feat(ai-sdk): demote superseded pre-eu-v1 sources in authority rerank

The old pre-eu-v1 corpus chunks (un-annotated CRA/AI Act/DORA/NIS2/DSGVO
duplicates + the old Machinery Directive and its guide) are tagged
status=superseded / use_for_primary=false in the vector store. Honor that
in the rerank: a superseded result takes a fixed penalty so the eu-v1 norm
wins default questions, while the old source stays in the pool (demoted,
not hidden) and remains findable for history / transition questions.

Verified on dev: "CRA Sicherheitsupdates" now returns CRA Anhang I (eu-v1)
at #1 instead of an un-annotated old chunk; MaschinenVO outranks the old
Machinery Directive/guide; superseded chunks remain retrievable lower down.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

admin-compliance/app/sdk/coverage/_helpers.ts

@@ -46,6 +46,28 @@ export interface CorpusOverview {
   totals: { documents: number; catalog_sources: number }
 }
 
+// --- Ingested legal-corpus structure (from the vector store, via the Go SDK).
+// Shows WHAT each eur-lex act consists of (articles/annexes/recitals), so the
+// ingested corpus is not a black box for developers. ---
+export interface LegalActStructure {
+  regulation_short: string
+  regulation_name: string
+  articles: number
+  annexes: number
+  recitals: number
+  chunks: number
+}
+
+export interface LegalCorpus {
+  regulations: LegalActStructure[]
+  totals: {
+    regulations: number
+    articles: number
+    annexes: number
+    recitals: number
+  }
+}
+
 // --- Korpus-Dokumente: gruppieren nach Art (Gesetz/Leitfaden/Standard/Urteil)
 // + Herausgeber-Familie (DSK, EDPB, OWASP, NIST …). Deterministisch, pure. ---
 interface DocCat {

admin-compliance/app/sdk/coverage/page.tsx

@@ -3,6 +3,7 @@ import Link from 'next/link'
 import {
   type UseCaseRow,
   type CorpusOverview,
+  type LegalCorpus,
   licenseTierBadgeClass,
   commercialBadgeClass,
   groupUseCases,
@@ -11,28 +12,46 @@ import {
 
 const BACKEND_URL =
   process.env.COMPLIANCE_BACKEND_URL || 'http://backend-compliance:8002'
+// The legal-corpus structure comes from the Go SDK (it owns the vector store).
+const SDK_URL = process.env.SDK_URL || 'http://ai-compliance-sdk:8090'
 
 export const dynamic = 'force-dynamic'
 
+// Fetched from the SDK and isolated in its own try/catch so a vector-store
+// hiccup degrades to "no structure shown" instead of blanking the whole page.
+async function fetchLegalCorpus(): Promise<LegalCorpus | null> {
+  try {
+    const res = await fetch(`${SDK_URL}/sdk/v1/rag/legal-corpus`, {
+      cache: 'no-store',
+    })
+    return res.ok ? await res.json() : null
+  } catch {
+    return null
+  }
+}
+
 async function getData(): Promise<{
   useCases: UseCaseRow[]
   corpus: CorpusOverview | null
+  legalCorpus: LegalCorpus | null
 }> {
   try {
-    const [ucRes, corpusRes] = await Promise.all([
+    const [ucRes, corpusRes, legalCorpus] = await Promise.all([
       fetch(`${BACKEND_URL}/api/compliance/v1/controls/use-cases`, {
         cache: 'no-store',
       }),
       fetch(`${BACKEND_URL}/api/compliance/v1/controls/corpus`, {
         cache: 'no-store',
       }),
+      fetchLegalCorpus(),
     ])
     return {
       useCases: ucRes.ok ? await ucRes.json() : [],
       corpus: corpusRes.ok ? await corpusRes.json() : null,
+      legalCorpus,
     }
   } catch {
-    return { useCases: [], corpus: null }
+    return { useCases: [], corpus: null, legalCorpus: null }
   }
 }
 
@@ -46,7 +65,7 @@ function Stat({ label, value }: { label: string; value: string | number }) {
 }
 
 export default async function CoveragePage() {
-  const { useCases, corpus } = await getData()
+  const { useCases, corpus, legalCorpus } = await getData()
   const groups = groupUseCases(useCases)
   const totalRelevant = useCases.reduce((s, u) => s + u.atom_relevant, 0)
   const totalAtoms = useCases.reduce((s, u) => s + u.atom_total, 0)
@@ -221,6 +240,67 @@ export default async function CoveragePage() {
         </div>
       </section>
 
