fix(ai-sdk): national-law subsidiarity in authority rerank (DSGVO > BDSG for general questions)
CI / detect-changes (pull_request) Successful in 11s
CI / branch-name (pull_request) Successful in 2s
CI / guardrail-integrity (pull_request) Successful in 9s
CI / secret-scan (pull_request) Successful in 11s
CI / dep-audit (pull_request) Failing after 54s
CI / sbom-scan (pull_request) Failing after 59s
CI / build-sha-integrity (pull_request) Successful in 8s
CI / validate-canonical-controls (pull_request) Successful in 8s
CI / loc-budget (pull_request) Successful in 23s
CI / go-lint (pull_request) Successful in 57s
CI / python-lint (pull_request) Failing after 16s
CI / nodejs-lint (pull_request) Failing after 1m11s
CI / nodejs-build (pull_request) Successful in 3m4s
CI / test-go (pull_request) Successful in 1m1s
CI / iace-gt-coverage (pull_request) Successful in 18s
CI / test-python-backend (pull_request) Successful in 25s
CI / test-python-document-crawler (pull_request) Successful in 14s
CI / test-python-dsms-gateway (pull_request) Successful in 12s

The authority reranker (wired in legal_rag_client.go:168) had no national-subsidiarity
dimension, so a general BDSG paragraph could outrank the primary DSGVO article. Surfaced by
the KB-2026.1 BDSG pilot (dp_05/08/11 + cr_07).

- authorityScore: DE binding_law in an EU-primary domain WITHOUT a co-primary topic match
  -> soft demote (subsidiarityPen 0.18), not exclusion. National special rules stay
  co-primary via the topic ontology (DSB Art.37+§38, special categories Art.9+§22, ...).
- queryDomain: fall back to a regulation-name mention (DSGVO/BDSG/CRA) so a question phrased
  around the act is domain-scoped even without a topical keyword (fixes cr_07: BDSG Teil-3 §64).
- data_protection keyword stem 'auftragsverarbeit' (catches Auftragsverarbeitungsvertrag).

Pure ranking logic, no data manipulation; soft demotes keep national rules visible.
Build result (DSGVO+BDSG): degraded=0, must_not=0. go build/vet/test ./... green; 6 new table tests.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Benjamin Admin
2026-06-26 21:28:12 +02:00
parent 8609b696c9
commit 623d80b6c8
3 changed files with 98 additions and 1 deletions
+21 -1
View File
@@ -112,7 +112,7 @@ var domains = []domainDef{
{"data_protection",
[]string{"DSGVO", "GDPR", "BDSG", "EDPB", "DSK", "BfDI", "BayLfD", "DPF"},
[]string{"personenbezogen", "betroffene", "datenschutz", "datenschutzbeauftrag", "dsb",
"datenpanne", "auskunft", "loesch", "lösch", "einwilligung", "besondere kategorien", "auftragsverarbeiter"}},
"datenpanne", "auskunft", "loesch", "lösch", "einwilligung", "besondere kategorien", "auftragsverarbeit"}},
{"cyber",
[]string{"CRA", "NIS2", "NIS-2", "ENISA", "DORA", "EUCC"},
[]string{"security update", "sicherheitsupdate", "sicherheitsaktualisierung", "schwachstelle", "sbom",
@@ -126,6 +126,16 @@ var domains = []domainDef{
nil},
}
// euPrimaryDomains are domains whose PRIMARY binding act is an EU regulation/directive
// (DSGVO, CRA/NIS2, AI Act, MaschinenVO). In these domains a NATIONAL implementing law
// (e.g. BDSG) is subsidiary for general questions — see nationalSubsidiarityPenalty.
var euPrimaryDomains = map[string]bool{
"data_protection": true,
"cyber": true,
"ai": true,
"product_safety": true,
}
func queryDomain(query string) string {
ql := strings.ToLower(query)
for _, d := range domains {
@@ -135,6 +145,16 @@ func queryDomain(query string) string {
}
}
}
// Fallback: an explicit regulation mention (e.g. "DSGVO", "BDSG", "CRA") also signals the
// domain — so a question phrased around the act ("... gilt die DSGVO ...") is scoped even
// without a topical keyword. Keyword match wins first (more specific).
for _, d := range domains {
for _, reg := range d.regs {
if strings.Contains(ql, strings.ToLower(reg)) {
return d.name
}
}
}
return ""
}