Skip to content
Kulular Teknoloji, Kulular Bilişim Teknolojileri Limited Şirketi

← Projects

KT / Dosya 02

Casuist

Evidence-Grounded Legal Intelligence with Verifiable Sourcing

Status
Seeking a partner
Year
2027
1

Question

2

Retrieval

3

Proposition binding

4

Support check

Legal language models have solved fluency without solving verifiability. In practice the binding constraint is not the quality of the prose but whether each proposition can be traced to a source in the correct jurisdiction, still in force, and correctly scoped.

The real danger is not the obvious fabrication. A plausible citation to a decision that genuinely exists but is not controlling on the point is the error a busy practitioner is least likely to catch. The system fails silently and the user does not notice.

That cost lands directly on the lawyer. The time the tool promised is largely handed back through the need to verify every line. Output that cannot be verified is unusable anyway under professional liability and duty-of-care rules.

Casuist is not an adjudication system. It is a grounding and provenance layer over legal corpora. Every generated proposition is bound to a specific span of the source it rests on, a statement without adequate support is flagged rather than completed, and the provenance chain can be opened and inspected by the user.

Abstention is a designed behaviour. Where there is no support, the system does not write the sentence. The final judgment stays with the professional in every case.

Scope

  • Legal technology
  • Evidence-grounded generation
  • Citation fidelity
  • Provenance
  • Auditable AI
  • Decision support

Target programme

TÜBİTAK 1709 EUREKA Eurostars, Çağrı 12 ↗
TÜBİTAKEUREKAEurostars

The project is at the application-preparation stage for this programme. These marks show the targeted support programme, not support that has been awarded.

Lead

Doç. Dr. Merve Ayşegül Kulular İbrahim

Ask about this project

Open call for a partner

Deadline 4 March 2027

We are looking for a partner to co-lead the evaluation, not a data supplier.

Türkiye is a full Eurostars member state; national funding is administered by TÜBİTAK under the 1709 programme. Call 12 opens for submission on 17 December 2026 and closes on 4 March 2027 at 14:00 Brussels time. Some countries have no allocated budget for this cut-off, so a partner should confirm its own national funding status early.

The partner leads

  1. 01Legal requirements, jurisdictional scoping and design of the evaluation framework
  2. 02Benchmark corpus construction and gold-standard annotation
  3. 03Validation, benchmarking and error analysis

What we are looking for

  • A research organisation, university law faculty or legal-data specialist holding, or able to lawfully obtain, a machine-readable case-law and legislation corpus in at least one European jurisdiction
  • Expert-jurist capacity for annotation work that can be named and committed contractually, not only doctoral students
  • At least one legal NLP benchmark, annotated dataset or empirical legal study with a published methodology
  • The ability to turn the distinction between cited and controlling into an operational scoring rule
  • Confirmed national funding eligibility for Call 12, or a written self-funding decision

Intellectual property

Background IP remains with its owner. Foreground IP in the core platform software remains with Kulular Teknoloji. In return, the evaluation methodology, the benchmark dataset and the resulting publications are designed as joint property; the partner holds research and teaching rights in its own field and a defined commercial route. This boundary is negotiated before drafting begins, not after.

How to get in touch

Write to us with a one-page capability note and your national funding status for Call 12. A thirty-minute introductory call, then a two-page joint concept note, then a go or no-go decision before drafting starts.

Write to us →

Mechanisms

A fluent answer is easy. Showing what the answer rests on is the hard part.

  1. FIG. 01

    Proposition-level source binding

    Problem
    A bibliography at the end of the text does not show which sentence came from which source. The verification burden stays with the user from beginning to end.
    Design
    The binding is made to individual propositions, not to the text as a whole. Every proposition is bound to a specific span of its source, and that span can be seen without leaving the interface.
  2. FIG. 02

    Controlling against merely cited

    Problem
    That a decision exists does not mean it is controlling on the point. Systems treat the two as equivalent, and the most dangerous error is born exactly there.
    Design
    A source is not merely retrieved; it is assessed for jurisdiction, whether it is still in force, and scope. A non-controlling citation is flagged separately.
  3. FIG. 03

    Abstention

    Problem
    A model prefers to complete the sentence when support is weak. It fills the gap by inference and does not make that visible.
    Design
    Where the support threshold is not met, no output is produced and the point is marked unsupported. An openly left gap instead of a silent guess.
  4. FIG. 04

    Independent evaluation

    Problem
    Performance measured by the team that built the system tells neither an evaluator nor a customer anything.
    Design
    Citation fidelity and error taxonomy are measured against an expert-jurist gold standard, by a team that does not build the core software. Independence here is a design decision, not a shortcoming.

Official project name: Hukuki Metin Üretiminde Önerme Düzeyinde Kaynak Bağlama, Dayanak Yeterliliği Denetimi ve Doğrulanabilir Atıf Üreten Yazılım. This page is for presentation and is not application text.