Projects · In preparation

French laws under scrutiny.

French law was built in successive layers. With tens of thousands of texts, nobody can reread everything to check that they remain consistent. This project hands that first reading to the machine, then measures what the texts cost and change.

Goal 1

Spot inconsistencies.

Review every law to find where texts contradict each other, repeat themselves or point to nothing.

Goal 2

Put figures on the impacts.

In a second phase, quantify the legislative and budget effects: how many texts a change touches, and what it means for public finances.

The approach

From PDF to analysis, in five steps.

  1. 01

    Gather the corpus

    Collect every existing law as PDF, from official public sources, with their dates and versions.

  2. 02

    Extract and split

    Turn each PDF into clean text, then split it into units that make legal sense: titles, chapters, articles, paragraphs.

  3. 03

    Vectorise

    Turn each article into a vector, a numerical fingerprint of its meaning, and store everything in a searchable vector database.

  4. 04

    Look for inconsistencies

    Bring together articles about the same thing, then spot contradictions, duplicates, references to repealed texts and diverging definitions.

  5. 05

    Quantify the impacts

    In a second phase, link the texts to public data to estimate their legislative effects (texts affected, dependencies) and budget effects.

Why vectorise

Search by meaning, not by words.

A classic search only finds texts that use the same terms. Two articles can deal with the same subject in different words, twenty years apart. Once vectorised, articles close in meaning end up next to each other even when their words differ: that is where inconsistencies hide.

What we want to spot

  • Contradiction

    Two articles in force that impose incompatible rules on the same subject.

  • Duplicate

    The same obligation written several times, with slightly different wording or thresholds.

  • Broken reference

    An article that refers to a text since amended or repealed.

  • Diverging definition

    The same notion defined in two ways depending on the text.

Points of attention

What the tool will not do on its own.

  • A lead, not legal advice. Every inconsistency the machine flags will have to be checked by a human reader.
  • The law in force keeps moving. Laws change constantly: the analysis will have to track versions and date every finding.
  • Imperfect PDFs. Layout, tables and appendices make text extraction harder, and the quality of the splitting drives everything else.
  • Figures to source. Budget impact estimates will rely on public data, cited for every result.

Progress

Section open, work ahead.

The project is starting. The technical choices (extraction tool, embedding model, vector database) will be made and described here as they come; non-public elements require private access.