 	raw, err := os.ReadFile(filepath.Join("testdata", "accepted_pins_warewashing.json"))
 	if err != nil {
 		t.Fatalf("read pins: %v", err)
 	}
 	var ps PinSet
 	if err := json.Unmarshal(raw, &ps); err != nil {
 		t.Fatalf("parse pins: %v", err)
 	}
 	_, _, kept := warewashingEngineOutput()
 	firedIDs := make([]string, 0, len(kept))
 	for _, pm := range kept {
 		firedIDs = append(firedIDs, pm.PatternID)
 	}
 	ok := 0
 	for _, r := range VerifyPins(ps.Pins, firedIDs) {
 		if r.OK {
 			ok++
 			continue
 		}
 		t.Errorf("PIN VIOLATED: %s (must_fire=%v) — %s [%s]", r.Pin.Pattern, r.Pin.MustFire, r.Detail, r.Pin.Reason)
 	}
 	t.Logf("accepted pins for %q: %d/%d hold", ps.Machine, ok, len(ps.Pins))
 }
@@ -1,10 +0,0 @@
 {
  "machine": "Gewerbliche Untertisch-Geschirrspuelmaschine (vernetzt)",
  "pins": [
    {"pattern": "HP016", "must_fire": false, "reason": "generic hot-surface (Formwerkzeuge/Auspuffleitung framing) superseded by HP2201", "from_proposal": "P1 thermal supersession"},
    {"pattern": "HP018", "must_fire": false, "reason": "actuator-burn superseded by HP2201", "from_proposal": "P1 thermal supersession"},
    {"pattern": "HP013", "must_fire": false, "reason": "stored-energy Batterie/USV framing superseded by HP144", "from_proposal": "P1 stored-energy supersession"},
    {"pattern": "HP2201", "must_fire": true, "reason": "warewashing hot-surface (Boiler/Tank/Spuelkammer) must remain — it is the clean equivalent that replaces HP016/HP018", "from_proposal": "P1 thermal supersession"},
    {"pattern": "HP144", "must_fire": true, "reason": "residual-voltage (Frequenzumrichter/Zwischenkreis) must remain — clean equivalent that replaces HP013", "from_proposal": "P1 stored-energy supersession"}
  ]
 }
@@ -18,10 +18,7 @@ func TestGuidanceFixE2E(t *testing.T) {
 		t.Skip("set RUN_E2E=1 + QDRANT_URL/OLLAMA_URL to run")
 	}
 	c := NewLegalRAGClient()
-	coll := os.Getenv("E2E_COLLECTION")
+	coll := "bp_compliance_kb_2026_1_build"
 	if coll == "" {
 		coll = "bp_compliance_kb_2026_1_build"
 	}
 	cases := []struct{ id, q, expect string }{
 		{"GQ-0012", "Welche neun Kriterien nennt WP248 fuer ein voraussichtlich hohes Risiko?", "WP248"},
 		{"GQ-0013", "Ab wie vielen der WP248-Kriterien ist in der Regel eine Datenschutz-Folgenabschaetzung erforderlich?", "WP248"},
@@ -86,10 +83,7 @@ func TestBenchE2E(t *testing.T) {
 		t.Fatal(err)
 	}
 	c := NewLegalRAGClient()
-	coll := os.Getenv("E2E_COLLECTION")
+	coll := "bp_compliance_kb_2026_1_build"
 	if coll == "" {
 		coll = "bp_compliance_kb_2026_1_build"
 	}
 	fmt.Printf("### BENCH n=%d hybrid=%v\n", len(bench.Questions), os.Getenv("RAG_HYBRID_SEARCH") != "false")
 	for _, q := range bench.Questions {
 		res, err := c.SearchCollection(context.Background(), coll, q.Question, nil, 8)
@@ -1,52 +0,0 @@
 package ucca
 import "strings"
 // kbScopeTopics are high-precision data-protection / compliance topic markers that place a query in
 // the KB-2026.1 authoritative slice even when it does NOT name a regulation. Conservative by design:
 // an unmatched query falls back to the broad CE default (no regression) — the slice is only used when
 // the query is confidently in-scope.
 var kbScopeTopics = []string{
 	// DP-Guidance-Marker, die IN der Slice liegen (EDPB/DSK/WP/GL) — bewusst NICHT die generischen
 	// Verben aus guidanceIntentSignals (sagt/laut/empfiehlt/auslegung) und NICHT enisa/bsi/nist/owasp
 	// (die liegen im breiten CE-Pool, nicht in der Slice).