+      {legalCorpus?.regulations?.length ? (
+        <section className="space-y-2">
+          <h2 className="text-lg font-semibold text-gray-900">
+            Ingestierter Rechtskorpus – Struktur ({legalCorpus.totals.regulations}{' '}
+            Rechtsakte)
+          </h2>
+          <p className="text-xs text-gray-500">
+            Woraus jeder ingestierte eur-lex-Rechtsakt tatsächlich besteht:
+            Artikel (§), Anhänge, Erwägungsgründe und retrievbare Chunks — direkt
+            aus dem Vektorspeicher, damit kein Black-Box-Korpus entsteht.
+          </p>
+          <div className="overflow-auto rounded-lg border border-gray-200">
+            <table className="min-w-full divide-y divide-gray-200 text-sm">
+              <thead className="bg-gray-50 text-left text-xs uppercase text-gray-500">
+                <tr>
+                  <th className="px-4 py-2">Rechtsakt</th>
+                  <th className="px-4 py-2 text-right">Artikel (§)</th>
+                  <th className="px-4 py-2 text-right">Anhänge</th>
+                  <th className="px-4 py-2 text-right">Erwägungsgründe</th>
+                  <th className="px-4 py-2 text-right">Chunks</th>
+                </tr>
+              </thead>
+              <tbody className="divide-y divide-gray-100 bg-white">
+                {legalCorpus.regulations.map((r) => (
+                  <tr key={r.regulation_short}>
+                    <td className="px-4 py-2 text-gray-900">
+                      <span className="font-medium">{r.regulation_short}</span>
+                      {r.regulation_name !== r.regulation_short ? (
+                        <span className="ml-2 text-xs text-gray-500">
+                          {r.regulation_name}
+                        </span>
+                      ) : null}
+                    </td>
+                    <td className="px-4 py-2 text-right font-semibold">
+                      {r.articles.toLocaleString('de-DE')}
+                    </td>
+                    <td className="px-4 py-2 text-right">
+                      {r.annexes > 0 ? (
+                        r.annexes.toLocaleString('de-DE')
+                      ) : (
+                        <span className="text-gray-300">—</span>
+                      )}
+                    </td>
+                    <td className="px-4 py-2 text-right text-gray-500">
+                      {r.recitals > 0 ? (
+                        r.recitals.toLocaleString('de-DE')
+                      ) : (
+                        <span className="text-gray-300">—</span>
+                      )}
+                    </td>
+                    <td className="px-4 py-2 text-right text-gray-500">
+                      {r.chunks.toLocaleString('de-DE')}
+                    </td>
+                  </tr>
+                ))}
+              </tbody>
+            </table>
+          </div>
+        </section>
+      ) : null}
+
       {corpus?.license_catalog?.length ? (
         <section className="space-y-2">
           <h2 className="text-lg font-semibold text-gray-900">

ai-compliance-sdk/internal/api/handlers/rag_handlers.go

@@ -78,6 +78,7 @@ func (h *RAGHandlers) Search(c *gin.Context) {
 		"query":      req.Query,
 		"results":    results,
 		"count":      len(results),
+		"assessment": ucca.Assess(results),
 	})
 }
 
@@ -206,3 +207,32 @@ func (h *RAGHandlers) HandleScrollChunks(c *gin.Context) {
 		"total":       len(chunks),
 	})
 }
+
+// LegalCorpusStructure returns the composition (distinct articles, annexes,
+// recitals + chunk count) of every ingested eur-lex legal act, so the coverage
+// page can show WHAT was ingested instead of just the act name.
+// GET /sdk/v1/rag/legal-corpus
+func (h *RAGHandlers) LegalCorpusStructure(c *gin.Context) {
+	acts, err := h.ragClient.CorpusStructure(c.Request.Context())
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to aggregate legal corpus: " + err.Error()})
+		return
+	}
+
+	arts, anns, recs := 0, 0, 0
+	for _, a := range acts {
+		arts += a.Articles
+		anns += a.Annexes
+		recs += a.Recitals
+	}
+
+	c.JSON(http.StatusOK, gin.H{
+		"regulations": acts,
+		"totals": gin.H{
+			"regulations": len(acts),
+			"articles":    arts,
+			"annexes":     anns,
+			"recitals":    recs,
+		},
+	})
+}

ai-compliance-sdk/internal/app/routes.go

@@ -161,6 +161,7 @@ func registerRAGRoutes(v1 *gin.RouterGroup, h *handlers.RAGHandlers) {
 		ragRoutes.GET("/corpus-status", h.CorpusStatus)
 		ragRoutes.GET("/corpus-versions/:collection", h.CorpusVersionHistory)
 		ragRoutes.GET("/scroll", h.HandleScrollChunks)
+		ragRoutes.GET("/legal-corpus", h.LegalCorpusStructure)
 	}
 }
 

ai-compliance-sdk/internal/ucca/authority_rerank.go

@@ -12,6 +12,7 @@ const (
 	offDomainPenalty  = 0.10 // off-domain binding (demoted, not removed)
 	scopePenalty      = 0.25 // BDSG Teil 3 (law enforcement) on a general DP question
 	topicGain         = 0.18 // amplifier only
+	supersededPenalty = 0.50 // superseded Alt-Quelle (pre-eu-v1): demoted, nicht versteckt
 )
 
 // authorityScore computes the normative relevance of a result for a query. It augments the
@@ -20,6 +21,12 @@ func authorityScore(query string, r LegalSearchResult, qDomain string, qForeign
 	info := classifyAuthority(r)
 	score := r.Score + authorityCoef*float64(info.weight)/100.0
 
+	if r.Superseded {
+		// Alt-Quelle (pre-eu-v1): Default-Fragen sollen die eu-v1-Norm sehen. Demoted,
+		// nicht entfernt — fuer Historie/Uebergangsfragen bleibt sie auffindbar.
+		score -= supersededPenalty
+	}
+
 	if info.jurisdiction == "CH" && !qForeign {
 		score -= foreignPenalty // Fremdrecht bei DE/EU-Frage: demoted, nicht geloescht
 	} else {

