How Lex is built, from top to bottom, in ten minutes: what it holds, how the data is cut, the indexes, the reference graph, the funnel that answers, how it is measured and how it is run. Each figure is read live from this site when the page opens, or comes from a named result file of the release that answers, with its date.
Lex holds the texts of the Luxembourg publisher (Legilux) and of the Publications Office of the European Union, each with its dated versions, and answers from them only.
…Luxembourg texts held
…of them in force, by the publisher’s data
…served as today’s law: today’s French wording is held
…in force, but held only in an older wording
…texts of the European Union, … of them served as today’s law
…old texts read from the OCR layer of a scanned page of the Mémorial (the publisher’s), and saying so
Of those …: … held only in their original wording, amended since (shown beside the texts that amended them, never as today’s wording); … whose newer version the publisher lists with no published file; … the publisher publishes only in German.
By kind of text
Kind
Held
In force
Served today
What Lex does not hold
Case law, the circulars of administrations and supervisors, parliamentary papers: when the reading or the choice of texts names one of them, code says that Lex holds no source of that kind.
Communal regulations (a commune’s police rules, for instance): not held, and the release that answers does not yet name them as such.
… of the … laws and … of the … grand-ducal regulations of the publisher’s catalogue are not held.
Today’s wording of … texts in force (above), among them the VAT law and the income-tax law.
… annexes of today’s versions exist only as PDF: they are counted on the page What Lex holds, and not read.
The shape of each part: typed, pile or case files
Three questions asked of each part of the corpus decide how it is read. The decision is written down, reversible, and stored with the data, one row per part.
One document type
The documents carry the same fields with one meaning: code filters on them (date asked, status, type) before the model picks the texts.
Pile
Documents with no field that says what they are about: they are reached through one routing line per text, then the text’s own table of contents.
Case files
The unit that answers is a bundle: a treaty and the law approving it, a directive and the law transposing it, an original text and the texts that amended it.
The content of the laws is not typed: no column says what a text is about, and the publisher’s subject codes stay out of the search. After the filter by code, the laws are therefore reached as a pile. These decisions were written after the build, not before it.
Technical detail: statuses, dated versions, the text check
By the publisher’s data: … Luxembourg texts in force, … repealed, … repealed in part, … no longer in force, … not yet in force, … whose status is not known.
Every one of the … Luxembourg versions is dated; … are open (still the law) and … have an end that is not known.
… of the … files pass the word-order check against the publisher’s file; the … failures are named.
Files of the European Union: … of … fail the word check after conversion; their texts are listed in the data file.
2
The data structure: from the text to the line
code
Each text is cut once, by code, into linked tables with stable identifiers: text, version, part of the table of contents, article, unit (alinéa, paragraph, point), line. No model takes part in this cut.
Text
One row per text, with the publisher’s fields: ELI address or CELEX number, kind, dates, status (in force, repealed…).
Version
Each dated wording. The dates are set by code, with no model, and each says where it comes from: the publisher’s record of that consolidation, the date in the file’s address, the entry into force, the next consolidation, the end of the text. An empty end means “still the law” or “end not known”, and the data says which.
Part of the table of contents
The publisher’s own headings (book, title, chapter, section), as they are: no table of contents is invented. This is the path the funnel walks down.
Article
The unit an answer cites: an article, an annex, text under a heading. Its identifier never changes; the same article in two versions is linked, and code knows since when its wording has not changed.
Unit
The alinéa, paragraph or point of the article, with its identifier. A reference to “paragraph two” is tied to its unit when code finds it.
Line
The text itself, numbered. The line numbers of an existing article never change: a citation points at lines, and code copies their words.
Links between texts
… links between texts read from the publisher’s data: amends, repeals, has as its legal basis, transposes…
Annexes
… annexes of today’s versions are held as text, … only as PDF.
Definitions
30,137 definitions found by code in the texts’ own words (« on entend par », « ci-après »…); 16,346 terms defined in today’s versions enter the keyword index.
Tables
… tables; header rows found by code in … of them. A row of a table keeps the line numbers of its header rows.
References
… reference links, from a line to an article: part four.
