fix: reject section-marker addresses and role+number orgs

.changeset/doc-scaffold-fp-guards.md

@@ -0,0 +1,5 @@
+---
+"@stll/anonymize": patch
+---
+
+Reject section-marker addresses and generic-role page-number organizations as false positives.

crates/anonymize-core/src/false_positives.rs

@@ -146,7 +146,13 @@ fn should_reject_entity(
   if exceeds_open_ended_word_count(entity) {
     return Ok(true);
   }
+  // Section markers (`6.1`, `3.2.4`) are never addresses, including when an
+  // address cue (place-of-performance) extracts them as trigger values.
+  if entity.label == ADDRESS_LABEL && is_section_number(text) {
+    return Ok(true);
+  }
   if entity.label != IP_ADDRESS_LABEL
+    && entity.label != ADDRESS_LABEL
     && is_section_number(text)
     && entity.source != DetectionSource::Trigger
   {
@@ -206,6 +212,14 @@ fn should_reject_entity(
   {
     return Ok(true);
   }
+  // Municipality / party-role prefix cues can capture page footers such as
+  // "Strana 7" / "Page 3". A generic-role head plus only a number is scaffolding,
+  // not an organization name.
+  if entity.label == ORGANIZATION_LABEL
+    && filters.is_some_and(|filters| is_generic_role_plus_number(text, filters))
+  {
+    return Ok(true);
+  }
   if entity.label == ADDRESS_LABEL && should_reject_address(entity, filters) {
     return Ok(true);
   }
@@ -444,6 +458,24 @@ fn role_exact_match(
       .contains(&entity.text.trim().to_lowercase())
 }
 
+/// True when the span is a generic role word followed only by a bare number
+/// (`Strana 7`, `Page 3`, `Contractor 1`). Uses the shared role vocabulary so
+/// the guard is not tied to any one language's page-footer spelling.
+fn is_generic_role_plus_number(
+  text: &str,
+  filters: &DenyListFilterData,
+) -> bool {
+  let trimmed = text.trim();
+  let Some((word_end, word)) = first_word(trimmed) else {
+    return false;
+  };
+  if !filters.generic_roles.contains(&word.to_lowercase()) {
+    return false;
+  }
+  let rest = trimmed.get(word_end..).unwrap_or("").trim();
+  !rest.is_empty() && rest.chars().all(|ch| ch.is_ascii_digit())
+}
+
 fn is_all_caps_candidate(text: &str) -> bool {
   let mut has_upper = false;
   for ch in text.chars().filter(|ch| ch.is_alphabetic()) {
@@ -1829,6 +1861,71 @@ mod tests {
     assert!(entities.is_empty());
   }
 
+  #[test]
+  fn rejects_trigger_address_section_numbers() {
+    let entities = filter_entity_false_positives(
+      vec![entity(
+        "6.1",
+        "6.1",
+        ADDRESS_LABEL,
+        DetectionSource::Trigger,
+      )],
+      "6.1. Place of performance",
+      Some(&DenyListFilterData::default()),
+    )
+    .unwrap();
+
+    assert!(entities.is_empty());
+  }
+
+  #[test]
+  fn rejects_generic_role_plus_number_organizations() {
+    let filters = DenyListFilterData {
+      // Vocabulary-driven: any generic role, not a Czech-only page word.
+      generic_roles: set(["contractor", "strana"]),
+      ..DenyListFilterData::default()
+    };
+    let text = "Contractor 1\nStrana 7\nAcme Industries";
+    let contractor = text.find("Contractor 1").unwrap();
+    let strana = text.find("Strana 7").unwrap();
+    let acme = text.find("Acme Industries").unwrap();
+    let entities = filter_entity_false_positives(
+      vec![
+        PipelineEntity::detected(
+          u32::try_from(contractor).unwrap(),
+          u32::try_from(contractor.saturating_add("Contractor 1".len()))
+            .unwrap(),
+          ORGANIZATION_LABEL,
+          "Contractor 1",
+          0.8,
+          DetectionSource::Trigger,
+        ),
+        PipelineEntity::detected(
+          u32::try_from(strana).unwrap(),
+          u32::try_from(strana.saturating_add("Strana 7".len())).unwrap(),
+          ORGANIZATION_LABEL,
+          "Strana 7",
+          0.8,
+          DetectionSource::Trigger,
+        ),
+        PipelineEntity::detected(
+          u32::try_from(acme).unwrap(),
+          u32::try_from(acme.saturating_add("Acme Industries".len())).unwrap(),
+          ORGANIZATION_LABEL,
+          "Acme Industries",
+          0.8,
+          DetectionSource::Trigger,
+        ),
+      ],
+      text,
+      Some(&filters),
+    )
+    .unwrap();
+
+    assert_eq!(entities.len(), 1);
+    assert_eq!(entities[0].text, "Acme Industries");
+  }
+
   #[test]
   fn keeps_ipv4_addresses_that_resemble_section_numbers() {
     let text = "192.0.2.1";