 	"edpb", "dsk", "datenschutzausschuss", "orientierungshilfe",
 	"wp2", "wp 2", "wp29", "working paper", "gl 0",
 	"datenschutz", "dsgvo", "gdpr", "dsfa", "folgenabschätzung", "folgenabschaetzung",
 	"einwilligung", "auftragsverarbeit", "betroffenenrecht", "auskunftsrecht",
 	"verarbeitungsverzeichnis", "datenschutzbeauftragt", "verzeichnis von verarbeitung",
 	"cookie", "tracking", "transparenzpflicht", "datenpanne", "meldepflicht",
 	"technische und organisatorische maßnahmen",
 	"cyber resilience", "schwachstelle", "vulnerability", "sicherheitsupdate",
 	"maschinensicherheit", "wesentliche veränderung", "wesentliche veraenderung",
 	"konformitätsbewertung", "konformitaetsbewertung", "ce-kennzeichnung",
 }
 // inKBScope reports whether the query belongs to the KB-2026.1 authoritative slice. True when it
 // names an in-slice regulation (detectRegulations), asks for guidance (EDPB/DSK/WP/GL), or hits a
 // data-protection / compliance topic marker.
 func inKBScope(query string) bool {
 	if len(detectRegulations(query)) > 0 {
 		return true
 	}
 	q := strings.ToLower(query)
 	for _, t := range kbScopeTopics {
 		if strings.Contains(q, t) {
 			return true
 		}
 	}
 	return false
 }
 // resolveCollection applies the Blue-Green „authoritative slice promotion" routing. An explicitly
 // requested collection is honoured unchanged; the DEFAULT (empty) request is routed to the KB-2026.1
 // slice when the query is in-scope, else to the broad CE default. Disable via RAG_KB_SCOPE_ROUTING=false.
 func (c *LegalRAGClient) resolveCollection(query, requested string) string {
 	if requested != "" {
 		return requested
 	}
 	if c.kbScopeRoutingEnabled && c.kbSliceCollection != "" && inKBScope(query) {
 		return c.kbSliceCollection
 	}
 	return c.collection
 }
@@ -1,101 +0,0 @@
 package ucca
 import (
 	"context"
 	"fmt"
 	"os"
 	"strings"
 	"testing"
 )
 func TestInKBScope(t *testing.T) {
 	inScope := []string{
 		"Welche neun Kriterien nennt WP248 fuer ein hohes Risiko?",
 		"Wie greifen CRA und Maschinenverordnung bei einer vernetzten Maschine ineinander?",
 		"Wann ist eine Datenschutz-Folgenabschaetzung erforderlich?",
 		"Welche Anforderungen stellt die DSGVO an die Einwilligung?",
 		"Brauche ich einen Datenschutzbeauftragten?",
 		"Wann muss eine aktiv ausgenutzte Schwachstelle gemeldet werden?",
 	}
 	outScope := []string{
 		"Welche OWASP-Kontrollen gibt es fuer Authentifizierung?",
 		"Was sagt NIST SP 800-53 zu Access Control?",
 		"Wie funktioniert ISO 27001 Zertifizierung?",
 		"Welche IFRS-Standards gelten fuer Leasing?",
 	}
 	for _, q := range inScope {
 		if !inKBScope(q) {
 			t.Errorf("inKBScope(%q) = false, want true", q)
 		}
 	}
 	for _, q := range outScope {
 		if inKBScope(q) {
 			t.Errorf("inKBScope(%q) = true, want false", q)
 		}
 	}
 }
 func TestResolveCollection(t *testing.T) {
 	c := &LegalRAGClient{collection: "bp_compliance_ce", kbSliceCollection: "kb_2026_1_build", kbScopeRoutingEnabled: true}
 	if got := c.resolveCollection("Welche Kriterien nennt WP248?", ""); got != "kb_2026_1_build" {
 		t.Errorf("in-scope default -> %s, want kb_2026_1_build", got)
 	}
 	if got := c.resolveCollection("Was sagt NIST SP 800-53?", ""); got != "bp_compliance_ce" {
 		t.Errorf("out-of-scope default -> %s, want bp_compliance_ce", got)
 	}
 	if got := c.resolveCollection("Welche Kriterien nennt WP248?", "explicit_coll"); got != "explicit_coll" {
 		t.Errorf("explicit request must be honoured -> %s", got)
 	}
 	c.kbScopeRoutingEnabled = false
 	if got := c.resolveCollection("Welche Kriterien nennt WP248?", ""); got != "bp_compliance_ce" {
 		t.Errorf("disabled routing -> %s, want bp_compliance_ce", got)
 	}
 }
 // TestKBScopeRoutingE2E (RUN_E2E=1) verifies the routing against the REAL collections: a default
 // Search() of an in-scope query must hit the KB-2026.1 slice (WP248/MaschVO live there but NOT in
 // the broad CE pool = clean discriminator); an out-of-scope query stays on CE.
