Tighten verifier: institutional-context tokens don't count as topical match

Closes the "blind spot A" surfaced during the gemma4 Kasus Ibam diagnostic:
the model cited Permen 24/2018 claiming an audit-mechanism topic, when the
actual title is "Pengelolaan Arsip Terjaga di Lingkungan Kementerian
Pendidikan dan Kebudayaan". The previous overlap check passed because claim
and title shared "kementerian / pendidikan / kebudayaan" — institutional
context, not topic.

_topics_overlap now requires at least one shared token to be NON-
institutional. Adds _INSTITUTIONAL_TOKENS frozenset covering generic
government context (kementerian / pemerintah / lingkungan / instansi /
lembaga / wilayah / kantor) and ministry/sector markers (pendidikan /
kebudayaan / kesehatan / keuangan / ketenagakerjaan / pertanian / etc.).

The acronym-expansion path is unaffected — institutional acronyms (MPR /
DPR / KPK) still match their spelled-out forms because their expansion
tokens (majelis / dewan / komisi / pemberantasan) are deliberately kept
out of the institutional list.

Tests: 4 new cases in test_citations.py — institutional-only rejection,
real topical match, acronym-vs-spelled-out, malformed-title fallback.

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

src/legawa/tools/citations.py

@@ -101,6 +101,34 @@ _STOP_TOKENS = frozenset(
     }
 )
 
+# Tokens that name an institution, sector, or generic government context.
+# These appear in the *title* of nearly every regulation issued by that
+# institution (e.g. "Kementerian Pendidikan" appears in every Kemendikbud
+# Permen) and in any *claim* that names the institution — so an overlap on
+# these tokens alone does not confirm topical alignment.
+#
+# Used by ``_topics_overlap`` to demand at least one *non-institutional*
+# token in common before declaring a match. Acronyms (MPR/DPR/KPK/...) are
+# kept out of this set because their expansions ARE the topic in regulations
+# whose subject is the institution itself (e.g. UU 17/2014 about MPR/DPR).
+_INSTITUTIONAL_TOKENS = frozenset(
+    {
+        # Generic government context.
+        "kementerian", "pemerintah", "pemerintahan", "negeri", "lingkungan",
+        "instansi", "lembaga", "publik", "pusat", "lingkup", "wilayah",
+        "kantor",
+        # Sector markers — when alone, signal sector not topic. Specific
+        # subject matter (audit / pengelolaan / arsip / outsourcing / etc.)
+        # is what should drive topical match, not the ministry name.
+        "pendidikan", "kebudayaan", "kesehatan", "keuangan",
+        "ketenagakerjaan", "pertanian", "perhubungan", "kehutanan",
+        "kelautan", "pertahanan", "agama", "sosial", "perindustrian",
+        "perdagangan", "ristek", "riset", "teknologi", "informasi",
+        "komunikasi", "perekonomian", "perencanaan", "pembangunan",
+    }
+)
+
+
 # Common Indonesian institutional acronyms — agents tend to use the acronym
 # while the regulation's actual title spells the institution out. We expand
 # claim-side acronyms before topical-overlap so e.g. "MPR" matches a title
@@ -169,7 +197,13 @@ def _topics_overlap(claimed: str, title: str) -> bool:
 
     claim_tokens = _significant_tokens(claimed)
     title_tokens = _significant_tokens(title)
-    if claim_tokens & title_tokens:
+
+    # Direct token overlap counts only if at least one shared token is
+    # non-institutional. This guards against false positives where claim
+    # and title share only ministry/sector context (e.g. "Kementerian
+    # Pendidikan") while their actual subjects are different.
+    overlap = claim_tokens & title_tokens
+    if overlap and (overlap - _INSTITUTIONAL_TOKENS):
         return True
 
     # Expand any claim-side acronyms and check for partial title overlap.

tests/test_citations.py

@@ -10,6 +10,7 @@ if str(SRC) not in sys.path:
 
 from legawa.tools.citations import (
     CitationCheck,
+    _topics_overlap,
     extract_citations,
     format_checks,
     verify_citation,
@@ -96,6 +97,32 @@ class CitationTests(unittest.TestCase):
         self.assertEqual(perda.query, "Perda 1/2024")
         self.assertEqual(perda.frbr_uri, "akn/id/act/perda/2024/1")
 
+    def test_topics_overlap_rejects_institutional_only_match(self) -> None:
+        # Real false-positive case from gemma4 memo run: model claimed
+        # Permen 24/2018 was about audit-mechanism, but it's about Pengelolaan
+        # Arsip. They share only ministry-context tokens.
+        claim = "Mekanisme Tindak Lanjut Hasil Audit Inspektorat Jenderal Kementerian Pendidikan dan Kebudayaan"
+        title = "Pengelolaan Arsip Terjaga di Lingkungan Kementerian Pendidikan dan Kebudayaan"
+        self.assertFalse(_topics_overlap(claim, title))
+
+    def test_topics_overlap_accepts_real_topical_match(self) -> None:
+        claim = "Pengadaan Barang/Jasa Pemerintah"
+        title = "Peraturan Presiden Nomor 12 Tahun 2021 tentang Perubahan Atas Pengadaan Barang/Jasa Pemerintah"
+        self.assertTrue(_topics_overlap(claim, title))
+
+    def test_topics_overlap_acronym_vs_spelled_out_still_works(self) -> None:
+        # Ensure the institutional filter doesn't break the acronym path —
+        # UU 17/2014 is genuinely about MPR/DPR/DPD/DPRD.
+        claim = "MPR, DPR, DPD, dan DPRD"
+        title = "Majelis Permusyawaratan Rakyat, Dewan Perwakilan Rakyat, Dewan Perwakilan Daerah, dan Dewan Perwakilan Rakyat Daerah"
+        self.assertTrue(_topics_overlap(claim, title))
+
+    def test_topics_overlap_uninformative_title_falls_back(self) -> None:
+        # Echoed pasal.id title — no usable signal — accept by existence-only.
+        claim = "Sistem Pendidikan Nasional"
+        title = "Undang-Undang Nomor 20 Tahun 2003 tentang Undang-Undang Nomor 20 Tahun 2003"
+        self.assertTrue(_topics_overlap(claim, title))
+
     def test_verify_citations_formats_mixed_results(self) -> None:
         fake = FakePasalClient(
             {