fix(ai-sdk): make interpretation-intent override reliably win

PR #34's fixed +0.25 guidance gain was too small live: for "Was empfiehlt der
EDPB zum DSB?" the binding Art. 37 (1.381) still edged out the boosted EDPB
guidance (1.348), because the live authority score gives the binding article a
topic/domain bonus the (partly English) guidance chunk does not match.

Replace the fixed gain with a deterministic lift: a semantically competitive
guideline (raw semantic >= best_binding_semantic - 0.05) is lifted just ABOVE
the best binding FINAL score (ordered by semantic), so authority/topic/domain
bonuses can no longer edge it out. Obligation questions (no intent signal) are
untouched — binding stays Top-1; off-topic guidance stays demoted.

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

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

@@ -7,17 +7,17 @@ import (
 
 // Re-ranking coefficients (validated in the offline golden harness; Phase A — conservative).
 const (
-	authorityCoef     = 0.40 // * weight/100
-	jurisdictionGain  = 0.05 // binding/guidance from DE or EU
-	foreignPenalty    = 0.60 // foreign law on a DE/EU question (demoted, not removed)
-	unknownPenalty    = 0.08
-	domainMatchGain   = 0.15
-	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-	supersededPenalty = 0.50 // superseded Alt-Quelle (pre-eu-v1): demoted, nicht versteckt
-	intentLiftGain    = 0.10 // epsilon a qualifying interpretative source is lifted ABOVE the best binding
-	intentLiftMargin  = 0.05 // ...only if that source is semantically competitive with binding
+	authorityCoef        = 0.40 // * weight/100
+	jurisdictionGain     = 0.05 // binding/guidance from DE or EU
+	foreignPenalty       = 0.60 // foreign law on a DE/EU question (demoted, not removed)
+	unknownPenalty       = 0.08
+	domainMatchGain      = 0.15
+	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+	supersededPenalty    = 0.50 // superseded Alt-Quelle (pre-eu-v1): demoted, nicht versteckt
+	guidanceIntentGain   = 0.10 // epsilon a qualifying guideline is lifted ABOVE the best binding hit
+	guidanceIntentMargin = 0.05 // ...only if the guideline is semantically competitive with binding
 )
 
 // guidanceIntentSignals mark a query that EXPLICITLY asks for an interpretation /
@@ -29,19 +29,10 @@ var guidanceIntentSignals = []string{
 	"auslegung", "empfiehlt", "empfehlung", "sagt", "laut",
 }
 
-// controlIntentSignals mark a query that asks HOW to implement / which controls or
-// measures fit — rather than WHAT the binding obligation is. Only then may a
-// (semantically competitive) technical_standard outrank the binding norm.
-var controlIntentSignals = []string{
-	"control", "controls", "maßnahme", "massnahme", "schutzmaßnahme",
-	"best practice", "best-practice", "umsetzen", "implementier", "absicher",
-	"härt", "haert", "hardening", "nist", "owasp", "grundschutz",
-	"ccm", "iso 27001", "isms",
-}
-
-func queryMatchesAny(query string, signals []string) bool {
+// queryWantsGuidance reports whether the query explicitly asks for guidance/interpretation.
+func queryWantsGuidance(query string) bool {
 	q := strings.ToLower(query)
-	for _, sig := range signals {
+	for _, sig := range guidanceIntentSignals {
 		if strings.Contains(q, sig) {
 			return true
 		}
@@ -49,17 +40,11 @@ func queryMatchesAny(query string, signals []string) bool {
 	return false
 }
 
-// queryWantsGuidance reports whether the query explicitly asks for guidance/interpretation.
-func queryWantsGuidance(query string) bool { return queryMatchesAny(query, guidanceIntentSignals) }
-
-// queryWantsControls reports whether the query asks for implementation controls/measures.
-func queryWantsControls(query string) bool { return queryMatchesAny(query, controlIntentSignals) }
-
 // bestBindingSemantic returns the highest RAW semantic score among binding-law
-// results (0 if none / no intent). Used as the guard threshold so an off-topic
-// interpretative source cannot ride the intent boost.
-func bestBindingSemantic(results []LegalSearchResult, wantsIntent bool) float64 {
-	if !wantsIntent {
+// results (0 if none / intent not requested). Used as the guard threshold so an
+// off-topic guideline cannot ride the interpretation-intent boost.
+func bestBindingSemantic(results []LegalSearchResult, wantsGuidance bool) float64 {
+	if !wantsGuidance {
 		return 0
 	}
 	best := 0.0
@@ -119,22 +104,15 @@ func rerankByAuthority(query string, results []LegalSearchResult) []LegalSearchR
 	qDomain := queryDomain(query)
 	qForeign := queryIsForeign(query)
 	wantsGuidance := queryWantsGuidance(query)
-	wantsControls := queryWantsControls(query)
-	bestBindingSem := bestBindingSemantic(results, wantsGuidance || wantsControls)
+	bestBindingSem := bestBindingSemantic(results, wantsGuidance)
 
 	out := make([]LegalSearchResult, len(results))
 	copy(out, results)
 	for i := range out {
 		out[i].Score = authorityScore(query, out[i], qDomain, qForeign)
 	}
-	// Explicit interpretation intent → a competitive guideline may outrank binding;
-	// explicit implementation intent → a competitive technical_standard may. Both lift
-	// ABOVE the best binding FINAL, so a pure norm question (neither intent) is untouched.
 	if wantsGuidance {
-		liftAboveBinding(out, results, bestBindingSem, "supervisory_guidance")
-	}
-	if wantsControls {
-		liftAboveBinding(out, results, bestBindingSem, "technical_standard")
+		applyGuidanceIntent(out, results, bestBindingSem)
 	}
 	sort.SliceStable(out, func(a, b int) bool {
 		return out[a].Score > out[b].Score
@@ -142,14 +120,13 @@ func rerankByAuthority(query string, results []LegalSearchResult) []LegalSearchR
 	return out
 }
 
