overthelex/ua-metodyka-audit

AMCU Методика Audit (2002-2026)

A guidance instrument, every citation of it in twenty four years of the issuing agency's and the courts' decisions, and the manifest of the corpus both were read from.

The instrument is the Методика визначення монопольного (домінуючого) становища суб'єктів господарювання на ринку, approved by розпорядження of the Antimonopoly Committee of Ukraine N 49-р of 5 March 2002, registered with the Ministry of Justice on 1 April 2002 under N 317/6605 (z0317-02). It was never amended — the register's card lists two entries under "Редакції документа" and the second is the repeal — and it lost force on 1 August 2026, superseded by z1043-26.

The method is the computational audit of a guidance instrument against the agency's own decisional record: which of its propositions the record uses, and which nobody uses.

Load

from datasets import load_dataset

props = load_dataset("overthelex/ua-metodyka-audit", "propositions", split="train")
cites = load_dataset("overthelex/ua-metodyka-audit", "citations", split="train")
corpus = load_dataset("overthelex/ua-metodyka-audit", "corpus", split="train")

Configs

propositions — 66 rows

The instrument cut into its numbered items. The numbered item is the unit because it is the unit the record cites: decisions say "пунктом 6.1 Методики", never "the second sentence of 6.1".

field
number	"6.1"
rozdil, rozdil_title	the section it belongs to
depth	2 for 6.1, 3 for 2.1.11
text	the item, from its own number to the next

citations — 2,315 rows

Every document found to cite the Методика: 167 decisions of the AMCU and 2,148 of the courts. One row per document.

field
corpus	amcu or court
doc_id	its identifier in the source register
mentions	how many times the instrument is named
cites	list of {kind, number, context}; kind is punkt or rozdil

labels — 1,657 rows

Every judgement made about a (proposition, passage) pair: whether the passage applies the rule, contradicts it, only recites it, or is about something else. Four annotators on the same written protocol (raw/protocol.md), which each received verbatim.

annotator	rows
claude-sonnet-4.6	528	the whole review packet
deepseek-r1	528	the whole review packet
mistral-large	522	run and then discarded, see below
human	79	the passages that decide the outcome

The human read 79 passages chosen because they decide the result — the 40 where the two retained models split on application, plus 39 drawn at random from the ones they agreed on, shuffled together and with the model labels withheld. On that blind control the reader agrees with each model on 97.4% (Cohen's κ = 0.95) and is stricter than either where they part, labelling 17.7% of passages "applies" against 35.4% and 22.8%.

mistral-large is published because a discarded annotator is evidence too. It matches on topic rather than on use: it called a verbatim copy of provision 2.1 "applies" because the passage "lists the actions described in provision 2.1". Do not pool it with the others.

field
number, doc_id, ord	the passage, keyed as in propositions and corpus
annotator	one of the four above
label	застосовує / суперечить / лише переказує / не про це
why	the model's one-sentence reason; empty for the human

corpus — 8,980 rows

The manifest of the documents the audit read, with the SHA of each normalised text. The decision texts themselves are not redistributed here: they are public at their sources, and the manifest is what lets a rebuild be checked document by document rather than by a row count — including which copy was kept of the 643 decisions the portal publishes twice.

field
doc_id, corpus, doc_ref, decision_date
chars, passages	length, and passages contributed
body_hash	MD5 of the text with non-alphanumerics collapsed
canonical	false for the second copy of a twice-published decision

Files

data/ holds the four configs as parquet. raw/ holds what datasets.py actually emits, byte for byte, which is also what data/metodyka/ in the code repository contains: propositions.json, goldset.jsonl, corpus-manifest.csv, plus coverage.json (the two crossed, per proposition) and checksums.txt over all of them, plus labels.jsonl and the labelling protocol protocol.md. The parquet files are a convenience for load_dataset; the raw files are the record.

What the labelling shows

41 of the 66 provisions show no application anywhere in the packet. This is a lower bound rather than an estimate: the human reader is the stricter judge, read only the passages that decide the outcome, and the reading only ever moved provisions into that set — the models alone gave 34, the reading added seven.

It is not an artefact of what the packet offered. The median number of non-recitation passages is 8 of 8 for the provisions with no application and 8 of 8 for the others. Provisions 1.2 (260 citations), 2.2 (258), 7.1 (203) and 3.1 (126) each had eight clean candidates and showed application in none.

Fourteen of the 41 rest on five citations or fewer in twenty four years, and are worth treating separately for that reason.

What the citations show

2,315 documents cite the instrument, and 62 of its 66 propositions are cited by number at least once.

Never cited by anyone in twenty four years: 10.1.6.1, 10.1.6.3, 10.2.4, and the one-word item 4.2.4. The first three are the whole machinery of collective dominance.

The agency and the courts use different halves. The agency cites the market-definition sections (6.1 fifty five times, 5.1 fifty two, 3.1 forty five) and hardly touches section 10, the dominance test itself: пункт 10.2 never, and 10.2.1, which states the 35% threshold, never.

That asymmetry is not an artefact of redaction. Of the 62 agency decisions that discuss the 35% threshold and mention the Методика, 52 cite статтю 12 of the Law and only 3 cite пункт 10.2. Section 10 restates the statute, so the agency goes to the statute; what the instrument contributes on its own authority is market definition, and that is where the agency uses it.

Finding the citations

"Методика" is a generic word, and the court register is full of методики for computing tariffs, damages and unmetered electricity. A mention counts as this instrument only when its title follows it, or when the document establishes the title and carries no competing one. A pinpoint binds only to the reference immediately beside the mention: without that, a statute's "п. 23" was attached to the Методика for sharing a sentence with it.

Three traps in the data, each of which changed a number:

• пункту 3 розділу 1 is item 1.3, not 3;
• п. 1.З. is a Cyrillic З typed for a 3;
• the register wraps lines inside a number, and 10.1\n.5 read as an item the instrument does not have.

Sources

Read from the issuing bodies, not from an intermediary.

• Agency decisions: the AMCU's own open data publication on data.gov.ua, dataset "Рішення та рекомендації Антимонопольного комітету України", package 8bdd45b8-0684-463a-ba76-26361c32841a. Each archive is named after the CKAN resource that published it, so every decision traces back to the published file it came out of.
• Court decisions: Єдиний державний реєстр судових рішень.
• The instrument: the Verkhovna Rada's legislation register.

Limits

• The agency corpus begins in 2017, because that is when the portal begins. The instrument dates from 2002, so the long half of the record is the court side, and nothing here supports a claim about agency practice before 2017.
• 2,498 agency decisions carry redaction markers ("Інформація, доступ до якої обмежено"); in 1,093 of them the markers are more than a tenth of the characters and in 52 more than a quarter. Names and figures are what is removed and the reasoning survives, but a claim about a particular figure in a particular decision cannot be checked here.
• The court side is restricted to decisions that name the Методика, so a proposition applied in a case that never names the instrument is invisible to this design. The restriction is deliberate: it keeps the measurement from drowning in competition cases decided on the statute alone.
• Retrieval over the corpus (bge-m3 CLS over 191,593 passages, union with keyword search) reaches proposition recall 0.887 at k=200 on the 3,673 citation pairs. The eight propositions it cannot reach are cited 20 times out of 3,699, half a percent of the record.

Code

Every file here is produced by scripts/metodyka/ and none of it was edited by hand.

stage
cut the instrument	metodyka.py
find the citations	goldset.py
cross the two	coverage.py
corpus, passages, retrieval	retrieve.py
embedding on a GPU	gpu_io.py, gpu_startup.sh, run_gpu.sh
the review packet	packet.py
export these files	datasets.py