 func TestKBScopeRoutingE2E(t *testing.T) {
 	if os.Getenv("RUN_E2E") != "1" {
 		t.Skip("set RUN_E2E=1 + QDRANT_URL/OLLAMA_URL/QDRANT_API_KEY")
 	}
 	c := NewLegalRAGClient()
 	cases := []struct {
 		q         string
 		wantToken string // expected in top-8 when routed to the slice
 		wantInKB  bool
 	}{
 		{"Welche neun Kriterien nennt WP248 fuer ein voraussichtlich hohes Risiko?", "WP248", true},
 		{"Welche grundlegenden Sicherheits- und Gesundheitsschutzanforderungen enthaelt Anhang III der Maschinenverordnung?", "MASCH", true},
 		{"Wie greifen CRA und Maschinenverordnung bei einer vernetzten Maschine ineinander?", "MASCH", true},
 		{"Was sagt NIST SP 800-53 zu Access Control?", "", false},
 	}
 	for _, tc := range cases {
 		routed := c.resolveCollection(tc.q, "")
 		res, err := c.Search(context.Background(), tc.q, nil, 8)
 		if err != nil {
 			t.Fatalf("%q: %v", tc.q, err)
 		}
 		codes := map[string]bool{}
 		for _, r := range res {
 			codes[strings.ToUpper(r.RegulationCode)] = true
 		}
 		hit := false
 		if tc.wantToken != "" {
 			for cd := range codes {
 				if strings.Contains(cd, tc.wantToken) {
 					hit = true
 					break
 				}
 			}
 		}
 		col := make([]string, 0, len(codes))
 		for cd := range codes {
 			col = append(col, cd)
 		}
 		fmt.Printf("inKB=%-5v routed=%-16s wantTok=%-6s found=%-5v | %v\n", tc.wantInKB, routed, tc.wantToken, hit, col)
 		if tc.wantInKB && tc.wantToken != "" && !hit {
 			t.Errorf("%q routed to %s but %s not in top-8 (slice not active?)", tc.q, routed, tc.wantToken)
 		}
 	}
 }
@@ -21,12 +21,6 @@ type LegalRAGClient struct {
 	textIndexEnsured map[string]bool
 	hybridEnabled    bool
 	graphEnabled     bool
 	// Blue-Green „authoritative slice promotion" (additiv, KEIN CE-Ersatz): faellt eine Query
 	// in den KB-2026.1-Scope (DP/CRA/MaschVO/NIS2/DataAct/DORA/AIAct + EDPB/DSK-Guidance), wird
 	// die hochwertige Slice-Collection abgefragt; sonst bleibt der breite Default (bp_compliance_ce).
 	kbSliceCollection     string
 	kbScopeRoutingEnabled bool
 }
 // NewLegalRAGClient creates a new Legal RAG client using Ollama bge-m3 embeddings.
@@ -51,14 +45,6 @@ func NewLegalRAGClient() *LegalRAGClient {
 	// zur Begruendung/Vollstaendigkeit genutzt, nicht zur Pool-Expansion (Default).
 	graphEnabled := os.Getenv("RAG_GRAPH_EXPANSION") == "true"
 	// KB-2026.1 authoritative slice (Blue-Green, additiv). Routing default AN; Rollback ohne
 	// Redeploy ueber RAG_KB_SCOPE_ROUTING=false (dann faellt alles auf den CE-Default zurueck).
 	kbSlice := os.Getenv("RAG_KB_SLICE_COLLECTION")
 	if kbSlice == "" {
 		kbSlice = "kb_2026_1_build"
 	}
 	kbScopeRouting := os.Getenv("RAG_KB_SCOPE_ROUTING") != "false"
 	return &LegalRAGClient{
 		qdrantURL:        qdrantURL,
 		qdrantAPIKey:     qdrantAPIKey,
@@ -68,8 +54,6 @@ func NewLegalRAGClient() *LegalRAGClient {
 		textIndexEnsured: make(map[string]bool),
 		hybridEnabled:    hybridEnabled,
 		graphEnabled:     graphEnabled,
 		kbSliceCollection:     kbSlice,
 		kbScopeRoutingEnabled: kbScopeRouting,
 		httpClient: &http.Client{
 			Timeout: 60 * time.Second,
 		},
@@ -79,32 +63,21 @@ func NewLegalRAGClient() *LegalRAGClient {
 // SearchCollection queries a specific Qdrant collection for relevant passages.
 // If collection is empty, it falls back to the default collection (bp_compliance_ce).
 func (c *LegalRAGClient) SearchCollection(ctx context.Context, collection string, query string, regulationIDs []string, topK int) ([]LegalSearchResult, error) {
-	return c.searchInternal(ctx, c.resolveCollection(query, collection), query, regulationIDs, topK)
+	if collection == "" {
 		collection = c.collection
 	}
 	return c.searchInternal(ctx, collection, query, regulationIDs, topK)
 }
-// Search queries the compliance corpus for relevant passages. The target collection is resolved by
+// Search queries the compliance CE corpus for relevant passages.
 // the Blue-Green slice routing: the KB-2026.1 slice for in-scope queries, else the broad CE default.
 func (c *LegalRAGClient) Search(ctx context.Context, query string, regulationIDs []string, topK int) ([]LegalSearchResult, error) {
-	return c.searchInternal(ctx, c.resolveCollection(query, ""), query, regulationIDs, topK)
+	return c.searchInternal(ctx, c.collection, query, regulationIDs, topK)
 }
 // searchInternal performs the actual search against a given collection.
