AI-Native Development Frameworks: A Comparison by AI Assessors
Year: 2026
Type: preprint
Abstract: This study examines how three independent model-generated assessments describe the documented coverage of different AI frameworks for software-engineering and delivery. One framework was noted as to be included in the Prompt (IntDEx v5) and the other frameworks were picked by the models on what they consider as more relevant and mature at time, according to the prompt instructions. In the prompt, the guideline to not favor any of the models and especially IntDEx (to the extent that IntDEx assessment may be considered stricter to the other frameworks) was provided. The selected models for the evaluation have been chosen based on what could be considered state-of-the-art at time, and are from three different model families to increase independency. The research material includes an evaluation prompt and three assessment records, retained unchanged in Annex A DATA. A descriptive secondary analysis extracted comparison selections, assessment criteria, scores on an ordered scale, evidence statements and stated limitations, and checked the arithmetic of the reported totals. IntDEx ranked first within all three assessments, with reported scores of 67/80 (83.75%), 53/80 (66.25%) and 63/75 (84%). The AI experts (assessors) selected different frameworks to compare and used different criteria and scoring standards. This is a design choice result. Their totals therefore do not form a common benchmark. The author deliberatively did not intervene in any phase of the AI expert’s assessment. The records match (for IntDEx) comparatively strong treatment of human accountability, the origin and traceability of evidence and delivery-engine governance, while identifying limited empirical validation. The broader delivery scoring additionally identifies weaker IntDEx product discovery, application architecture and application operations. These findings describe the AI expert judgments rather than independently verified framework capabilities or demonstrated effectiveness. The main contribution is a transparent account of agreement, differences in assessment methods and limits of the evidence in model-generated comparative assessment. A common more detailed evaluation framework registered in advance, versioned references and human review of each AI assessor’s results are proposed as requirements for a later validation study.
Source: Figshare
Author: SPYRIDON KTENAS
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225460343