One real path, end to end
TextLoi du 27 juillet 1997 sur le contrat d'assurance: /eli/etat/leg/loi/1997/07/27/n10, kind LOI, in force since 07/09/1997.
VersionVersion read today: from 25/11/2025, no end (still the law); 148 articles. 10 versions held, the first from 25/12/2006.
Table of contentsTITRE I LE CONTRAT D'ASSURANCE EN GENERAL (75 articles) › Chapitre II - Dispositions communes à tous les contrats (41 articles) › Section IV - Exécution du contrat (12 articles)
ArticleArticle 21, “Défaut de paiement de la prime”: its wording is the same since 07/09/1997, identical in the 10 versions held.
UnitOne alinéa: art_21/a1. Elsewhere in the same section: Article 20 has an alinéa and points (art_20/a1, art_20/a1/pt1, art_20/a1/pt2, art_20/a1/pt3, art_20/a1/pt4); Article 26 has paragraphs and alinéas (art_26/p1/a1, art_26/p1/a2, art_26/p2/a1).
LineLine 1: A défaut de paiement d'une prime, ou d'une fraction de prime, dans les dix jours de son échéance et indépendamment du droit pour l'assureur de poursuivre l'exécution du contrat en justice, la garantie peut être suspendue à l'expiration d'un délai d'au moins trente jours suivant l'envoi au preneur d'assurance d'une lettre recommandée au dernier domicile connu.
Cited by Article 24, “Assurances sur la vie”: “ne sont pas applicables” (articles 21 à 23); a sure link. Part four follows it.
Read from this site on 22/11/2026 (GET /api/next/text-outline, /api/next/article-by, /api/next/context); the live read did not answer.
The tables of the store, and what never changes
act, version, current
the texts, their dated versions, and which version is today’s (and why not, when it is not)
node, outline
the headings of each version, as the publisher gives them
article, unit, line
the cited units, their alinéas and points, their numbered lines
ref, edge, edge_unit
every reference found in the text; one link per article aimed at, typed; the unit it is tied to
definition, annex, table_row
the definitions, the annexes (as text or PDF only), the rows of tables with their headers
the part of each text and the shape of each part; each column’s meaning, fill rate and rule for an empty value; the facts a reader must be told (OCR, German only…)
What never changes: an article’s identifier; the line numbers of an existing article; a row of the meaning index, which is only appended, recomputed or zeroed in place.
3
The indexes: four ways to find, each with its use
codemodel
The indexes decide nothing: each proposes candidates in its own column, and the model and code both see which index proposed what.
Keyword index
How
SQLite FTS5, ranked by BM25, accents folded, each word matched by its first letters (a prefix search: « suspendre » also finds « suspension »)
Size
… articles of the versions applicable today; 16,346 terms the texts define themselves
One row is
one article: its heading, its path in the table of contents, its text
Used for
the “words” detector inside the texts picked and over the whole store; the texts of the best articles at the corpus level; the counts of a proof of absence (in how many articles each word appears); the definition of a term of the question
Written by
code
Meaning index
How
…, … dimensions
Size
… rows for … units, 189,063 in use
One row is
the text’s title, the article’s number, heading and path, then its text: a first row of 3,000 characters, then stretches of about 2,500 characters along line boundaries
Used for
the “meaning” detector inside the texts picked and over the whole store; the texts of the closest articles at the corpus level. A main passage found by meaning alone is checked again before it is asserted.
Written by
code, which calls the embedding model; no language model writes a row
Left out
articles that only amend another text (…), texts that no longer apply (…), individual measures that name persons (…)
Routing lines
How
one line per text, and one per book or title of the large texts; each with its vector and in a keyword index
Size
… lines: … for whole texts (every text held has one; the line of a text that has left the store stays, ignored) and … for parts of large texts
One row is
what the text governs in legal terms, the same in plain words, what it does not deal with, English and German key words, its nature; its status comes from the publisher, by code
Used for
the corpus level: the 30 best lines by words and the 30 best by meaning, 24 of them shown to the model that picks the texts
Written by
a model (gpt-4.1-mini, prompt route-v1), from the title, the headings and the first lines of the text
Status
drafts: nobody has reviewed them
Dictionary of everyday words
How
a table read by code when the question is parsed
Size
… rows: … notions and … names of texts; …
One row is
a reader’s words in French, German and English, and the term the law uses
Used for
matched by code before any model; its legal terms feed a search of their own (a legal phrase rare in the store proposes the articles that hold it) and the proof of absence; for a question not in French they also join the searched words (weight 0.8)
Written by
a model or the code’s own lists, checked by code against the store
Status
drafts: no row is signed by a person
Each searched word carries a weight, which orders the tally of the article step (the keyword searches take the words as a set): the question’s own words 1.0 when the question is in French (less in another language, the store being French), the dictionary’s terms 0.8, the terms a model proposes 0.6.