 // If hybrid search is enabled, it uses the Qdrant Query API with RRF fusion
 // (dense + full-text). Falls back to dense-only /points/search on failure.
 func (c *LegalRAGClient) searchInternal(ctx context.Context, collection string, query string, regulationIDs []string, topK int) ([]LegalSearchResult, error) {
 	// Multi-Regulation-Retrieval: nennt die Query EXPLIZIT >=2 Regelwerke (z.B. "CRA und
 	// Maschinenverordnung"), wird pro Regelwerk separat retrieved + gemergt, damit BEIDE
 	// Domaenen im Prompt landen statt nur der keyword-dominanten. Generisch (Query->Regelwerke,
 	// keine doc-spezifische Logik); nur wenn der Caller nicht ohnehin schon auf Regulierungen
 	// filtert. Best-effort: leeres/fehlerhaftes Multi-Ergebnis faellt auf die Standardsuche zurueck.
 	if len(regulationIDs) == 0 {
 		if regs := detectRegulations(query); len(regs) >= 2 {
 			if mr, mErr := c.searchMultiRegulation(ctx, collection, query, regs, topK); mErr == nil && len(mr) > 0 {
 				return mr, nil
 			}
 		}
 	}
 	embedding, err := c.generateEmbedding(ctx, query)
 	if err != nil {
 		return nil, fmt.Errorf("failed to generate embedding: %w", err)
@@ -150,7 +123,43 @@ func (c *LegalRAGClient) searchInternal(ctx context.Context, collection string,
 		hits = c.expandViaGraph(ctx, collection, hits)
 	}
-	results := hitsToResults(hits)
+	results := make([]LegalSearchResult, len(hits))
 	for i, hit := range hits {
 		// Legal-Metadaten nach rag_reingest_spec.md §2: bevorzugt die normalisierten Felder
 		// (article_label/regulation_code/article/...); Fallback auf alte Feldnamen, solange der
 		// Korpus noch nicht re-ingestiert ist (regulation_id, section="§ 38").
 		regCode := getString(hit.Payload, "regulation_code")
 		if regCode == "" {
 			regCode = getString(hit.Payload, "regulation_id")
 		}
 		article := getString(hit.Payload, "article")
 		if article == "" {
 			article = getString(hit.Payload, "section")
 		}
 		results[i] = LegalSearchResult{
 			Text:            getString(hit.Payload, "chunk_text"),
 			RegulationCode:  regCode,
 			RegulationName:  getString(hit.Payload, "regulation_name_de"),
 			RegulationShort: getString(hit.Payload, "regulation_short"),
 			Category:        getString(hit.Payload, "category"),
 			ArticleLabel:    getString(hit.Payload, "article_label"),
 			Article:         article,
 			Paragraph:       getString(hit.Payload, "paragraph"),
 			Sub:             getString(hit.Payload, "sub"),
 			IsRecital:       getBool(hit.Payload, "is_recital"),
 			CitationStyle:   getString(hit.Payload, "citation_style"),
 			Pages:           getIntSlice(hit.Payload, "pages"),
 			SourceURL:       getString(hit.Payload, "source"),
 			Score:           hit.Score,
 			AuthorityWeight: getInt(hit.Payload, "authority_weight"),
 			SourceClass:     getString(hit.Payload, "source_class"),
 			Jurisdiction:    getString(hit.Payload, "jurisdiction"),
 			CitationUnit:    getString(hit.Payload, "citation_unit"),
 			ReferencesOut:   getStringSlice(hit.Payload, "references_out"),
 			ReferencesIn:    getStringSlice(hit.Payload, "references_in"),
 			Superseded:      getString(hit.Payload, "status") == "superseded",
 		}
 	}
 	// Authority-aware Re-Ranking: bindendes Recht der passenden Jurisdiktion/Domaene nach
 	// oben, Guidance/Fremdrecht/Off-Domain runter (nichts wird geloescht). Reihenfolge only,
@@ -122,14 +122,12 @@ func (c *LegalRAGClient) searchHybrid(ctx context.Context, collection string, em
 	}
 	if len(regulationIDs) > 0 {
-		// Match BOTH the legacy field (regulation_id) and the normalized field
+		conditions := make([]qdrantCondition, len(regulationIDs))
-		// (regulation_code) so per-regulation filtering works on the re-ingested corpus too.
+		for i, regID := range regulationIDs {
-		conditions := make([]qdrantCondition, 0, len(regulationIDs)*2)
+			conditions[i] = qdrantCondition{
-		for _, regID := range regulationIDs {
+				Key:   "regulation_id",
-			conditions = append(conditions,
+				Match: qdrantMatch{Value: regID},
-				qdrantCondition{Key: "regulation_id", Match: qdrantMatch{Value: regID}},
+			}
 				qdrantCondition{Key: "regulation_code", Match: qdrantMatch{Value: regID}},
 			)
 		}
 		queryReq.Filter = &qdrantFilter{Should: conditions}
 	}
@@ -177,14 +175,12 @@ func (c *LegalRAGClient) searchDense(ctx context.Context, collection string, emb
 	}
 	if len(regulationIDs) > 0 {
-		// Match BOTH the legacy field (regulation_id) and the normalized field
+		conditions := make([]qdrantCondition, len(regulationIDs))
-		// (regulation_code) so per-regulation filtering works on the re-ingested corpus too.
+		for i, regID := range regulationIDs {
-		conditions := make([]qdrantCondition, 0, len(regulationIDs)*2)
+			conditions[i] = qdrantCondition{
-		for _, regID := range regulationIDs {
+				Key:   "regulation_id",
-			conditions = append(conditions,
+				Match: qdrantMatch{Value: regID},
-				qdrantCondition{Key: "regulation_id", Match: qdrantMatch{Value: regID}},
+			}
 				qdrantCondition{Key: "regulation_code", Match: qdrantMatch{Value: regID}},
 			)
 		}
 		searchReq.Filter = &qdrantFilter{Should: conditions}
 	}
@@ -1,143 +0,0 @@
 package ucca
 import (
 	"context"
 	"fmt"
 	"strings"
 )
 // multiRegMinPerRegulation is the minimum number of hits fetched per named regulation, so
 // each domain is fairly represented even when topK/len(regs) would be tiny.
 const multiRegMinPerRegulation = 3
 // regulationCatalog maps a regulation to (a) the aliases that signal it is EXPLICITLY named
 // in a query and (b) the regulation_code/regulation_id values used to filter the corpus.
