How NoDelulu Works

What NoDelulu does

NoDelulu analyses text for hallucinations, factual errors, fabricated sources, logical inconsistencies, and other reliability issues commonly found in AI-generated content. It uses a sequential adversarial verification process:

1. The Sweep — The first AI model reads your entire document and flags every potential issue it finds across 8 hallucination categories.
2. The Challenge — A second AI model analyses the same document independently, forming its own view. Only then does it see the first model's findings — and challenges every one. Findings both models agree on carry strong confidence. Findings the second model disputes carry less. This adversarial debate is what separates NoDelulu from simple multi-model voting.
3. Web Grounding — Factual, numerical, citation, and temporal findings are verified against the live web. The models never have web access during their analysis — grounding happens strictly after, so the adversarial review stays independent. Analytical findings (logic, opinion, contradiction, omission) stand on the adversarial review alone, because no search engine can tell you whether a conclusion follows from its premises.
4. The Report — A synthesis model translates the adversarial findings and web evidence into a structured report with your NoDelulu Index, individually scored findings, clickable sources, and a correction prompt you can paste back into your AI.

The 8 types of AI hallucination

Not all errors are the same. A hallucinated statistic is a different kind of failure from a logical contradiction or a misleading framing. NoDelulu classifies findings across 8 distinct categories:

1. Factual DeLulu — Claims that contradict established facts
2. Number DeLulu — Wrong dates, statistics, measurements, or quantities
3. Made Up DeLulu — Citations, URLs, or references that don't exist
4. Self-Contradiction — The text says one thing, then contradicts it
5. Logical Leap — Conclusions that don't follow from the premises
6. Opinion As Fact — Subjective claims presented as objective truth
7. Time/Date DeLulu — Anachronisms, wrong timelines, or outdated information presented as current
8. Missing Context — Critical context that changes the meaning is left out

The first 4 categories are checked against the live web. The last 4 are structural — about the text's own logic and internal consistency — and are evaluated through the adversarial model debate.

How the report is generated

NoDelulu's findings are generated by AI models. While the adversarial approach and web verification significantly improve accuracy compared to a single model reviewing its own work, the system is not infallible.

Like all AI-powered tools, NoDelulu's analysis may occasionally:

• Flag something as incorrect when it is actually correct (false positive)
• Miss a genuine error (false negative)
• Misinterpret context, nuance, or domain-specific language
• Provide source links that don't fully support the stated finding

What you should do

NoDelulu is designed to assist your review process, not replace it. We recommend:

• Reviewing each finding and using the provided source links to verify claims independently
• Applying your own judgment, especially for domain-specific or subjective content
• Treating the confidence score as a guide, not an absolute measure

Why adversarial verification matters

When you ask a single AI to check its own work, you are asking the same system — with the same biases and blind spots — to evaluate itself. Research shows this produces confirmation bias. NoDelulu's adversarial design uses a different approach: independent models that challenge each other, with web evidence as an objective tiebreaker. The architecture maps to research in developmental psychology (peer accountability), epistemology (external grounding), and safety engineering (redundant independent verification).

For the full research foundations and methodology, visit How NoDelulu Works — Methodology. For the science behind AI hallucination and why it matters, visit The Science.

Feedback

If you encounter a finding that seems incorrect or if the system misses something important, we want to know. Contact contact@nodelulu.ai.