-// liftAboveBinding lifts a semantically-competitive interpretative source (the given
-// sourceClass — supervisory_guidance or technical_standard) just ABOVE the best binding
-// hit, ordered by semantic, so an EXPLICIT guidance/implementation question can return
-// that source Top-1. A pure norm question (no intent → not called) keeps binding on top.
-// Sources below the semantic margin are left untouched, so an off-topic source can never
-// ride the override — and the lift is from the binding FINAL score, so authority/topic/
-// domain bonuses cannot edge it out.
-func liftAboveBinding(out, raw []LegalSearchResult, bestBindingSem float64, sourceClass string) {
+// applyGuidanceIntent lifts semantically-competitive guidance just ABOVE the best
+// binding hit (ordered by semantic), so an EXPLICIT interpretation question can
+// return guidance Top-1. Obligation questions (no intent → not called) keep
+// binding on top. Guidance below the semantic margin is left untouched, so an
+// off-topic guideline can never ride the override — and the lift is computed from
+// the binding FINAL score, so authority/topic/domain bonuses cannot edge it out.
+func applyGuidanceIntent(out, raw []LegalSearchResult, bestBindingSem float64) {
 	bestBindingFinal := 0.0
 	for i := range out {
 		if out[i].SourceClass == "binding_law" && out[i].Score > bestBindingFinal {
@@ -157,10 +134,10 @@ func liftAboveBinding(out, raw []LegalSearchResult, bestBindingSem float64, sour
 		}
 	}
 	for i := range out {
-		if out[i].SourceClass != sourceClass || raw[i].Score < bestBindingSem-intentLiftMargin {
+		if out[i].SourceClass != "supervisory_guidance" || raw[i].Score < bestBindingSem-guidanceIntentMargin {
 			continue
 		}
-		lifted := bestBindingFinal + intentLiftGain + (raw[i].Score - bestBindingSem)
+		lifted := bestBindingFinal + guidanceIntentGain + (raw[i].Score - bestBindingSem)
 		if lifted > out[i].Score {
 			out[i].Score = lifted
 		}

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

@@ -70,65 +70,3 @@ func TestRerank_OffTopicGuidance_BlockedByGuard(t *testing.T) {
 		t.Errorf("off-topic guidance must not win even with intent, got %s", out[0].SourceClass)
 	}
 }
-
-func TestQueryWantsControls(t *testing.T) {
-	wants := []string{
-		"Welche Controls passen zu Security Updates?",
-		"Welche Maßnahmen sollten wir umsetzen?",
-		"Wie härten wir den Server ab?",
-		"Gibt es NIST-Controls dafür?",
-		"OWASP Best Practice für Logging?",
-		"BSI Grundschutz Bausteine",
-	}
-	plain := []string{
-		"Welche Anforderungen bestehen an Security Updates?",
-		"Ab wann braucht man einen Datenschutzbeauftragten?",
-	}
-	for _, q := range wants {
-		if !queryWantsControls(q) {
-			t.Errorf("should detect control/implementation intent: %q", q)
-		}
-	}
-	for _, q := range plain {
-		if queryWantsControls(q) {
-			t.Errorf("should NOT detect control intent (norm question): %q", q)
-		}
-	}
-}
-
-func TestRerank_ControlQuestion_StandardMayWin(t *testing.T) {
-	// Explicit implementation intent + standard semantically competitive → standard wins.
-	results := []LegalSearchResult{
-		intentRes("NIST SP 800-82", "technical_standard", 0.62, 80),
-		intentRes("CRA", "binding_law", 0.58, 100),
-	}
-	out := rerankByAuthority("Welche Controls passen zu Security Updates?", results)
-	if out[0].SourceClass != "technical_standard" {
-		t.Errorf("control question: technical_standard should win Top-1, got %s", out[0].SourceClass)
-	}
-}
-
-func TestRerank_NormQuestion_BindingOverStandard(t *testing.T) {
-	// "Anforderungen" → no control intent → binding stays Top-1 over the standard.
-	results := []LegalSearchResult{
-		intentRes("NIST SP 800-82", "technical_standard", 0.62, 80),
-		intentRes("CRA", "binding_law", 0.58, 100),
-	}
-	out := rerankByAuthority("Welche Anforderungen bestehen an Security Updates?", results)
-	if out[0].SourceClass != "binding_law" {
-		t.Errorf("norm question: binding must stay Top-1 over standard, got %s", out[0].SourceClass)
-	}
-}
-
-func TestRerank_OffTopicStandard_BlockedByGuard(t *testing.T) {
-	// Control intent present, but the standard is semantically far below binding →
-	// the margin guard keeps binding Top-1 (no off-topic standard override).
-	results := []LegalSearchResult{
-		intentRes("NIST SP 800-82", "technical_standard", 0.40, 80),
-		intentRes("CRA", "binding_law", 0.58, 100),
-	}
-	out := rerankByAuthority("Welche Controls passen zu Security Updates?", results)
-	if out[0].SourceClass != "binding_law" {
-		t.Errorf("off-topic standard must not win even with control intent, got %s", out[0].SourceClass)
-	}
-}