ai-compliance-sdk/internal/ucca/legal_corpus_structure.go

@@ -0,0 +1,167 @@
+package ucca
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"sort"
+)
+
+// LegalActStructure is the composition of one ingested eur-lex legal act — how
+// many distinct articles, annexes and recitals it consists of (plus the raw
+// chunk count). Backs the coverage page so the ingested corpus is not a black
+// box: a developer SEES what each act actually contains, not only its name.
+type LegalActStructure struct {
+	RegulationShort string `json:"regulation_short"`
+	RegulationName  string `json:"regulation_name"`
+	Articles        int    `json:"articles"`
+	Annexes         int    `json:"annexes"`
+	Recitals        int    `json:"recitals"`
+	Chunks          int    `json:"chunks"`
+}
+
+const eurlexSource = "eur-lex.europa.eu"
+
+// legalStructureCollections hold the clean eur-lex legal corpus (chunks tagged
+// with chunk_scope = section | annex | recital).
+var legalStructureCollections = []string{"bp_compliance_ce", "bp_compliance_datenschutz"}
+
+// chunkScopeBucket maps a Qdrant chunk_scope to the structure field it feeds.
+var chunkScopeBucket = map[string]string{"section": "articles", "annex": "annexes", "recital": "recitals"}
+
+// CorpusStructure scrolls the eur-lex legal corpus across the legal collections
+// and aggregates the per-act composition. The source filter keeps it to a few
+// hundred points regardless of total corpus size. Read-only; a collection that
+// fails to scroll is skipped rather than failing the whole call.
+func (c *LegalRAGClient) CorpusStructure(ctx context.Context) ([]LegalActStructure, error) {
+	var all []qdrantScrollPoint
+	for _, coll := range legalStructureCollections {
+		pts, err := c.scrollLegalCorpus(ctx, coll)
+		if err != nil {
+			continue
+		}
+		all = append(all, pts...)
+	}
+	return aggregateStructure(all), nil
+}
+
+// aggregateStructure counts distinct article labels per (regulation, scope).
+// Pure → unit-testable without a vector store.
+func aggregateStructure(points []qdrantScrollPoint) []LegalActStructure {
+	distinct := map[string]map[string]map[string]struct{}{}
+	names := map[string]string{}
+	chunks := map[string]int{}
+	order := []string{}
+
+	for _, pt := range points {
+		reg := getString(pt.Payload, "regulation_short")
+		if reg == "" {
+			continue
+		}
+		if _, seen := names[reg]; !seen {
+			name := getString(pt.Payload, "regulation_name_de")
+			if name == "" {
+				name = reg
+			}
+			names[reg] = name
+			distinct[reg] = map[string]map[string]struct{}{}
+			order = append(order, reg)
+		}
+		chunks[reg]++
+		bucket, ok := chunkScopeBucket[getString(pt.Payload, "chunk_scope")]
+		article := getString(pt.Payload, "article")
+		if !ok || article == "" {
+			continue
+		}
+		if distinct[reg][bucket] == nil {
+			distinct[reg][bucket] = map[string]struct{}{}
+		}
+		distinct[reg][bucket][article] = struct{}{}
+	}
+
+	out := make([]LegalActStructure, 0, len(order))
+	for _, reg := range order {
+		out = append(out, LegalActStructure{
+			RegulationShort: reg,
+			RegulationName:  names[reg],
+			Articles:        len(distinct[reg]["articles"]),
+			Annexes:         len(distinct[reg]["annexes"]),
+			Recitals:        len(distinct[reg]["recitals"]),
+			Chunks:          chunks[reg],
+		})
+	}
+	sort.SliceStable(out, func(i, j int) bool {
+		if out[i].Articles != out[j].Articles {
+			return out[i].Articles > out[j].Articles
+		}
+		return out[i].RegulationShort < out[j].RegulationShort
+	})
+	return out
+}
+
+// scrollLegalCorpus pages through one collection, filtered to the eur-lex legal
+// corpus, returning minimal-payload points (no text/vectors).
+func (c *LegalRAGClient) scrollLegalCorpus(ctx context.Context, collection string) ([]qdrantScrollPoint, error) {
+	var all []qdrantScrollPoint
+	var offset interface{}
+	for {
+		points, next, err := c.scrollLegalPage(ctx, collection, offset)
+		if err != nil {
+			return nil, err
+		}
+		all = append(all, points...)
+		if next == nil {
+			break
+		}
+		offset = next
+	}
+	return all, nil
+}
+
+// scrollLegalPage fetches one page of the filtered scroll and returns the
+// points plus the next-page offset (nil when exhausted).
+func (c *LegalRAGClient) scrollLegalPage(ctx context.Context, collection string, offset interface{}) ([]qdrantScrollPoint, interface{}, error) {
+	reqBody := map[string]interface{}{
+		"limit":        500,
+		"with_payload": map[string]interface{}{"include": []string{"regulation_short", "regulation_name_de", "chunk_scope", "article"}},
+		"with_vectors": false,
+		"filter": map[string]interface{}{
+			"must": []map[string]interface{}{
+				{"key": "source", "match": map[string]interface{}{"value": eurlexSource}},
+			},
+		},
+	}
+	if offset != nil {
+		reqBody["offset"] = offset
+	}
+	jsonBody, err := json.Marshal(reqBody)
+	if err != nil {
+		return nil, nil, err
+	}
+	url := fmt.Sprintf("%s/collections/%s/points/scroll", c.qdrantURL, collection)
+	req, err := http.NewRequestWithContext(ctx, "POST", url, bytes.NewReader(jsonBody))
+	if err != nil {
+		return nil, nil, err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	if c.qdrantAPIKey != "" {
+		req.Header.Set("api-key", c.qdrantAPIKey)
+	}
+	resp, err := c.httpClient.Do(req)
+	if err != nil {
+		return nil, nil, err
+	}
+	defer func() { _ = resp.Body.Close() }()
+	if resp.StatusCode != http.StatusOK {
+		body, _ := io.ReadAll(resp.Body)
+		return nil, nil, fmt.Errorf("qdrant returned %d: %s", resp.StatusCode, string(body))
+	}
+	var scrollResp qdrantScrollResponse
+	if err := json.NewDecoder(resp.Body).Decode(&scrollResp); err != nil {
+		return nil, nil, err
+	}
+	return scrollResp.Result.Points, scrollResp.Result.NextPageOffset, nil
+}