Two other structures work as indexes: the table of contents (part two), which the funnel reads one level at a time, and the table of references (part four).
4
The reference graph
codemodel
Laws keep pointing at other articles and other laws. Code reads each reference in the text, resolves it to a text, an article, sometimes an alinéa, and types it by the words before it; the search uses it after the reading (one bounded hop out, the reverse index in), and its effect is measured.
From the written reference to a usable link
references found by code in the text (all versions)
1,371,994
links, one row per article aimed at
…
links that reach an article of the store
944,930
links tied to a unit (alinéa, paragraph, point)
…
References to another text, from today’s versions of the Luxembourg texts
Status
References
Share
resolved to the article
…
…
resolved to the text only (no article named, or a range)
…
…
text found, article not found in it
…
…
to a text Lex does not hold: named, never guessed
…
…
text not identified: kept as unresolved
…
…
The kind of a link, read by code in the words before the reference
Kind
Words that set it
Links
Role in an answer
exception
« par dérogation à », « sauf », « nonobstant », « à l’exception de »
…
at the article cited: an exception to its rule, found through the “cited by” index
sets aside
« ne sont pas applicables », « ne s’applique pas »
…
as for an exception
definition
« au sens de », « tel que défini à »
…
the definition is read with the rule
preserves
« sans préjudice de », « ne porte pas atteinte à »
…
the rule stays whole; shown as a condition
makes applicable
« s’appliquent », « sont applicables à »
…
a condition, followed when the reading flags it
pointer
« visé à », « prévu à », « conformément à », « en vertu de »
…
a condition, followed when the reading flags it
untyped
none of those words
…
a condition, followed when the reading flags it
When the reverse index proposes an exception, code reads the citing line again and the word nearest before the reference decides: « sans préjudice de » leaves the rule whole, « sous réserve de » yields to it, « par dérogation à » sets it aside.
… links inside the same text are marked sure and … to be checked; … links to another text are marked likely and … to be checked. Code checks each one before it is followed; the reverse index uses only the sure ones.
Around the article of part two
The links around the example article, as the store holds them:
Article 24 → Article 21: sets aside, “ne sont pas applicables” (articles 21 à 23), sure.
Article 21 → Article 1er: a term the text defines (“Prime”), pulled from the text’s own definitions.
How the search uses it
1
The reading flags pending references model
While reading, the model marks only the references that would change the answer if they were read (pending_references).
2
One bounded hop code
Before deciding that an answer is partial, code follows those references: 1 hop, at most 3 articles added. Each link is checked: the citing line names the target, the target has live lines; a reference to more than 5 articles is listed under “see also”, never followed. Another Luxembourg text is read in its version at the date asked; a text of the Union that Lex holds is followed; a text not held is named, never guessed.
3
A second reading model
The model reads again with the added articles, told why each is there (“cited by article …, alinéa …”).
4
The reverse “cited by” index code
After an answer, code looks for the articles that set the rule aside or depart from it: in law the exception is often written in the article that cites the rule, not in the rule. This is an extension for law. A term of the question that the text defines also pulls the text’s own definition.
What the references are worth, measured
On the … sealed questions, with the references switched off: … right instead of … (18 questions right only with them, 7 only without; p …). With the model’s choice of texts, it is one of the two levels whose effect is measured.