 // Deterministic + generic: a query naming >=2 regulations triggers per-regulation retrieval
 // so a cross-regulation question returns every named domain — NOT a doc-specific rule.
 var regulationCatalog = []struct {
 	Canonical  string
 	Aliases    []string
 	CodeValues []string
 }{
 	{"CRA", []string{"cra", "cyber resilience"}, []string{"CRA"}},
 	{"MaschVO", []string{"maschinenverordnung", "maschvo", "machinery regulation"}, []string{"MASCHVO", "MaschVO"}},
 	{"NIS2", []string{"nis2", "nis-2", "nis 2"}, []string{"NIS2"}},
 	{"DORA", []string{"dora"}, []string{"DORA"}},
 	{"Data Act", []string{"data act", "datengesetz"}, []string{"DATA ACT", "DataAct"}},
 	{"AI Act", []string{"ai act", "ki-vo", "ki-verordnung", "ai-verordnung"}, []string{"AI ACT", "AIAct"}},
 	{"DSGVO", []string{"dsgvo", "gdpr"}, []string{"DSGVO"}},
 	{"TDDDG", []string{"tdddg"}, []string{"TDDDG"}},
 	{"BDSG", []string{"bdsg"}, []string{"BDSG"}},
 }
 type detectedRegulation struct {
 	Canonical  string
 	CodeValues []string
 }
 // detectRegulations returns the DISTINCT regulations explicitly named in the query. >=2 of
 // them is the trigger for multi-regulation retrieval. Pure + deterministic, no LLM.
 func detectRegulations(query string) []detectedRegulation {
 	q := strings.ToLower(query)
 	var out []detectedRegulation
 	for _, r := range regulationCatalog {
 		for _, a := range r.Aliases {
 			if strings.Contains(q, a) {
 				out = append(out, detectedRegulation{Canonical: r.Canonical, CodeValues: r.CodeValues})
 				break
 			}
 		}
 	}
 	return out
 }
 func hitID(h qdrantSearchHit) string { return fmt.Sprintf("%v", h.ID) }
 // searchMultiRegulation retrieves each explicitly-named regulation SEPARATELY (per-regulation
 // filter) and merges, so a cross-regulation query ("Wie greifen CRA und MaschVO ineinander?")
 // returns BOTH domains in the prompt instead of only the keyword-dominant one. Generic over any
 // named pair (DSGVO+TDDDG, CRA+NIS2, DORA+NIS2, AI Act+DSGVO, ...). The merged pool is
 // authority-reranked once. Pure pool-construction; topK contract preserved.
 func (c *LegalRAGClient) searchMultiRegulation(ctx context.Context, collection, query string, regs []detectedRegulation, topK int) ([]LegalSearchResult, error) {
 	embedding, err := c.generateEmbedding(ctx, query)
 	if err != nil {
 		return nil, fmt.Errorf("failed to generate embedding: %w", err)
 	}
 	perReg := topK / len(regs)
 	if perReg < multiRegMinPerRegulation {
 		perReg = multiRegMinPerRegulation
 	}
 	var merged []qdrantSearchHit
 	seen := make(map[string]bool)
 	for _, r := range regs {
 		var hits []qdrantSearchHit
 		if c.hybridEnabled {
 			if h, hErr := c.searchHybrid(ctx, collection, embedding, r.CodeValues, perReg); hErr == nil {
 				hits = h
 			}
 		}
 		if hits == nil {
 			if h, dErr := c.searchDense(ctx, collection, embedding, r.CodeValues, perReg); dErr == nil {
 				hits = h
 			}
 		}
 		for _, h := range hits {
 			id := hitID(h)
 			if seen[id] {
 				continue
 			}
 			seen[id] = true
 			merged = append(merged, h)
 		}
 	}
 	if len(merged) == 0 {
 		return nil, fmt.Errorf("multi-regulation search returned no hits")
 	}
 	results := hitsToResults(merged)
 	results = rerankByAuthority(query, results)
 	if topK > 0 && len(results) > topK {
 		results = results[:topK]
 	}
 	return results, nil
 }
 // hitsToResults maps raw Qdrant hits to LegalSearchResult, preferring the normalized payload
 // fields (regulation_code/article_label/...) with fallback to the legacy names (regulation_id,
 // section) while the corpus is mid-re-ingestion. Shared by searchInternal + searchMultiRegulation.