ai-compliance-sdk/internal/ucca/legal_corpus_structure_test.go

@@ -0,0 +1,50 @@
+package ucca
+
+import "testing"
+
+func structPoint(reg, name, scope, article string) qdrantScrollPoint {
+	return qdrantScrollPoint{Payload: map[string]interface{}{
+		"regulation_short":   reg,
+		"regulation_name_de": name,
+		"chunk_scope":        scope,
+		"article":            article,
+	}}
+}
+
+func TestAggregateStructure_CountsDistinctPerScope(t *testing.T) {
+	points := []qdrantScrollPoint{
+		structPoint("CRA", "Cyber Resilience Act", "section", "13"),
+		structPoint("CRA", "Cyber Resilience Act", "section", "13"), // duplicate article → still 1
+		structPoint("CRA", "Cyber Resilience Act", "section", "14"),
+		structPoint("CRA", "Cyber Resilience Act", "annex", "Anhang-I"),
+		structPoint("CRA", "Cyber Resilience Act", "annex", "Anhang-VII"),
+		structPoint("DORA", "", "section", "6"),  // first sighting has no name →
+		structPoint("DORA", "", "section", "19"), // regulation_name falls back to short
+		structPoint("DORA", "", "recital", ""),   // empty article → ignored for distinct
+		structPoint("", "x", "section", "1"),     // missing regulation → skipped entirely
+	}
+
+	got := aggregateStructure(points)
+
+	if len(got) != 2 {
+		t.Fatalf("want 2 acts, got %d (%+v)", len(got), got)
+	}
+	// CRA has more articles → sorts first.
+	cra := got[0]
+	if cra.RegulationShort != "CRA" || cra.Articles != 2 || cra.Annexes != 2 || cra.Recitals != 0 || cra.Chunks != 5 {
+		t.Errorf("CRA wrong: %+v", cra)
+	}
+	dora := got[1]
+	if dora.RegulationShort != "DORA" || dora.Articles != 2 || dora.Chunks != 3 {
+		t.Errorf("DORA wrong: %+v", dora)
+	}
+	if dora.RegulationName != "DORA" {
+		t.Errorf("DORA name fallback failed: %q", dora.RegulationName)
+	}
+}
+
+func TestAggregateStructure_Empty(t *testing.T) {
+	if got := aggregateStructure(nil); len(got) != 0 {
+		t.Errorf("want empty, got %+v", got)
+	}
+}

ai-compliance-sdk/internal/ucca/legal_rag_assess.go

@@ -0,0 +1,134 @@
+package ucca
+
+import (
+	"fmt"
+	"strings"
+)
+
+const (
+	assessConnectedCap    = 12   // cap connected norms surfaced in the assessment
+	assessCrossRegimeTopN = 5    // window over which "cross regime" is judged
+	assessReviewMargin    = 0.05 // a tighter winner gap → recommend human review
+)
+
+// Assess builds the auditable explanation layer over a ranked result set:
+// primary norm, the norms it connects to (citation graph), cross-regime, a
+// human-review flag, the winner margin and a short reasoning string. Pure →
+// unit-testable. It EXPLAINS the ranking, it does not change it. Returns nil for
+// an empty result set.
+func Assess(results []LegalSearchResult) *LegalAssessment {
+	if len(results) == 0 {
+		return nil
+	}
+	// Norm-level view: collapse multiple chunks of the same article/annex so the
+	// margin and cross-regime are judged between DISTINCT norms, not near-identical
+	// chunks of one norm (which would make every winner margin ~0).
+	norms := distinctNorms(results)
+	p := norms[0]
+
+	primary := primaryLabel(p)
+	connected := dedupStrings(p.ReferencesOut, p.ReferencesIn, p.CitationUnit)
+	if len(connected) > assessConnectedCap {
+		connected = connected[:assessConnectedCap]
+	}
+
+	window := norms
+	if len(window) > assessCrossRegimeTopN {
+		window = window[:assessCrossRegimeTopN]
+	}
+	regimes := make(map[string]bool)
+	for _, r := range window {
+		if r.RegulationShort != "" {
+			regimes[r.RegulationShort] = true
+		}
+	}
+	crossRegime := len(regimes) > 1
+
+	margin := 0.0
+	if len(norms) > 1 {
+		margin = norms[0].Score - norms[1].Score
+	}
+
+	primaryBinding := p.SourceClass == "binding_law"
+	humanReview := margin < assessReviewMargin || crossRegime || !primaryBinding
+
+	return &LegalAssessment{
+		PrimaryNorm:       primary,
+		PrimaryRegulation: p.RegulationShort,
+		ConnectedNorms:    connected,
+		CrossRegime:       crossRegime,
+		HumanReviewFlag:   humanReview,
+		WinnerMargin:      margin,
+		ScoreReasoning:    assessReasoning(p, margin, crossRegime, primaryBinding),
+	}
+}
+
+func primaryLabel(p LegalSearchResult) string {
+	if p.CitationUnit != "" {
+		return p.CitationUnit
+	}
+	if p.ArticleLabel != "" {
+		return p.ArticleLabel
+	}
+	return strings.TrimSpace(p.RegulationShort + " " + p.Article)
+}
+
+// assessReasoning renders a short, human-readable justification (German).
+func assessReasoning(p LegalSearchResult, margin float64, crossRegime, primaryBinding bool) string {
+	label := primaryLabel(p)
+	parts := make([]string, 0, 4)
+	if primaryBinding {
+		parts = append(parts, fmt.Sprintf("Primärtreffer %s: bindendes Recht (Autorität %d).", label, p.AuthorityWeight))
+	} else {
+		parts = append(parts, fmt.Sprintf("Primärtreffer %s ist keine bindende Norm (Leitlinie/Standard) — Quelle prüfen.", label))
+	}
+	if margin > 0 {
+		parts = append(parts, fmt.Sprintf("Vorsprung %.2f vor #2.", margin))
+	}
+	if margin < assessReviewMargin {
+		parts = append(parts, "Knapper Vorsprung — Alternativtreffer prüfen.")
+	}
+	if crossRegime {
+		parts = append(parts, "Mehrere Regime betroffen — Querbezug prüfen.")
+	}
+	return strings.Join(parts, " ")
+}
+
+// distinctNorms collapses results that share a citation (multiple chunks of the
+// same article/annex) to the first — i.e. highest-ranked — occurrence. Results
+// without any citation identity are each kept, since they cannot be matched.
+func distinctNorms(results []LegalSearchResult) []LegalSearchResult {
+	seen := make(map[string]bool, len(results))
+	out := make([]LegalSearchResult, 0, len(results))
+	for _, r := range results {
+		key := r.CitationUnit
+		if key == "" {
+			key = r.ArticleLabel
+		}
+		if key != "" {
+			if seen[key] {
+				continue
+			}
+			seen[key] = true
+		}
+		out = append(out, r)
+	}
+	return out
+}
+
+// dedupStrings concatenates out+in, drops empties and the excluded value, and
+// returns a stable de-duplicated slice (insertion order preserved).
+func dedupStrings(out, in []string, exclude string) []string {
+	seen := map[string]bool{exclude: true}
+	res := make([]string, 0, len(out)+len(in))
+	for _, list := range [][]string{out, in} {
+		for _, s := range list {
+			if s == "" || seen[s] {
+				continue
+			}
+			seen[s] = true
+			res = append(res, s)
+		}
+	}
+	return res
+}