On the development set (where the builders tune, never shown as proof), measured on the earlier pipeline (the closest passages, before the funnel), its report half (112 questions, 28 needing a reference, mean of two runs): right 64.5 → 74.0 (noise 5); wrong asserted 13.0 → 14.5 (noise 2); reference questions with every needed article quoted 10.0 → 12.5 (noise 1). Two hops instead of one: 75.0 right, within the noise, so the budget stays one hop. Cost: 0.0016 USD more per question; the second reading starts on 15% of the questions. The funnel calls the same module, with its own second reading.
The precision of the resolver
31 right of 31 resolved references drawn at random once the resolver was frozen (6 rightly refused by the checks, 3 not served).
37 right of 38 hops actually followed on the development set; the error found (a range of articles read as one number) is refused by a code check since.
Each of these samples was marked by one reader, an AI model session, never by a person. Two earlier samples, drawn before the resolver’s fixes, had 28 right of 33 and 26 of 30; each error found became a code check. A sheet of 40 references drawn at random is ready for a person; it is not filled in yet.
Is this “graph RAG”?
Not in the usual sense. Lex holds a typed graph of references that the law writes itself and that code reads; the search walks it one hop, and only on a typed signal (the reading’s flag, the reverse index of exceptions, a term of the question that the text defines). It is not a knowledge graph of entities extracted by a model: no guessed entities or relations, no community summaries, no walk of several hops. Lex combines the structure (the table of contents), words, meaning and references; the evaluation measures each, switched off one at a time (part six).
What is not right yet
The typing of exceptions sometimes errs: « Sauf en cas d’application de l’article … » taken as an exception to that article; an entry-into-force clause shown as an exception; unrelated exceptions added to an answer. Seen by the final review; not corrected in the release that answers.
5
The funnel, on a real question
codemodel
Lex does not answer from “the k closest passages”: it picks the texts, walks down their tables of contents, has the candidates of several searches judged (each search proposes a bounded number), reads, follows the references, and code decides. How much is read comes out of the structure, inside written budgets.
Asked to the live site on 22/11/2026 (code t11, data 20261122c):
Je n'ai pas payé la prime de mon assurance habitation à l'échéance. L'assureur peut-il suspendre ma garantie, et à partir de quand ?
1
Reading the question
codemodel
Code reads the language (French), the date (no date in the question: today, 22/11/2026), the articles and texts named (none); the model gives the kind of question (a rule), a French rendering and legal terms; code recognises 4 rows of the dictionary (drafts).
2
Picking the texts
codemodel
93 candidate texts (95 entries, counting parts of large texts), proposed by the routing lines by words (30), the routing lines by meaning (30), the best articles by meaning (11), the best articles by words (15), the transposition of a directive (28); code removes 29 entries (repealed at the date asked (23), no longer in force at the date asked (5), not in the store (1)); 24 routing lines are shown to the model, which keeps 1 text: “La loi générale sur le contrat d’assurance peut contenir les règles relatives au défaut de paiement des primes et à la suspension de la garantie.”
What the trace holds at this level
Routing line
Proposed by (rank)
Question words in it
T1/eli/etat/leg/rgd/2010/02/23/n3
the routing lines by words #1
4
T2/eli/etat/leg/loi/1997/07/27/n10
the routing lines by meaning #25, the best articles by meaning #1, the best articles by words #1, the transposition of a directive #35
the routing lines by words #9, the routing lines by meaning #3
1
T7/eli/etat/leg/rgd/1999/05/05/n1
the routing lines by meaning #13, the best articles by meaning #3
2
T8celex:32015R0035 › titre_i__chapitre_iii
the best articles by words #3
3
T9/eli/etat/leg/code/civil › title_30
the routing lines by meaning #4
1
T10/eli/etat/leg/code/securite_sociale › n00773
the best articles by meaning #4
2
T11/eli/etat/leg/code/route › n00419
the best articles by meaning #5, the best articles by words #4
2
T12/eli/etat/leg/rgd/2003/11/11/n2
the best articles by words #5
3
T13/eli/etat/leg/loi/1937/07/09/n1
the routing lines by words #6, the routing lines by meaning #30, the transposition of a directive #27
4
T14/eli/etat/leg/loi/2020/07/10/a582
the routing lines by meaning #6
1
T15/eli/etat/leg/loi/2003/04/16/n2
the routing lines by words #15, the best articles by meaning #6, the best articles by words #12
1
T16/eli/etat/leg/loi/1994/12/08/n2
the best articles by words #6, the transposition of a directive #32
3
T17/eli/etat/leg/loi/2024/03/29/a137
the routing lines by words #7, the routing lines by meaning #8
4
T18/eli/etat/leg/rgd/2019/12/04/a824
the routing lines by meaning #7
0
T19/eli/etat/leg/loi/2023/08/07/a554
the best articles by meaning #7, the best articles by words #15
0
T20/eli/etat/leg/loi/1906/05/29/n4
the best articles by words #7
1
T21/eli/etat/leg/rgd/1969/03/07/n2
the routing lines by words #8
4
T22/eli/etat/leg/rgd/2009/06/25/n4
the best articles by meaning #8
3
T23celex:31991L0674
the best articles by words #8
3
T24/eli/etat/leg/loi/1873/09/04/n2
the routing lines by meaning #9
1
3
Walking down the table of contents
model
3 levels, one at a time: at each level the model sees the headings with their number of articles and the question’s words they hold, and picks at most 3 parts, each with a reason. It stops at a leaf; 24 articles lie under the parts kept.