 func hitsToResults(hits []qdrantSearchHit) []LegalSearchResult {
 	results := make([]LegalSearchResult, len(hits))
 	for i, hit := range hits {
 		regCode := getString(hit.Payload, "regulation_code")
 		if regCode == "" {
 			regCode = getString(hit.Payload, "regulation_id")
 		}
 		article := getString(hit.Payload, "article")
 		if article == "" {
 			article = getString(hit.Payload, "section")
 		}
 		results[i] = LegalSearchResult{
 			Text:            getString(hit.Payload, "chunk_text"),
 			RegulationCode:  regCode,
 			RegulationName:  getString(hit.Payload, "regulation_name_de"),
 			RegulationShort: getString(hit.Payload, "regulation_short"),
 			Category:        getString(hit.Payload, "category"),
 			ArticleLabel:    getString(hit.Payload, "article_label"),
 			Article:         article,
 			Paragraph:       getString(hit.Payload, "paragraph"),
 			Sub:             getString(hit.Payload, "sub"),
 			IsRecital:       getBool(hit.Payload, "is_recital"),
 			CitationStyle:   getString(hit.Payload, "citation_style"),
 			Pages:           getIntSlice(hit.Payload, "pages"),
 			SourceURL:       getString(hit.Payload, "source"),
 			Score:           hit.Score,
 			AuthorityWeight: getInt(hit.Payload, "authority_weight"),
 			SourceClass:     getString(hit.Payload, "source_class"),
 			Jurisdiction:    getString(hit.Payload, "jurisdiction"),
 			CitationUnit:    getString(hit.Payload, "citation_unit"),
 			ReferencesOut:   getStringSlice(hit.Payload, "references_out"),
 			ReferencesIn:    getStringSlice(hit.Payload, "references_in"),
 			Superseded:      getString(hit.Payload, "status") == "superseded",
 		}
 	}
 	return results
 }
@@ -1,92 +0,0 @@
 package ucca
 import (
 	"context"
 	"fmt"
 	"os"
 	"strings"
 	"testing"
 )
 // TestDetectRegulations is a pure unit test of the multi-regulation TRIGGER (no Qdrant):
 // only an explicit naming of >=2 regulations enables multi-regulation retrieval. A single
 // named regulation, or a topical question that doesn't name one, stays single-domain.
 func TestDetectRegulations(t *testing.T) {
 	cases := []struct {
 		q    string
 		want int
 	}{
 		{"Welche neun Kriterien nennt WP248 fuer ein voraussichtlich hohes Risiko?", 0},
 		{"Welche Anforderungen gelten fuer wesentliche Veraenderungen einer Maschine?", 0}, // "Maschine" != MaschVO
 		{"Benoetigt eine SPS ohne Netzwerkanschluss eine CRA-Bewertung?", 1},               // 1 -> single
 		{"Wie greifen CRA und Maschinenverordnung bei einer vernetzten Maschine ineinander?", 2},
 		{"Wie greifen DSGVO und TDDDG bei der Nutzung von Cookies ineinander?", 2},
 		{"Wie verhalten sich DORA und NIS2 fuer ein Finanzunternehmen?", 2},
 		{"Wie greifen AI Act und DSGVO bei einem KI-System ineinander?", 2},
 	}
 	for _, c := range cases {
 		if got := len(detectRegulations(c.q)); got != c.want {
 			t.Errorf("detectRegulations(%q) = %d, want %d", c.q, got, c.want)
 		}
 	}
 }
 // TestMultiRegE2E (RUN_E2E=1) verifies against the build collection that an explicit
 // cross-regulation query returns BOTH named domains in the top-K — the core acceptance
 // gate for multi-regulation retrieval.
 func TestMultiRegE2E(t *testing.T) {
 	if os.Getenv("RUN_E2E") != "1" {
 		t.Skip("set RUN_E2E=1 + QDRANT_URL/OLLAMA_URL")
 	}
 	c := NewLegalRAGClient()
 	coll := os.Getenv("E2E_COLLECTION")
 	if coll == "" {
 		coll = "bp_compliance_kb_2026_1_build"
 	}
 	cases := []struct {
 		id   string
 		q    string
 		want []string
 	}{
 		{"GQ-0070 CRA+MaschVO", "Wie greifen CRA und Maschinenverordnung bei einer vernetzten Maschine ineinander?", []string{"CRA", "MASCH"}},
 		{"DSGVO+TDDDG", "Wie greifen DSGVO und TDDDG bei der Nutzung von Cookies und Tracking-Technologien ineinander?", []string{"DSGVO", "TDDDG"}},
 		{"CRA+NIS2", "Wie verhalten sich CRA und NIS2 bei einem vernetzten Produkt eines wichtigen Unternehmens zueinander?", []string{"CRA", "NIS2"}},
 		{"DORA+NIS2", "Wie greifen DORA und NIS2 bei einem Finanzunternehmen ineinander?", []string{"DORA", "NIS2"}},
 		{"AI Act+DSGVO", "Wie greifen AI Act und DSGVO bei einem KI-System ineinander, das personenbezogene Daten verarbeitet?", []string{"AI ACT", "DSGVO"}},
 	}
 	for _, tc := range cases {
 		res, err := c.SearchCollection(context.Background(), coll, tc.q, nil, 8)
 		if err != nil {
 			t.Fatalf("%s: %v", tc.id, err)
 		}
 		present := map[string]bool{}
 		for _, r := range res {
 			present[strings.ToUpper(r.RegulationCode)] = true
 		}
 		ok := true
 		for _, w := range tc.want {
 			found := false
 			for cd := range present {
 				if strings.Contains(cd, w) {
 					found = true
 					break
 				}
 			}
 			if !found {
 				ok = false
 			}
 		}
 		codes := make([]string, 0, len(present))
 		for cd := range present {
 			codes = append(codes, cd)
 		}
 		status := "OK"
 		if !ok {
 			status = "FAIL"
 		}
 		fmt.Printf("%-22s want=%v  present=%v  %s\n", tc.id, tc.want, codes, status)
 		if !ok {
 			t.Errorf("%s: not all named regulations in top-8 (want %v, got %v)", tc.id, tc.want, codes)
 		}
 	}
 }
@@ -162,7 +162,7 @@ async def update_ai_system(
    db: Session = Depends(get_db),
 ):
    """Update an AI system."""