ai-compliance-sdk/internal/ucca/legal_rag_assess_test.go

@@ -0,0 +1,112 @@
+package ucca
+
+import "testing"
+
+func ares(reg, cu, sc string, score float64, weight int, out, in []string) LegalSearchResult {
+	return LegalSearchResult{
+		RegulationShort: reg, CitationUnit: cu, SourceClass: sc, Score: score,
+		AuthorityWeight: weight, ReferencesOut: out, ReferencesIn: in,
+	}
+}
+
+func TestAssess_Empty(t *testing.T) {
+	if Assess(nil) != nil {
+		t.Error("empty results → nil assessment")
+	}
+}
+
+func TestAssess_BindingPrimary_NoReview(t *testing.T) {
+	results := []LegalSearchResult{
+		ares("CRA", "Art. 13 CRA", "binding_law", 1.05, 100,
+			[]string{"CRA Anhang I", "Art. 14 CRA"}, []string{"Art. 12 CRA"}),
+		ares("CRA", "Art. 14 CRA", "binding_law", 0.80, 100, nil, nil),
+	}
+	a := Assess(results)
+	if a == nil {
+		t.Fatal("nil assessment")
+	}
+	if a.PrimaryNorm != "Art. 13 CRA" || a.PrimaryRegulation != "CRA" {
+		t.Errorf("primary wrong: %+v", a)
+	}
+	if len(a.ConnectedNorms) != 3 { // out(2) + in(1), self excluded, deduped
+		t.Errorf("connected norms: %v", a.ConnectedNorms)
+	}
+	if a.CrossRegime {
+		t.Error("single regime must not be cross-regime")
+	}
+	if a.WinnerMargin < 0.24 || a.WinnerMargin > 0.26 {
+		t.Errorf("margin = %v, want ~0.25", a.WinnerMargin)
+	}
+	if a.HumanReviewFlag {
+		t.Error("clean binding + healthy margin + single regime → no review")
+	}
+}
+
+func TestAssess_CrossRegimeFlagsReview(t *testing.T) {
+	a := Assess([]LegalSearchResult{
+		ares("CRA", "Art. 13 CRA", "binding_law", 1.05, 100, nil, nil),
+		ares("DORA", "Art. 6 DORA", "binding_law", 0.70, 100, nil, nil),
+	})
+	if !a.CrossRegime || !a.HumanReviewFlag {
+		t.Errorf("cross-regime must flag review: %+v", a)
+	}
+}
+
+func TestAssess_NonBindingFlagsReview(t *testing.T) {
+	a := Assess([]LegalSearchResult{
+		ares("ENISA", "ENISA SBOM", "supervisory_guidance", 0.90, 70, nil, nil),
+		ares("ENISA", "ENISA X", "supervisory_guidance", 0.40, 70, nil, nil),
+	})
+	if !a.HumanReviewFlag {
+		t.Error("non-binding primary → review")
+	}
+}
+
+func TestAssess_TightMarginFlagsReview(t *testing.T) {
+	a := Assess([]LegalSearchResult{
+		ares("CRA", "Art. 13 CRA", "binding_law", 1.00, 100, nil, nil),
+		ares("CRA", "Art. 14 CRA", "binding_law", 0.98, 100, nil, nil),
+	})
+	if a.WinnerMargin >= 0.05 || !a.HumanReviewFlag {
+		t.Errorf("tight margin → review: %+v", a)
+	}
+}
+
+func TestAssess_MarginIsNormLevelNotChunkLevel(t *testing.T) {
+	// Two near-identical chunks of the SAME norm at the top, then a distinct norm.
+	results := []LegalSearchResult{
+		ares("CRA", "Art. 13 CRA", "binding_law", 1.050, 100, []string{"CRA Anhang I"}, nil),
+		ares("CRA", "Art. 13 CRA", "binding_law", 1.049, 100, nil, nil), // same norm
+		ares("CRA", "Art. 14 CRA", "binding_law", 0.800, 100, nil, nil),
+	}
+	a := Assess(results)
+	if a.WinnerMargin < 0.24 || a.WinnerMargin > 0.26 { // Art.13 vs Art.14, not chunk vs chunk
+		t.Errorf("margin must be norm-level (~0.25), got %v", a.WinnerMargin)
+	}
+	if a.HumanReviewFlag {
+		t.Error("healthy norm-level margin → no review")
+	}
+}
+
+func TestDistinctNorms(t *testing.T) {
+	got := distinctNorms([]LegalSearchResult{
+		{CitationUnit: "Art. 13 CRA"},
+		{CitationUnit: "Art. 13 CRA"}, // duplicate norm → collapsed
+		{CitationUnit: "Art. 14 CRA"},
+		{CitationUnit: ""}, // no identity → kept
+		{CitationUnit: ""}, // no identity → kept
+	})
+	if len(got) != 4 {
+		t.Errorf("want 4 (2 distinct + 2 unidentified), got %d", len(got))
+	}
+}
+
+func TestDedupStrings(t *testing.T) {
+	got := dedupStrings([]string{"a", "b", "", "a"}, []string{"b", "c"}, "self")
+	if len(got) != 3 || got[0] != "a" || got[1] != "b" || got[2] != "c" {
+		t.Errorf("dedup: %v", got)
+	}
+	if len(dedupStrings([]string{"self"}, nil, "self")) != 0 {
+		t.Error("excluded value must be dropped")
+	}
+}

