SlopOne: Winning the code quality fight with Jev
Software teams are drowning in AI code slop. We’ve developed a new, superior way to prevent it and fix it.
Today, we are releasing SlopOne as a free research preview. SlopOne is an AI-powered model to differentiate quality, maintainable code from slop that erodes codebases over time. In our testing, SlopOne outperformed every available algorithm in terms of its ability to accurately predict the maintainability of code at the level of expert human software engineers.
We achieved this result by training a hybrid model that combines statistics from traditional static analysis with classifications from Typesafe AI’s new Jev decision model. These data sources work together to boost precision and recall to a level that neither can achieve on their own.
Along with SlopOne, we’re also releasing a new code quality trend report experience that visualizes SlopOne maintainability over time, and makes it easier than ever to kick off an AI-lead refactoring effort. All of this is available today for free in the Qlty CLI.
A maintainability score with intelligence, courtesy of Jev
Automated maintainability scores have historically come from static analysis tools, which can understand the structure of code but not its meaning. For example, for years Qlty has calculated scores based on statistics like function complexity, deeply nested conditional logic, and similar or identical blocks of code.
By contrast, SlopOne goes further and combines signals from traditional static analysis with decisions from Jev. Jev is a new System One model from TypeSafe AI which makes intelligent decisions quickly and with low cost. For example, we ask Jev questions like (shortened for clarity):
- Would an experienced developer consider this code difficult to maintain?
- Would a developer need to keep many interacting conditions or mutable values in mind to understand it?
- Does this code keep the same information in more than one place, where it must be kept consistent by hand?
Jev makes a reviewer’s judgments cheap and fast enough to generate for every file and feed to an ML model. We combined answers like these with static analysis statistics and used them as inputs (features) to train a machine-learning model trained on human-labeled code quality data.
As far as we know, nobody has scored maintainability this way before. Our internal benchmarks show SlopOne outperforming all other maintainability scoring systems we have tested it against. We believe this is the best maintainability score available today, and we’ll be sharing more about this in a follow up.
We’ve been using SlopOne internally, and it’s noticeably better than anything we’ve had before. The improved accuracy and precision makes it an excellent input for automated refactoring loops.
Point, click, refactor
Once you’ve perused the report, you can act on those findings right away. Select the files to clean up, click “Copy refactor prompt”, and paste the resulting prompt into your favorite coding agent.
The prompt combines everything your agent needs: which files you selected, the findings from both static analysis and Jev’s answers, and SlopOne’s weighting of which problems pull the score down the most.
Because Jev’s answers are based on a semantic reading of the code, the prompt gives your coding agent better hints on its refactoring task. Give it a try and tell us how the refactors come out.
Get code quality trends and start auto-refactoring with SlopOne
Using SlopOne is free. You bring your own API key for Jev (we support Typesafe, Vercel, and OpenRouter). The cost of running a historical SlopOne analysis is typically less than $1 in Jev usage credits.
curl https://qlty.sh | bash
export TYPESAFE_API_KEY="apikey_..."
qlty slop-oneIf you need an API key, you can sign up for TypeSafe AI.
We’ll publish the methodology and our measurements soon. Until then, run SlopOne on your own code and tell us where the scores match your sense of it and where they don’t, in the SlopOne discussion on GitHub: github.com/orgs/qltysh/discussions/2871
