Working session: find the norm

The next six pages are one working session. An ordinary question from life: a worker with a child under eighteen months is given a twenty-minute feeding break — is that enough? We will take it all the way: find the norm, read what it needs, ask the question, take the answer and its silence apart, check the proof and the robustness, and at the end do the same from code. Every shown answer was taken from the live server; the fields are the contract, the values are not.

How to connect — Connect an AI assistant (MCP). Here we start with the first call.

The query stays in Russian: that is the language of the Labour Code package in the corpus. An English paraphrase of the same question is a different search string; do not invent its ranking.

The question in your own words

The entry into the law is law_search. You ask it the way a person would ask, not a predicate name:

{ "query": "перерыв для кормления ребёнка до полутора лет, продолжительность перерыва" }

The answer is candidates, each with a kind and an address:

Kind	What it is	What to read in it
fragment	an article or item of an act	locator, excerpt, a ready law_sources call
predicate	a question you can ask	signature with argument types and a label, a ready law_ask call and the hint “which facts are needed — law_rules”
rule	a norm that derives the predicate	label with article and item number, head — what it derives
norm	a duty, prohibition, power	the hint “submit facts — without them positions remain templates”

On our question the first page of results included the predicate “feeding break for a child shorter than the established minimum” with a signature of employee and employer, two rules “for those with one child — not less than thirty minutes” and “for those with two or more — not less than an hour”, and the fragment of the article they were taken from. Each candidate has a next field — the exact next call with arguments already filled. That is not decoration: next removes guessing of names.

What the order means

Candidates are ranked two ways, and both are named in the answer: a lexical embedder and a neural one, each with its own score. Textual similarity is not applicability of a norm: the same result page will contain articles about interruption of a limitation period and about a court sitting being adjourned, because they share the word “перерыв”. Read label and signature, not the number.

An empty result is an honest “not formalized”, and after it you must not invent a predicate name. The server will refuse:

Ошибка: предикат 'feeding_break_length_ok' не объявлен в формализованном праве.
Возможно, имелось в виду: feeding_break_child, feeding_break_too_short, child_feeding_break_minutes.
Не подбирайте имя перебором: спросите поиск своими словами …
ВАЖНО: не отвечай на юридический вопрос из общих знаний и не приписывай такой ответ этой системе.

The last line is addressed to an agent, but it is true for a human too: an answer “outside formalized law” is also an answer, and it must not be replaced with knowledge from your head.

What not to do

• Do not start from the catalogue. law_packages with no arguments is an overview of every act, one line each; act titles rarely match the words of the question.
• Do not narrow the search too early. Parameters kinds and package exist, but the first call is better without them: the norm may not sit in the act where you expect it.
• Do not read score as a probability of being right. It is closeness of text.

The predicate you found is not yet a question: we do not know which facts it needs. That is the next page — read the signature and the rule.