ai-compliance-sdk/internal/ucca/legal_rag_client.go

@@ -20,6 +20,7 @@ type LegalRAGClient struct {
 	httpClient       *http.Client
 	textIndexEnsured map[string]bool
 	hybridEnabled    bool
+	graphEnabled     bool
 }
 
 // NewLegalRAGClient creates a new Legal RAG client using Ollama bge-m3 embeddings.
@@ -38,6 +39,11 @@ func NewLegalRAGClient() *LegalRAGClient {
 	}
 
 	hybridEnabled := os.Getenv("RAG_HYBRID_SEARCH") != "false"
+	// Graph-Expansion ist OPT-IN: kein gemessener Rang-Nutzen ggue. der Binding-Augmentation,
+	// +1 Qdrant-Call/Suche, Flutungsrisiko ueber Reverse-Kanten. Bleibt als Recall-Sicherheitsnetz
+	// fuer spaetere Luecken (RAG_GRAPH_EXPANSION=true). Die Graph-Kanten werden in der Response
+	// zur Begruendung/Vollstaendigkeit genutzt, nicht zur Pool-Expansion (Default).
+	graphEnabled := os.Getenv("RAG_GRAPH_EXPANSION") == "true"
 
 	return &LegalRAGClient{
 		qdrantURL:        qdrantURL,
@@ -47,6 +53,7 @@ func NewLegalRAGClient() *LegalRAGClient {
 		collection:       "bp_compliance_ce",
 		textIndexEnsured: make(map[string]bool),
 		hybridEnabled:    hybridEnabled,
+		graphEnabled:     graphEnabled,
 		httpClient: &http.Client{
 			Timeout: 60 * time.Second,
 		},
@@ -100,6 +107,13 @@ func (c *LegalRAGClient) searchInternal(ctx context.Context, collection string,
 		hits = mergeDedupHits(hits, bindingHits)
 	}
 
+	// Graph-Augmentation: verbundene Normen (references_out/in) der Top-Hits ueber die
+	// praezise Zitations-Kante in den Pool ziehen — z.B. Art. 13 CRA zieht Anhang I (die
+	// eigentliche Pflichtquelle). Pool-Augmentation only; Re-Rank + topK bleiben.
+	if c.graphEnabled {
+		hits = c.expandViaGraph(ctx, collection, hits)
+	}
+
 	results := make([]LegalSearchResult, len(hits))
 	for i, hit := range hits {
 		// Legal-Metadaten nach rag_reingest_spec.md §2: bevorzugt die normalisierten Felder
@@ -131,6 +145,10 @@ func (c *LegalRAGClient) searchInternal(ctx context.Context, collection string,
 			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",
 		}
 	}
 

