Skip to content
Kulular Teknoloji, Kulular Bilişim Teknolojileri Limited Şirketi
TR

← Projects

KT / Dosya 01

Conflict

Evidence Gate for Text-Derived Improvement Claims

Status
Application preparation
Year
2026
1

Multilingual text

2

Classification

3

Evidence basis

4

Human decision

Complaint volume fell. There are two possible explanations. The problem was genuinely fixed, or the customer gave up on complaining. Both situations leave the same data signature and produce opposite conclusions.

No tool in the sector makes this distinction. That is not a technical gap, it is a commercial one. No supplier wants to tell its client that complaints fell because people stopped believing complaining would help.

Conflict is built to do exactly this. It extracts conflict patterns from text, measures the effect of an intervention, and decides whether an improvement claim clears an evidence threshold. Its answer is, more often than not, that it does not.

The project is framed not as a single software product, but as an audit protocol for text-derived corporate improvement claims, together with a reference implementation of that protocol.

Scope

  • Natural language processing
  • Multilingual model
  • Audit protocol
  • Multilingual text analysis
  • Evidence sufficiency
  • Causal effect measurement

Target programme

TÜBİTAK 1709 EUREKA Eurostars ↗
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

Mechanisms

The party being measured wants to distort the measurement. The design accounts for that.

  1. FIG. 01

    Separating exit from silence

    Problem
    A system that only measures complaints mistakes a customer quietly leaving for success.
    Design
    Complaint volume is read against transaction volume, not as a raw count. Whether a complaining customer keeps transacting afterwards is tracked. If silence continues while transactions stop, that is exit, not improvement.
  2. FIG. 02

    Pre-registration

    Problem
    Choosing a metric and an evaluation window after seeing the result is the easiest way to produce the answer you want.
    Design
    Before an intervention, the organisation declares the target pattern, the expected direction of change, the evaluation window and the primary metric. The declaration is timestamped and its summary is stored. Anything changed afterwards is visible in the evidence file. Pre-registration cannot be done retroactively.
  3. FIG. 03

    A recurrence measure

    Problem
    Closing a ticket counts as resolution. When the customer writes again, a new ticket opens, and both records show as resolved.
    Design
    Success is measured not by a fall in complaint count but by how long a resolution actually holds. The test is whether the same customer returns with the same problem within a defined period. The customer's return is outside the organisation's control and cannot be manipulated.
  4. FIG. 04

    An external anchor

    Problem
    Internal data verifies itself. An organisation's own ticket, with its own label, produces its own report.
    Design
    Unresolved consumer disputes are escalated externally. Applications to arbitration bodies, court cases and regulatory complaints are verifiable signals generated outside the organisation. If a pattern is genuinely resolved, escalation should fall too.
  5. FIG. 05

    A source of natural experiments

    Problem
    Measuring effect needs a real intervention with a recorded date and scope. Few organisations keep this record properly.
    Design
    A product release is, by definition, a complete record of an intervention. Its release date is fixed and its scope is written in the release notes. It provides the text flow of user comments before and after, including across languages.
  6. FIG. 06

    A catalogue of failure

    Problem
    Everyone talks about what worked. A catalogue of successes suffers from selection bias.
    Design
    Cycles that fail to clear the evidence gate are recorded too. Which intervention failed to produce an effect, in which sector, against which pattern, accumulates over time. This catalogue is far harder to copy than the model itself. This is where the real intellectual property sits.

Official project name: Çok Dilli Müşteri Şikâyeti Metinlerinde Çatışma Örüntüsü Tespiti, Müdahale Etkisinin Ölçümü ve İyileşme İddiaları için Kanıt Yeterliliği Değerlendirmesi Yapan Yazılım. This page is for presentation and is not application text.