What the trace holds at this level
Level 1: 5 parts shown, picked by the model.
Part picked
Words
The model’s reason
Stop
TITRE II DES ASSURANCES DE DOMMAGES
8
Ce titre traite des assurances de dommages, ce qui inclut l'assurance habitation, et contient les mots clés 'prime', 'assurance', 'échéance', 'assureur', 'suspendre' liés à la question.
continue
TITRE I LE CONTRAT D'ASSURANCE EN GENERAL
11
Ce titre général sur le contrat d'assurance contient des articles sur la prime, l'assurance, l'échéance et l'assureur, ce qui est pertinent pour comprendre les règles générales de suspension de garantie.
continue
TITRE III DES ASSURANCES DE PERSONNES
10
Ce titre concerne les assurances de personnes, moins pertinent pour l'assurance habitation, mais contient des mots clés liés à la prime et à l'échéance.
continue
Level 2: 12 parts shown, picked by the model.
Part picked
Words
The model’s reason
Stop
Chapitre II - Des contrats d'assurance de choses
7
Ce chapitre traite des contrats d'assurance de choses, incluant l'assurance habitation, et contient les mots clés 'prime', 'assurance', 'assureur', 'suspendre', 'garantie' liés à la question.
continue
Chapitre II - Dispositions communes à tous les contrats
11
Ce chapitre contient des dispositions communes à tous les contrats, incluant le paiement de la prime et l'échéance, ce qui est central pour la question sur la suspension de garantie.
continue
Chapitre IV - Des contrats d'assurance de la protection juridique
6
Ce chapitre mentionne 'prime', 'échéance', 'assureur', 'garantie', ce qui est pertinent pour la question sur la suspension de la garantie à l'échéance du paiement.
a leaf
Level 3: 13 parts shown, picked by the model.
Part picked
Words
The model’s reason
Stop
Section IV - Exécution du contrat
11
Cette section traite de l'exécution du contrat, incluant le paiement de la prime et les conséquences en cas de non-paiement, ce qui est central pour la suspension de la garantie.
a leaf
Section VIII - Prise d'effet de la garantie et durée du contrat
5
Cette section aborde la prise d'effet de la garantie et la durée du contrat, donc potentiellement le moment à partir duquel la garantie peut être suspendue.
a leaf
Section X - Prescription
5
Cette section mentionne explicitement la suspension et l'assureur, ce qui est pertinent pour la question de la suspension de la garantie.
a leaf
4
The detectors, side by side
code
33 candidate articles; each keeps its rank in each search: table of contents (24), meaning (6), meaning over the whole store (8), dictionary (8), words (6), words over the whole store (6). Skipped: the exact path (no article named in the question).
5
The arbiter
model
The model sees 18 candidates (the list is cut at 18) and gives each a role and a reason: 2 rules, 2 contexts, 2 conditions, 12 set aside.