packages/anonymize/src/__test__/document-scaffold-fp.test.ts

@@ -0,0 +1,81 @@
+/**
+ * False-positive guards for document scaffolding mistaken as PII:
+ * section-marker addresses and generic-role + page-number organizations.
+ * Motivated by Czech Registr smluv mandate contracts; rules are vocabulary-
+ * driven / structural, not Czech-token checks in core logic.
+ */
+import { describe, expect, setDefaultTimeout, test } from "bun:test";
+import { DEFAULT_ENTITY_LABELS } from "../constants";
+import type { Dictionaries, PipelineConfig } from "../types";
+import { detectNative } from "./native-detect";
+import { loadTestDictionaries } from "./load-dictionaries";
+
+setDefaultTimeout(60_000);
+
+const CONFIG: PipelineConfig = {
+  threshold: 0.3,
+  enableTriggerPhrases: true,
+  enableRegex: true,
+  enableLegalForms: true,
+  enableNameCorpus: true,
+  enableDenyList: true,
+  enableGazetteer: false,
+  enableConfidenceBoost: true,
+  enableCoreference: true,
+  enableHotwordRules: true,
+  denyListCountries: ["CZ"],
+  nameCorpusLanguages: ["cs"],
+  labels: [...DEFAULT_ENTITY_LABELS],
+  workspaceId: "doc-scaffold-fp-test",
+};
+
+let cachedDictionaries: Dictionaries | undefined;
+const detect = async (text: string) => {
+  cachedDictionaries ??= await loadTestDictionaries();
+  return detectNative({ ...CONFIG, dictionaries: cachedDictionaries }, text);
+};
+
+describe("document scaffolding false positives", () => {
+  test("place-of-performance cue does not keep a section marker as address", async () => {
+    const text =
+      "Místo plnění\n\n6.1. Místem plnění se rozumí: Technická správa města";
+    const entities = await detect(text);
+    expect(
+      entities.some((e) => e.label === "address" && e.text.trim() === "6.1"),
+    ).toBe(false);
+  });
+
+  test("generic-role page footer is not an organization (vocabulary-driven)", async () => {
+    // Role word is English `contractor` from generic-roles.json, not hardcoded
+    // in detector logic.
+    const text = "starosta města\n\nContractor 1 (of 3)\n";
+    const entities = await detect(text);
+    expect(
+      entities.some(
+        (e) => e.label === "organization" && /^Contractor\s+1$/u.test(e.text),
+      ),
+    ).toBe(false);
+  });
+
+  test("ordinary municipality organization trigger is unchanged", async () => {
+    const text = "město Brandýs nad Labem, IČO: 00240066";
+    const entities = await detect(text);
+    expect(
+      entities.some(
+        (e) =>
+          e.label === "organization" && e.text.includes("Brandýs nad Labem"),
+      ),
+    ).toBe(true);
+  });
+
+  test("Czech page footer Strana N is not an organization", async () => {
+    const text =
+      "Mgr. et Bc. Milan Rychtařík\n         starosta města\n\n                                               Strana 7 (celkem 7)\n";
+    const entities = await detect(text);
+    expect(
+      entities.some(
+        (e) => e.label === "organization" && /^Strana\s+7$/u.test(e.text),
+      ),
+    ).toBe(false);
+  });
+});