-    from datetime import datetime, timezone
+    from datetime import datetime
    system = db.query(AISystemDB).filter(AISystemDB.id == system_id).first()
    if not system:
@@ -226,7 +226,7 @@ async def assess_ai_system(
    db: Session = Depends(get_db),
 ):
    """Run AI Act risk assessment for an AI system."""
-    from datetime import datetime, timezone
+    from datetime import datetime
    system = db.query(AISystemDB).filter(AISystemDB.id == system_id).first()
    if not system:
@@ -47,8 +47,6 @@ from compliance.services.canonical_control_service import (
    _control_row,  # re-exported for legacy test imports
 )
 logger = logging.getLogger(__name__)
 router = APIRouter(prefix="/v1/canonical", tags=["canonical-controls"])
@@ -14,7 +14,7 @@ Endpoints:
 """
 import logging
-from datetime import datetime, date, timedelta, timezone
+from datetime import datetime, date, timedelta
 from calendar import month_abbr
 from typing import Optional, Dict, Any, List
 from decimal import Decimal
@@ -26,11 +26,10 @@ versions). Module-level helpers re-exported for legacy tests.
 import logging
 from typing import Any, List, Optional
-from fastapi import APIRouter, Depends, HTTPException, Query
+from fastapi import APIRouter, Depends, Query
 from pydantic import BaseModel
 from fastapi.responses import Response
 from sqlalchemy.orm import Session
 from sqlalchemy import text
 from classroom_engine.database import get_db
 from compliance.api._http_errors import translate_domain_errors
@@ -485,7 +484,6 @@ async def list_dsfas(
 async def create_dsfa(
    request: DSFACreate,
    tenant_id: Optional[str] = Query(None),
    db: Session = Depends(get_db),
    service: DSFAService = Depends(get_dsfa_service),
 ) -> dict[str, Any]:
    """Neue DSFA erstellen."""
@@ -16,11 +16,6 @@ from the legacy path.
 """
 import logging
 import os
 import json
 import hashlib
 import uuid as uuid_module
 from datetime import datetime, timedelta
 from typing import Any, Optional
 from fastapi import APIRouter, Depends, File, HTTPException, Query, UploadFile
@@ -35,15 +30,14 @@ from ..db import (
    EvidenceConfidenceEnum,
    EvidenceTruthStatusEnum,
 )
-from ..db.models import EvidenceDB, AuditTrailDB
+from ..db.models import EvidenceDB, ControlDB, AuditTrailDB
 from ..services.auto_risk_updater import AutoRiskUpdater
-from ..services.evidence_service import EvidenceService, _update_risks as _update_risks_impl
+from ..services.evidence_service import EvidenceService
 from .schemas import (
    EvidenceCreate, EvidenceResponse, EvidenceListResponse,
    EvidenceRejectRequest,
 )
 from .audit_trail_utils import log_audit_trail
 from ._http_errors import translate_domain_errors
 logger = logging.getLogger(__name__)
 router = APIRouter(tags=["compliance-evidence"])
@@ -152,7 +146,6 @@ async def list_evidence(
    status: Optional[str] = None,
    page: Optional[int] = Query(None, ge=1, description="Page number (1-based)"),
    limit: Optional[int] = Query(None, ge=1, le=500, description="Items per page"),
    db: Session = Depends(get_db),
    service: EvidenceService = Depends(get_evidence_service),
 ) -> EvidenceListResponse:
    """List evidence with optional filters and pagination."""
@@ -193,11 +186,9 @@ async def list_evidence(
@router.post("/evidence", response_model=EvidenceResponse)
 async def create_evidence(
    evidence_data: EvidenceCreate,
    db: Session = Depends(get_db),
    service: EvidenceService = Depends(get_evidence_service),
 ) -> EvidenceResponse:
    """Create new evidence record."""
    dsms_cid = None
    repo = EvidenceRepository(db)
    # Get control UUID
@@ -266,7 +257,6 @@ async def create_evidence(
@router.delete("/evidence/{evidence_id}")
 async def delete_evidence(
    evidence_id: str,
    db: Session = Depends(get_db),
    service: EvidenceService = Depends(get_evidence_service),
 ) -> dict[str, Any]:
    """Delete an evidence record."""
@@ -285,7 +275,6 @@ async def upload_evidence(
    title: str = Query(...),
    file: UploadFile = File(...),
    description: Optional[str] = Query(None),
    db: Session = Depends(get_db),
    service: EvidenceService = Depends(get_evidence_service),
 ) -> EvidenceResponse:
    """Upload evidence file."""