What the trace holds at this level
Article
Role
The arbiter’s reason
Found by (rank)
Article 21
rule
Définit suspension garantie après non-paiement prime dix jours.
table of contents #12, words #5, meaning #1, words over the whole store #5, meaning over the whole store #1, dictionary #1
Article 25
rule
Précise effets suspension garantie et conditions primes ultérieures.
table of contents #16, words #1, meaning #2, words over the whole store #1, meaning over the whole store #2
Article 20
context
Modalités paiement prime, pas suspension garantie.
table of contents #11, meaning #6, meaning over the whole store #6, dictionary #2
Article 23
condition
Condition remise en vigueur garantie après paiement prime.
table of contents #14, words #2, meaning #4, words over the whole store #2, meaning over the whole store #4
Article 22
condition
Condition mise en demeure avant résiliation contrat.
table of contents #13, words #3, meaning #5, words over the whole store #3, meaning over the whole store #5
Article 105
set aside
Assurance vie, pas habitation.
words #4, meaning #3, words over the whole store #4, meaning over the whole store #3
Article 70
context
Créanciers privilégiés, pas suspension garantie.
words #6, words over the whole store #6
Article 45
set aside
Prescription, hors sujet suspension garantie.
table of contents #24, meaning over the whole store #7
Article 9
set aside
Convention collective, hors sujet.
dictionary #3
Article 93
set aside
Protection juridique, hors sujet.
table of contents #1
Article 93-1
set aside
Protection juridique, hors sujet.
table of contents #2
Article 94
set aside
Protection juridique, hors sujet.
table of contents #3
Article 95
set aside
Protection juridique, hors sujet.
table of contents #4
Article 95-1
set aside
Protection juridique, hors sujet.
table of contents #5
Article 96
set aside
Protection juridique, hors sujet.
table of contents #6
Article 97
set aside
Protection juridique, hors sujet.
table of contents #7
Article 97-1
set aside
Protection juridique, hors sujet.
table of contents #8
Article 18
set aside
Déchéance partielle, pas suspension garantie.
table of contents #9
6
Reading
model
The model reads 6 articles, an amount set by the part of the table of contents where the walk stopped, and points at 1 line; code copies the law’s words and refuses a line it does not know. Pending references flagged: none.
7
References
code
Trigger: the reverse index of exceptions.
Article 24, al. 1 écarte Article 21 (« ne sont pas applicables »)
« Prime » est défini par Article 1er de ce texte
8
The decision, by code
code
Rule answer_with_exception_by_code → state “answer with an exception”; 3 passages checked for their date and status, none dropped.
8 calls (7 to a language model, 1 to the embedding model), 0.0148 USD at list price, 10.8 s.
The answer Lex gave answer with an exception
The law’s words:la garantie peut être suspendue à l'expiration d'un délai d'au moins trente jours(Article 21)
Value read by code in those words: 30 jours (from “au moins trente jours”).
The model’s summary of the quoted lines, checked by code (every number is in those lines):Si la prime reste impayée dix jours après l’échéance, l’assureur peut suspendre la garantie au plus tôt trente jours après l’envoi d’une mise en demeure recommandée au dernier domicile connu.
ruleArticle 21Loi du 27 juillet 1997 sur le contrat d'assurance.
Wording in force from 25/11/2025, no end (still the law). why (model): A1 fixe les délais de suspension pour impayé et A4 précise le contenu obligatoire de la mise en demeure.
1 A défaut de paiement d'une prime, ou d'une fraction de prime, dans les dix jours de son échéance et indépendamment du droit pour l'assureur de poursuivre l'exécution du contrat en justice, la garantie peut être suspendue à l'expiration d'un délai d'au moins trente jours suivant l'envoi au preneur d'assurance d'une lettre recommandée au dernier domicile connu.
conditionArticle 22Loi du 27 juillet 1997 sur le contrat d'assurance.
Wording in force from 25/11/2025, no end (still the law). why (model): Condition mise en demeure avant résiliation contrat.
1 La lettre recommandée doit comporter mise en demeure du preneur de payer la prime échue, rappeler la date d'échéance et le montant de cette prime et indiquer les conséquences du défaut de paiement à l'expiration du délai visé ci-dessus.
2 L'assureur a le droit de résilier le contrat dix jours après l'expiration du délai visé ci-dessus.
exceptionArticle 24Loi du 27 juillet 1997 sur le contrat d'assurance.