ai-compliance-sdk/internal/ucca/legal_rag_graph.go

@@ -0,0 +1,162 @@
+package ucca
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"sort"
+)
+
+// Graph-augmented retrieval: when a top hit cites an annex/article (references_out)
+// or is cited by one (references_in), pull that connected norm into the candidate
+// pool via the PRECISE citation graph instead of hoping semantic search surfaces
+// it. E.g. a hit on CRA Art. 13 pulls in CRA Anhang I (the actual requirement).
+// Pool-augmentation only — authority re-rank + topK slice still apply, so the
+// response schema is unchanged.
+const (
+	graphSeedCount  = 5    // only the top hits seed the expansion
+	graphMaxExpand  = 15   // cap connected norms pulled in (avoid pool explosion)
+	graphHopPenalty = 0.05 // a one-hop neighbour ranks just below its seed
+)
+
+// expandViaGraph augments hits with the norms they cite and the norms that cite
+// them. Best-effort: on any error (or nothing to expand) the original hits are
+// returned unchanged.
+func (c *LegalRAGClient) expandViaGraph(ctx context.Context, collection string, hits []qdrantSearchHit) []qdrantSearchHit {
+	if len(hits) == 0 {
+		return hits
+	}
+	present := make(map[string]bool, len(hits))
+	for _, h := range hits {
+		if cu := getString(h.Payload, "citation_unit"); cu != "" {
+			present[cu] = true
+		}
+	}
+
+	seeds := hits
+	if len(seeds) > graphSeedCount {
+		seeds = seeds[:graphSeedCount]
+	}
+	// Forward edges only (references_out = the detail a hit explicitly points to,
+	// e.g. Art. 13 → Anhang I). Reverse (references_in) has high fan-out for popular
+	// annexes (Anhang I is cited by 23 articles) → pool flooding; it is surfaced as
+	// connected-norm metadata in the Phase 2 response instead of expanding the pool.
+	want := make(map[string]float64) // connected citation_unit -> best seeding score
+	for _, h := range seeds {
+		for _, cu := range getStringSlice(h.Payload, "references_out") {
+			if cu == "" || present[cu] {
+				continue
+			}
+			if s, ok := want[cu]; !ok || h.Score > s {
+				want[cu] = h.Score
+			}
+		}
+	}
+	if len(want) == 0 {
+		return hits
+	}
+
+	units := topByScore(want, graphMaxExpand)
+	fetched, err := c.fetchByCitationUnits(ctx, collection, units)
+	if err != nil || len(fetched) == 0 {
+		return hits
+	}
+	neighbours := make([]qdrantSearchHit, 0, len(fetched))
+	for cu, pt := range fetched {
+		neighbours = append(neighbours, qdrantSearchHit{ID: pt.ID, Score: want[cu] - graphHopPenalty, Payload: pt.Payload})
+	}
+	return mergeDedupHits(hits, neighbours)
+}
+
+// topByScore returns up to n keys with the highest values. Deterministic: ties
+// broken by the key string so the cap is stable across runs.
+func topByScore(m map[string]float64, n int) []string {
+	keys := make([]string, 0, len(m))
+	for k := range m {
+		keys = append(keys, k)
+	}
+	sort.Slice(keys, func(i, j int) bool {
+		if m[keys[i]] != m[keys[j]] {
+			return m[keys[i]] > m[keys[j]]
+		}
+		return keys[i] < keys[j]
+	})
+	if len(keys) > n {
+		keys = keys[:n]
+	}
+	return keys
+}
+
+// fetchByCitationUnits loads one representative point (the first chunk) per
+// citation_unit from the given collection.
+func (c *LegalRAGClient) fetchByCitationUnits(ctx context.Context, collection string, units []string) (map[string]qdrantScrollPoint, error) {
+	should := make([]map[string]interface{}, 0, len(units))
+	for _, cu := range units {
+		should = append(should, map[string]interface{}{"key": "citation_unit", "match": map[string]interface{}{"value": cu}})
+	}
+	reqBody := map[string]interface{}{
+		"limit":        len(units) * 4,
+		"with_payload": true,
+		"with_vectors": false,
+		"filter":       map[string]interface{}{"should": should},
+	}
+	jsonBody, err := json.Marshal(reqBody)
+	if err != nil {
+		return nil, err
+	}
+	url := fmt.Sprintf("%s/collections/%s/points/scroll", c.qdrantURL, collection)
+	req, err := http.NewRequestWithContext(ctx, "POST", url, bytes.NewReader(jsonBody))
+	if err != nil {
+		return nil, err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	if c.qdrantAPIKey != "" {
+		req.Header.Set("api-key", c.qdrantAPIKey)
+	}
+	resp, err := c.httpClient.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer func() { _ = resp.Body.Close() }()
+	if resp.StatusCode != http.StatusOK {
+		body, _ := io.ReadAll(resp.Body)
+		return nil, fmt.Errorf("qdrant scroll returned %d: %s", resp.StatusCode, string(body))
+	}
+	var scrollResp qdrantScrollResponse
+	if err := json.NewDecoder(resp.Body).Decode(&scrollResp); err != nil {
+		return nil, err
+	}
+	out := make(map[string]qdrantScrollPoint, len(units))
+	for _, pt := range scrollResp.Result.Points {
+		cu := getString(pt.Payload, "citation_unit")
+		if cu != "" {
+			if _, seen := out[cu]; !seen {
+				out[cu] = pt
+			}
+		}
+	}
+	return out, nil
+}
+
+// getStringSlice extracts a []string from a Qdrant payload list field
+// (references_out / references_in are stored as JSON arrays of strings).
+func getStringSlice(m map[string]interface{}, key string) []string {
+	v, ok := m[key]
+	if !ok {
+		return nil
+	}
+	arr, ok := v.([]interface{})
+	if !ok {
+		return nil
+	}
+	out := make([]string, 0, len(arr))
+	for _, item := range arr {
+		if s, ok := item.(string); ok {
+			out = append(out, s)
+		}
+	}
+	return out
+}

