Most of the corporate repositories examined in the analysis were classified at the lowest level of a four-level scale for committed AI practice. Of 441 repositories, 294, or 66.7%, were Level 1. Another 109, or 24.7%, were Level 2, and 38, or 8.6%, were Level 3. None was classified at Level 4. In plain terms, the sample was dominated by repositories with no committed AI configuration, while more developed forms of the practice were less common.
A scale built from files in the repository
RAMP is a four-level cumulative model grounded in artifacts committed to a repository. Cumulative means the categories are designed to build on one another, so the scale records an ordered pattern rather than a single yes-or-no flag. To identify the relevant artifacts, the classifier combines filename-pattern heuristics with embedding-based semantic classification, which compares representations of files and categories by their meaning. The pipeline spans 12 AI coding tools, 1,046 validated artifacts and the 441-repository development frame.
That choice gives the study a concrete object to count: files and other repository artifacts that are actually committed. It also lets the researchers examine evidence of AI practice across multiple tools with one classification pipeline. The approach is therefore a repository-artifact measure, and the paper uses it as the basis for comparisons between repositories and across adoption patterns.
Two studies, one question about maturity
The authors paired two repository studies. Study 1 characterizes how AI practices are adopted over time, using a 196-repository temporal subset with reconstructed git histories. Study 2 applies RAMP to an existing agent-adoption panel and examines whether repositories at different maturity levels show different outcomes after adoption. Together, the studies connect the visible configuration landscape with the question of what happens after a coding agent enters the picture.
The corporate development frame contained 441 repositories from 27 organizations, but the first study's time-based analysis used only the 196-repository subset. That distinction matters because the broad repository frame and the historical analysis serve different purposes: one maps where repositories sit on the scale, while the other follows changes over time. The second study extends the framework to post-adoption comparisons rather than treating the initial corporate frame as the only evidence.
Because these are comparisons among repositories at different maturity levels, the design does not by itself establish that committed practices caused any later outcome. The study's contribution at this stage is the measurement framework and the pattern it makes visible, not a verdict on cause and effect.
The model held up in checks
Researchers tested whether the four stages behaved as a cumulative scale through Guttman validation, a formal check of whether the observed artifact patterns fit the proposed ordering. The reported coefficient of reproducibility was 0.997, and the coefficient of scalability was 0.983. A mapping-permutation test then compared the result with 10,000 rearranged assignments. The observed values landed at the 99.3rd and 99.5th null percentiles, with permutation p-values of 0.0074 and 0.0051.
A separate held-out check used human annotation of 195 artifacts from 35 repositories. The repository-level label matched the human result in 34 repositories, or 97.1%. At file level, agreement was 81.7%. The contrast is useful: the model reproduced the overall maturity label in nearly all of the held-out repositories, while individual artifacts produced more room for disagreement.
That validation supports using RAMP to organize committed repository evidence into maturity groups. It also sets the boundary of the result: the reported tests evaluate the scale's structure and its repository-level labels. They do not, on their own, say that a Level 2 or Level 3 repository has better configuration than another repository, or that the scale predicts outcomes beyond the comparisons described.
A baseline for future comparisons
The central finding is the spread of maturity in the corporate frame: Level 1 was the norm, Level 2 was a minority, Level 3 was smaller still, and Level 4 was absent. RAMP gives researchers a shared way to describe that spread from committed artifacts, while the paired studies provide a way to examine adoption dynamics and post-adoption outcomes in separate analyses. Whether the scale can explain those outcomes remains an open question.
Paper data and sources
Original title: A Few Pages of Markdown: Committed AI Configuration and Lower Quality Cost after Coding-Agent Adoption
Authors: Yegor Denisov-Blanch, Shyam Agarwal, Pavel Azaletskiy et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text