think
16px
820px

Akta mislabel rate — full verification of all 15 distinct deeds

Date: 2026-08-13 · Method: read each deed's operative decision (Memutuskan / menyetujui items, excluding anggaran-dasar boilerplate) and compare against the authoritative label SubmissionDocument.classifiedType. Read-only.

Headline

Distinct akta deeds 15
Deeds carrying a submission label 14
Mislabelled 5
Rate ~36% (5 of 14)
Correct 8
Unresolved (sparse OCR, nothing contradicts) 1
Earlier estimate 13% — that was a floor, as flagged

Two failure patterns, one root behaviour

Every mislabel is a RUPS-decision deed collapsed into AKTA_PERUBAHAN (or, in one case, into another wrong bucket), regardless of what the shareholders actually resolved:

Pattern A — dissolution read as amendment (2):

hash stored actually resolved correct
b940a54b AKTA_PERUBAHAN membubarkan Perseroan Terbatas PT FAMILY LOMBARD pembubaran
d90b6736 AKTA_PERUBAHAN acara rapat adalah: I. Pembubaran Perseroan; II. Penunjukan Likuidatorkeadaan "Bubar" pembubaran

Pattern B — annual report read as amendment (3, new):

hash stored actually resolved correct
0bd27829 AKTA_PERUBAHAN Menerima dan mengesahkan Laporan Tahunan Perseroan untuk tahun buku … 31-12-2025 … Neraca dan Laporan Laba Rugi + Direksi accountability report laporan_rups_tahunan
31731874 AKTA_PERUBAHAN menyetujui Laporan Tahunan Perseroan untuk tahun buku 2025 + Menyetujui pembagian dividen laporan_rups_tahunan
38358b45 AKTA_PERALIHAN_PP_PT Menyetujui dan mengesahkan laporan tahunan untuk tahun buku yang berakhir 31-12-2025; the term Perseroan Perorangan appears 0 times, so nothing supports a PP→PT conversion laporan_rups_tahunan

laporan_rups_tahunan / AKTA_RUPS_TAHUNAN already exists in the taxonomy and is already a built flow, so these are misroutes into a wrong existing bucket, not gaps in coverage.

Verified correct (8)

hash label evidence
54c8b6da AKTA_RUPS_TAHUNAN menerima Laporan Direksi … tahun buku 2025, approves Neraca/Laba Rugi
25dcede7 AKTA_RUPS_TAHUNAN full annual RUPS (laporan tahunan×8, tahun buku×27) plus extra agenda (corporate guarantee, remuneration delegation)
9b0fd428 AKTA_PERUBAHAN Menyetujui perubahan maksud dan tujuan serta kegiatan usaha … Pasal 3 Anggaran Dasar diubah — a genuine amendment
81c76eb5 AKTA_PERALIHAN_PP_PT Perseroan Perorangan×4, filename perubahan-pp-ke-pt.pdf
1e64150a AKTA_BERAKHIRNYA liquidator acquittal
112c9474 AKTA_PENGGABUNGAN post-merger PKRUPSLB
fe8107e6, bf5eaa7c AKTA_PENDIRIAN determinative PENDIRIAN PERSEROAN TERBATAS titles

Unresolved: 8d1fa0d7 (stored AKTA_PERUBAHAN) — OCR too sparse to locate any decision text; filename akta-perubahan.pdf and nothing contradicts the label. Not counted as a mislabel.

Not a mislabel but worth noting: 9e829396 has no SubmissionDocument row at all — it carries no authoritative label. Document.aktaSubtype says pendirian. Worth checking why a processed document is unattached to any submission.

What this means

The engine session's fix (51a56fa5) added deterministic routing for the dissolution lifecycle — that closes Pattern A. Pattern B is not covered: an annual-report RUPS still depends on the LLM, and the LLM currently reads it as perubahan.

The good news is that Pattern B has an unusually clean deterministic signal — the decision items literally say Laporan Tahunan … tahun buku <year> / mengesahkan … Neraca dan Laporan Laba Rugi. That is at least as strong as the pembubaran signal, and it sits in the same operative-decision span decisionSpan() already extracts. The guard needed is the mirror of the one already written: the anggaran dasar of every deed contains an article about annual RUPS duties (Persetujuan laporan tahunan dan pengesahan laporan keuangan oleh RUPS tahunan berarti memberikan pelunasan…), and an unbounded scan for laporan tahunan would route founding deeds to laporan_rups_tahunan. I hit exactly that false positive while doing this audit — my first extraction pass matched that boilerplate on all three pendirian deeds.

Confidence and limits

These are my readings of the decision recitals, not notaris-verified adjudications. Pattern B's three cases are unambiguous (explicit annual-report approval for a named financial year). Pattern A's two were independently confirmed by the engine session. The single unresolved deed and the one unlabelled document would need a human to close.

Sample is 15 distinct deeds — small, and it is the entire distinct akta corpus in staging, so it cannot be widened without new documents.