diff --git a/.github/workflows/audit.yml b/.github/workflows/audit.yml
new file mode 100644
index 0000000..2dc8164
--- /dev/null
+++ b/.github/workflows/audit.yml
@@ -0,0 +1,22 @@
+name: Registry audit
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    paths: ['sources/**', 'scripts/**']
+
+concurrency:
+  group: audit-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  audit:
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    steps:
+      - uses: actions/checkout@v4
+      - uses: ruby/setup-ruby@v1
+        with:
+          ruby-version: '3.3'
+      - run: ruby scripts/registry_audit.rb
diff --git a/scripts/registry_audit.rb b/scripts/registry_audit.rb
new file mode 100755
index 0000000..d6eac3b
--- /dev/null
+++ b/scripts/registry_audit.rb
@@ -0,0 +1,107 @@
+#!/usr/bin/env ruby
+# Registry integrity audit for the SIGNATIF requirements/conformance data.
+# Exits non-zero on any failure. Run from the repository root.
+require 'yaml'
+require 'set'
+
+ROOT = File.expand_path('..', __dir__)
+DATA = File.join(ROOT, 'sources/data')
+SECTIONS = File.join(ROOT, 'sources/sections')
+failures = []
+fail = ->(msg) { failures << msg }
+
+rc_files = Dir.glob("#{DATA}/*-rc.yaml").sort
+cc_files = Dir.glob("#{DATA}/*-cc.yaml").sort
+
+# 1. parse + collect
+req_ids = {}
+req_names = {}
+tests = 0
+test_targets = []
+rc_files.each do |f|
+  d = YAML.safe_load(File.read(f), permitted_classes: [], aliases: false)
+  (d['groups'] || []).each do |g|
+    (g['scopes'] || []).each do |sc|
+      (sc['requirements'] || []).each do |r|
+        id = "#{sc['identifier']}/#{r['identifier_fragment']}"
+        req_ids[id] = f
+        (req_names[r['name']] ||= []) << sc['identifier']
+        stmt = r['statement'].to_s
+        fail "NO-SHALL #{id}" unless stmt =~ /\bshall\b/
+        fail "THIN #{id}" if stmt.split(/\s+/).length < 6
+        fail "NON-KEBAB #{id}" unless r['identifier_fragment'] =~ /\A[a-z0-9]+(-[a-z0-9]+)*\z/
+      end
+    end
+  end
+end
+cc_files.each do |f|
+  d = YAML.safe_load(File.read(f), permitted_classes: [], aliases: false)
+  (d['groups'] || []).each do |g|
+    (g['scopes'] || []).each do |sc|
+      (sc['tests'] || []).each do |t|
+        tests += 1
+        (t['targets'] || []).each { |tg| test_targets << tg }
+        fail "TEST-NO-METHOD #{sc['identifier']}/#{t['identifier_fragment']}" if t['method'].to_s.strip.empty?
+      end
+    end
+  end
+end
+req_count = req_ids.size
+
+# 2. conformance classes match requirements classes per file pair
+rc_classes = Set.new
+rc_files.each do |f|
+  YAML.safe_load(File.read(f))['groups'].each { |g| g['scopes'].each { |sc| rc_classes << sc['identifier'].sub('/req/', '') } }
+end
+cc_classes = Set.new
+cc_files.each do |f|
+  YAML.safe_load(File.read(f))['groups'].each { |g| g['scopes'].each { |sc| cc_classes << sc['identifier'].sub('/conf/', '') } }
+end
+fail "class set mismatch: req-only #{(rc_classes - cc_classes).to_a} conf-only #{(cc_classes - rc_classes).to_a}" unless rc_classes == cc_classes
+
+# 3. dependency closure
+defined = Set.new(rc_classes.map { |c| "/req/#{c}" })
+rc_files.each do |f|
+  YAML.safe_load(File.read(f))['groups'].each do |g|
+    g['scopes'].each do |sc|
+      (sc['dependencies'] || []).each do |dep|
+        fail "dep target missing #{dep} (from #{sc['identifier']})" unless defined.include?(dep)
+      end
+    end
+  end
+end
+
+# 4. test targets resolve
+test_targets.each { |t| fail "test target missing #{t}" unless req_ids.key?(t) }
+
+# 5. unique requirement names
+req_names.each { |n, cs| fail "duplicate requirement name '#{n}' in #{cs.uniq.join(', ')}" if cs.size > 1 }
+
+# 6. requirements == tests
+fail "count mismatch: #{req_count} requirements vs #{tests} tests" unless req_count == tests
+
+# 7. bibliography: entries all cited (adoc + yaml)
+bib = File.read("#{SECTIONS}/99-bibliography.adoc")
+ids = bib.scan(/^\* \[\[\[([a-zA-Z0-9_-]+),/).flatten
+corpus = Dir.glob("#{SECTIONS}/*.adoc").reject { |f| File.basename(f) == '99-bibliography.adoc' }.map { |f| File.read(f) }.join("\n")
+corpus += Dir.glob("#{DATA}/*.yaml").map { |f| File.read(f) }.join("\n")
+ids.each { |i| fail "bib entry uncited: #{i}" unless corpus =~ /<<#{i}(?:,|>>)/ }
+
+# 8. anchors unique outside code fences (fence-aware)
+anchors = Hash.new(0)
+Dir.glob("#{SECTIONS}/*.adoc").each do |f|
+  fence = false
+  File.foreach(f) do |line|
+    fence = !fence if line.start_with?('----')
+    next if fence
+    line.scan(/^\[\[([a-zA-Z0-9_-]+)\]\]/).flatten.each { |a| anchors[a] += 1 }
+  end
+end
+anchors.each { |a, n| fail "duplicate anchor #{a} (#{n}x)" if n > 1 }
+
+if failures.any?
+  warn "REGISTRY AUDIT FAILED (#{failures.size}):"
+  failures.each { |m| warn " - #{m}" }
+  exit 1
+end
+puts "registry audit passed: #{rc_classes.size} classes, #{req_count} requirements, #{tests} tests, #{ids.size} cited bibliography entries"