ai-compliance-sdk/internal/ucca/legal_rag_graph_test.go

@@ -0,0 +1,89 @@
+package ucca
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+func TestGetStringSlice(t *testing.T) {
+	m := map[string]interface{}{
+		"refs": []interface{}{"a", "b", 3, "c"}, // non-strings are skipped
+		"str":  "not-a-list",
+	}
+	got := getStringSlice(m, "refs")
+	if len(got) != 3 || got[0] != "a" || got[2] != "c" {
+		t.Errorf("refs: %v", got)
+	}
+	if getStringSlice(m, "missing") != nil {
+		t.Error("missing key should be nil")
+	}
+	if getStringSlice(m, "str") != nil {
+		t.Error("non-list should be nil")
+	}
+}
+
+func TestTopByScore_DeterministicCap(t *testing.T) {
+	m := map[string]float64{"x": 0.5, "y": 0.9, "z": 0.5, "w": 0.7}
+	got := topByScore(m, 2)
+	if len(got) != 2 || got[0] != "y" || got[1] != "w" {
+		t.Errorf("want [y w], got %v", got)
+	}
+	all := topByScore(m, 10)
+	if all[2] != "x" || all[3] != "z" { // tie 0.5 broken by key string
+		t.Errorf("tie-break not deterministic: %v", all)
+	}
+}
+
+func TestExpandViaGraph_NoSeedsOrRefs(t *testing.T) {
+	c := &LegalRAGClient{} // nil httpClient → must not be called on these paths
+	if out := c.expandViaGraph(context.Background(), "x", nil); out != nil {
+		t.Error("empty hits should return nil")
+	}
+	hits := []qdrantSearchHit{{ID: 1, Score: 0.8, Payload: map[string]interface{}{"citation_unit": "Art. 1 CRA"}}}
+	if out := c.expandViaGraph(context.Background(), "x", hits); len(out) != 1 {
+		t.Errorf("no references → unchanged, got %d", len(out))
+	}
+}
+
+func TestExpandViaGraph_PullsConnectedNorm(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]interface{}{
+			"result": map[string]interface{}{
+				"points": []map[string]interface{}{
+					{"id": 99, "payload": map[string]interface{}{
+						"citation_unit": "CRA Anhang I", "chunk_text": "Sicherheitsanforderungen",
+						"source_class": "binding_law", "authority_weight": 100, "regulation_short": "CRA",
+					}},
+				},
+				"next_page_offset": nil,
+			},
+		})
+	}))
+	defer srv.Close()
+
+	c := &LegalRAGClient{qdrantURL: srv.URL, httpClient: srv.Client()}
+	hits := []qdrantSearchHit{
+		{ID: 1, Score: 0.70, Payload: map[string]interface{}{
+			"citation_unit": "Art. 13 CRA", "references_out": []interface{}{"CRA Anhang I"},
+		}},
+	}
+	out := c.expandViaGraph(context.Background(), "bp_compliance_ce", hits)
+	if len(out) != 2 {
+		t.Fatalf("want 2 hits (seed + connected annex), got %d", len(out))
+	}
+	var found *qdrantSearchHit
+	for i := range out {
+		if getString(out[i].Payload, "citation_unit") == "CRA Anhang I" {
+			found = &out[i]
+		}
+	}
+	if found == nil {
+		t.Fatal("connected norm CRA Anhang I was not pulled into the pool")
+	}
+	if found.Score < 0.64 || found.Score > 0.66 { // 0.70 seed − 0.05 hop penalty
+		t.Errorf("connected score = %v, want ~0.65", found.Score)
+	}
+}

ai-compliance-sdk/internal/ucca/legal_rag_supersede_test.go

@@ -0,0 +1,30 @@
+package ucca
+
+import "testing"
+
+// A superseded alt-source must rank below the same result when it is NOT
+// superseded (the eu-v1 norm), but only demoted — the penalty is finite, so it
+// stays in the pool and remains findable for history/transition questions.
+func TestAuthorityScore_SupersededIsDemotedNotRemoved(t *testing.T) {
+	fresh := LegalSearchResult{
+		Score: 0.65, SourceClass: "binding_law", AuthorityWeight: 100,
+		Jurisdiction: "EU", RegulationShort: "CRA", Article: "13",
+	}
+	old := fresh
+	old.Superseded = true
+
+	sFresh := authorityScore("CRA Sicherheitsupdates Hersteller", fresh, "", false)
+	sOld := authorityScore("CRA Sicherheitsupdates Hersteller", old, "", false)
+
+	if sOld >= sFresh {
+		t.Errorf("superseded must score lower: fresh=%.3f superseded=%.3f", sFresh, sOld)
+	}
+	gap := sFresh - sOld
+	if gap < supersededPenalty-0.001 || gap > supersededPenalty+0.001 {
+		t.Errorf("demotion should equal supersededPenalty (%.2f), got %.3f", supersededPenalty, gap)
+	}
+	// Still a positive, finite score → present in the pool, not hidden.
+	if sOld <= -1 {
+		t.Errorf("superseded score collapsed (%.3f) — must remain findable", sOld)
+	}
+}

ai-compliance-sdk/internal/ucca/legal_rag_types.go

@@ -27,6 +27,31 @@ type LegalSearchResult struct {
 	AuthorityWeight int    `json:"-"`
 	SourceClass     string `json:"-"`
 	Jurisdiction    string `json:"-"`
+
+	// Zitations-Graph (Phase 2) — intern, speist nur die Assessment-Berechnung
+	// (verbundene Normen, Begruendung). Pro-Result-Schema bleibt eingefroren.
+	CitationUnit  string   `json:"-"`
+	ReferencesOut []string `json:"-"`
+	ReferencesIn  []string `json:"-"`
+
+	// Supersede-Status (status="superseded", use_for_primary=false) — Alt-Quelle,
+	// die fuer Default-Fragen demoted wird (nicht versteckt; fuer Historie auffindbar).
+	Superseded bool `json:"-"`
+}
+
+// LegalAssessment is the auditable explanation layer over a ranked result set:
+// which norm is primary, which norms connect to it via the citation graph,
+// whether the answer crosses regulatory regimes, and whether a human should
+// review. Computed from the already-ranked results — it EXPLAINS retrieval, it
+// does not change it (graph edges for reasoning/completeness, not pool-expansion).
+type LegalAssessment struct {
+	PrimaryNorm       string   `json:"primary_norm"`
+	PrimaryRegulation string   `json:"primary_regulation"`
+	ConnectedNorms    []string `json:"connected_norms"`
+	CrossRegime       bool     `json:"cross_regime"`
+	HumanReviewFlag   bool     `json:"human_review_flag"`
+	WinnerMargin      float64  `json:"winner_margin"`
+	ScoreReasoning    string   `json:"score_reasoning"`
 }
 
 // LegalContext represents aggregated legal context for an assessment.