Wording in force from 25/11/2025, no end (still the law). why (code): Article 24, al. 1 écarte Article 21 (« ne sont pas applicables »)
1 Les dispositions des articles 21 à 23 ne sont pas applicables aux assurances sur la vie.
definitionArticle 1erLoi du 27 juillet 1997 sur le contrat d'assurance.
Wording in force from 25/11/2025, no end (still the law). why (code): « Prime » est défini par Article 1er de ce texte
14 G. Prime: toute espèce de rémunération demandée par l'assureur en contrepartie de ses engagements.
The rule was found by several searches at once: table of contents #12, words #5, meaning #1, words over the whole store #5, meaning over the whole store #1, dictionary #1.
The arbiter set aside 12 of the 18 candidates, 9 of which only the walk had proposed: the walk also opened parts that do not fit.
No pending reference was flagged, so there was no outbound hop; the exception comes from the reverse index (Article 24, al. 1 écarte Article 21 (« ne sont pas applicables »)). Code does not judge whether an exception concerns the reader’s case: it shows it, with its reason and the law’s words.
Code read “30 jours” in the law’s words “au moins trente jours”: the value alone can lose words of the law, which is why the answer always shows the law’s words first. A known fault.
“How many articles to read?” is not a setting
The number read comes out of the structure: the walk stops at a part with no sub-part, at a part of at most 6 articles (16,000 characters), or at a list; the articles under the parts kept are read whole up to twice that size, a list up to 30 articles; past those sizes, an article step has the model pick articles from their headings, and a list too long is built from its headings. The reading sees at most 16 articles (30 for a list) and adds no article past 48,000 characters; a very long article is shown by windows around the lines that hold the question’s words. Here, 6 articles were read; on the 260 sealed questions, 9.0 on average.
What the walk is worth, measured: switched off, the score does not move (210 right against 208, p = 0.80); it costs 2.9 calls, 3.4 s (on a PC, four answers at a time) and 0.0029 USD per answer. What it brings: a second, independent line of evidence for the passages found by meaning, and a path a reader can follow. Part six.
6
Evaluation
code
The figures that count come from a sealed set of questions, written by others than the builders, used once, read by slice, with each brick switched off in turn and two points of comparison.
…right of … (…; 95% interval …–…)
…wrong answer asserted (…–…)
…the right article read, of the … questions that expect one
…missed: corpus gap …, search …, reading …, decision …
The same questions, every run
Right answers of …, with the 95% interval. Each brick switched off is the same engine, the same data, the same questions. An effect is said only where the exact McNemar test gives p < 0.05 against both full runs.
Run
Right
Right, with its 95% interval50%60%70%80%90%
Against the full run
By slice of questions
Slice
Questions
Right
95% interval
Run two
Previous Lex
Wrong
What the walk down the table of contents is worth
Switched off, the score does not move: 210 right against 208 (of 260); what the walk adds lies between −3.9 and +2.3 points, so nothing measurable on this set (it can only see an effect of 4.3 points or more). The walk’s own column holds the expected article 171 times of 222; all the other searches together, 200. When the right text was kept, it finds less in the large codes, where it should help most (34 of 44), than in the small texts (127 of 143). What it brings: a second, independent line of evidence (answers resting on meaning alone fall from 59 to 14) and a path a reader can follow. It costs 2.9 calls, 3.4 s (on a PC, four answers at a time) and 0.0029 USD per answer.
Development set and sealed set
Development set: 226 questions; the builders tune on it (the funnel and reference experiments); its figures are never shown as proof.
Sealed set: … questions, written by …, fixed before the run and used once; the protocol and the scorer were fixed before the runs. The runs ran locally on the same engine files and data (equal fingerprints), four answers at a time, not on the live revision.
A second scorer read the wrong answers and corrected the expected answer where the law justified it, for every run at once: the release went from 203 right and 5 wrong to 208 and 1; 40 right answers drawn at random were read again too.
30 new questions written on 22/11/2026 by a scorer independent of the builders (an AI model session), asked one at a time to the revision now live, before it took the public traffic: 29 right by the sealed set’s rules; 23 firm answers, none of them wrong.