@@ -685,7 +674,6 @@ async def collect_ci_evidence(
 async def get_ci_evidence_status(
    control_id: Optional[str] = Query(None, description="Filter by control ID"),
    days: int = Query(30, description="Look back N days"),
    db: Session = Depends(get_db),
    service: EvidenceService = Depends(get_evidence_service),
 ) -> dict[str, Any]:
    """Get CI/CD evidence collection status overview."""
@@ -693,8 +681,70 @@ async def get_ci_evidence_status(
        return service.ci_status(control_id, days)
-# (Alte CI-Status-Implementierung entfernt — unerreichbarer Code nach `return
+# ----------------------------------------------------------------------------
-# service.ci_status(...)`; durch den Service ersetzt, `query` war nie initialisiert.)
+# Legacy re-exports for tests that import helpers directly.
 # ----------------------------------------------------------------------------
    if control_id:
        ctrl_repo = ControlRepository(db)
        control = ctrl_repo.get_by_control_id(control_id)
        if control:
            query = query.filter(EvidenceDB.control_id == control.id)
    evidence_list = query.order_by(EvidenceDB.collected_at.desc()).limit(100).all()
    # Group by control and calculate stats
    control_stats = defaultdict(lambda: {
        "total": 0,
        "valid": 0,
        "failed": 0,
        "last_collected": None,
        "evidence": [],
    })
    for e in evidence_list:
        # Get control_id string
        control = db.query(ControlDB).filter(ControlDB.id == e.control_id).first()
        ctrl_id = control.control_id if control else "unknown"
        stats = control_stats[ctrl_id]
        stats["total"] += 1
        if e.status:
            if e.status.value == "valid":
                stats["valid"] += 1
            elif e.status.value == "failed":
                stats["failed"] += 1
        if not stats["last_collected"] or e.collected_at > stats["last_collected"]:
            stats["last_collected"] = e.collected_at
        # Add evidence summary
        stats["evidence"].append({
            "id": e.id,
            "type": e.evidence_type,
            "status": e.status.value if e.status else None,
            "collected_at": e.collected_at.isoformat() if e.collected_at else None,
            "ci_job_id": e.ci_job_id,
        })
    # Convert to list and sort
    result = []
    for ctrl_id, stats in control_stats.items():
        result.append({
            "control_id": ctrl_id,
            "total_evidence": stats["total"],
            "valid_count": stats["valid"],
            "failed_count": stats["failed"],
            "last_collected": stats["last_collected"].isoformat() if stats["last_collected"] else None,
            "recent_evidence": stats["evidence"][:5],
        })
    result.sort(key=lambda x: x["last_collected"] or "", reverse=True)
    return {
        "period_days": days,
        "total_evidence": len(evidence_list),
        "controls": result,
    }
 # ============================================================================
@@ -722,7 +772,6 @@ async def review_evidence(
    approval_status='first_approved'. A second (different) reviewer then
    sets second_reviewer and approval_status='approved'.
    """
    dsms_cid = None
    evidence = db.query(EvidenceDB).filter(EvidenceDB.id == evidence_id).first()
    if not evidence:
        raise HTTPException(status_code=404, detail=f"Evidence {evidence_id} not found")
@@ -802,7 +851,6 @@ async def reject_evidence(
    db: Session = Depends(get_db),
 ):
    """Reject evidence (sets approval_status='rejected')."""
    dsms_cid = None
    evidence = db.query(EvidenceDB).filter(EvidenceDB.id == evidence_id).first()
    if not evidence:
        raise HTTPException(status_code=404, detail=f"Evidence {evidence_id} not found")
@@ -24,7 +24,6 @@ from fastapi.responses import FileResponse
 from sqlalchemy.orm import Session
 from classroom_engine.database import get_db
 from ..db.models import EvidenceDB
 from .audit_trail_utils import log_audit_trail
 from ..db import (
@@ -311,7 +310,6 @@ async def list_controls_paginated(
 )
 async def get_control(
    control_id: str,
    db: Session = Depends(get_db),
    svc: ControlExportService = Depends(get_ctrl_export_service),
 ) -> ControlResponse:
    """Get a specific control by control_id."""
@@ -356,7 +354,6 @@ async def get_control(
 async def update_control(
    control_id: str,
    update: ControlUpdate,
    db: Session = Depends(get_db),
    svc: ControlExportService = Depends(get_ctrl_export_service),
 ) -> ControlResponse:
    """Update a control."""
@@ -446,7 +443,6 @@ async def update_control(
 async def review_control(
    control_id: str,
    review: ControlReviewRequest,
    db: Session = Depends(get_db),
    svc: ControlExportService = Depends(get_ctrl_export_service),
 ) -> ControlResponse:
    """Mark a control as reviewed with new status."""
@@ -21,7 +21,7 @@ Phase 1 Step 4 refactor: handlers delegate to VVTService.
 import logging
 from typing import Any, List, Optional
-from fastapi import APIRouter, Depends, HTTPException, Query, Request
+from fastapi import APIRouter, Depends, Query, Request
 from fastapi.responses import StreamingResponse
 from sqlalchemy.orm import Session