Each model call has a role, a named model, a price and a time, in the trace of every answer; the server keeps no question and no answer, only, apart, each word of a question that no text holds (below).
The models, by role (read from the server)
…
Where the calls run
…
What the server keeps
…
Cost and time per answer
Measured on
USD per answer
Model calls
Median time
Nine in ten under
The sealed set, on a PC, four answers at a time
…
…
…
…
New questions, one at a time, on the release now live
0.021
7.9
10.4 s
13.0 s
The question of part five
0.0148
7
10.8 s
—
Replaying an answer
An answer’s link carries a token signed by the server (HMAC-SHA256); the server draws the answer again with no model call, and refuses a token it did not sign. Measured on the answer of part five: no model call. It redraws, it does not reproduce: nothing is logged, by choice, and the models do not answer twice the same way.
The engine measured is the engine that answers
The server publishes the fingerprint of the files that compute an answer: …. The fingerprint of the engine measured on the sealed set: ….
What no person has checked yet
The dictionary: … rows, ….
The routing lines: …, all written by a model, reviewed by nobody.
The precision of the references: marked by a model, never by a person.
The shape decisions of the corpus: written after the build, by a builder.
Three review sheets are ready since 22/11/2026: the dictionary (86 rows), the routing lines of 40 texts chosen for an insurer, insurance first (148 lines), 40 references drawn at random. None is filled in; the script that applies them is not part of the release that answers.
8
What I would do next
Have an expert sign the dictionary and the insurance routing lines, and have a person mark the references drawn at random: the sheets are ready.
Type exceptions by a rule on the nearest words (« sauf en cas d’application de » is not an exception), and stop adding unrelated exceptions.
Never cut the law’s words in a typed value (« trois mois et quarante jours », « majoré de quarante pour cent »).
Name what Lex does not hold when a question aims at it (communal regulations), and hold today’s wording of the VAT and income-tax laws.
Make the walk earn its place: run it only on large texts or when the searches disagree, rank its finds into the arbiter’s list instead of after the others, or measure it on a new set fixed in advance, rich in large codes and lists.
Add a grouping level above the routing lines, and publish how often the choice of texts keeps the right one on the sealed set.
A log the owner chooses, so that an answer can be reproduced and not only redrawn.
A candidate release already carries fixes for the exceptions, the typed values and the texts named, and data with today’s VAT and income-tax wordings (the tax administrations’ coordinated texts); it has passed only a smoke test of a few questions, has not been measured on a question set, and does not answer.
Where each figure comes from
Read live when the page opened:
GET /api/next/corpus (reading…)
GET /api/next/config (reading…)
GET /api/next/coverage (reading…)
GET /api/next/results (reading…)
GET /api/next/shapes (reading…)
GET /api/next/dictionary (reading…)
GET /api/next/text-outline?work=loi-1997-07-27-n10 (reading…)
GET /api/next/article-by?work=loi-1997-07-27-n10&anchor=art_21 (reading…)
From the result files of the release that answers (code t11, data 20261122c, engine 01f1b08c054c474c):
The final counts of the data: build-log\repair-data-final-counts.out, 22/11/2026.
The reference experiment and the precision samples: build-log\refs.md, experiments\refs\tables-R.md, experiments\refs\REPORT-from-reply.md, 21/11/2026.
The value of the walk, from the saved sealed runs: experiments\walk\WALK-VALUE.md, experiments\walk\walk_metrics.json, 22/11/2026.
The sealed set, re-checked: experiments\sealed\tables-checked.md, 22/11/2026.
The new questions of the final review: VERDICT-FINAL.md, 22/11/2026.
The budgets written in the code: app\funnel.py, app\refs.py, app\build_routing.py, app\embed_index.py, app\build_lexical.py (git tag rc3), 22/11/2026.
The example path (used when the live read does not answer): GET /api/next/text-outline, /api/next/article-by, /api/next/context, 22/11/2026.
The question of part five: POST /api/next/ask, POST /api/next/replay, 22/11/2026.
The review sheets: review\DICTIONARY-REVIEW.md, ROUTING-REVIEW.md, REFERENCES-REVIEW.md (review-manifest.json), 22/11/2026.