I published my AI test generation platform to Zenodo. Not for clout — for a paper trail that survives the speed.
Every AI repo has the same origin story.
One person. One problem. One weekend. A working prototype that shouldn’t exist yet but does.
Then it compounds. A new framework gets added. The architecture gets split. A Human-in-the-Loop gate lands. An evaluation pipeline shows up. Before long, what started as a clever script is running in CI pipelines at companies you’ve never heard of.
That’s exactly what happened with ai-natural-language-tests.
And somewhere in that acceleration, most builders forget to do one thing: prove they built it first.
Every AI Repo Needs a DOI
The Problem With Building in Public at AI Speed
Open source in the AI era has a trust problem — not with the code, but with the clock.
When you ship fast, the gap between “I built this” and “anyone can verify I built this” gets very wide very quickly. GitHub timestamps exist, yes. But tags are mutable. Repos get renamed. Organisations disappear. The commit history is only as permanent as the platform hosting it.
In a space where ideas travel at the speed of a retweet and “inspired by” often quietly replaces “built by,” that gap is a real risk.
The fix is simple, free, and takes five minutes. Most builders still don’t do it.
What Actually Shipped in v5.0.0
Before the fix, a bit of context on what was worth protecting.
ai-natural-language-tests is an enterprise-grade platform that converts plain English requirements into production-ready end-to-end tests. Type a requirement. Point it at a URL. Get a working Cypress, Playwright, or WebdriverIO spec — generated by GPT-4, orchestrated by LangGraph, informed by a FAISS vector store that learns from every test you've ever generated.
Version 5.0.0, shipped May 3, 2026, brought three architectural changes that moved the platform from “capable tool” to “something you can actually operate at scale”:
The runtime became modular. What was one heavy file became three focused modules. qa_config owns framework definitions and LLM settings. qa_runtime owns observability, pattern store lifecycle, and failure formatting. qa_workflow owns the LangGraph graph itself. The entrypoint got thin. The codebase became maintainable.
A human approval gate landed. The --approve flag pauses generation after the AI produces a test, shows you what it wrote, and waits for sign-off before anything touches the filesystem. Fast teams leave it off. Compliance-sensitive teams make it mandatory. Same platform, different trust levels.
Evaluation scoring became structural. An offline NLP evaluator now scores every generated test using ROUGE and string similarity — no API key, no cost, blocks CI if quality drops below threshold. A Ragas evaluator goes further, scoring faithfulness, relevancy, and context recall against the live page HTML. The overall quality score ships with every test file.
This is the version that earned the permanent record.
Why Zenodo. Why Now.
Zenodo is a research data repository run by CERN. Scientists have used it for years to make their datasets citable. Software engineers are increasingly using it for something adjacent: making their releases permanently attributable.
When v5.0.0 was published there, three things happened automatically.
A DOI was minted — 10.5281/zenodo.20002379 — a globally unique, permanent identifier that resolves forever and doesn't depend on GitHub continuing to exist.
The full source tree was archived by Software Heritage, the world’s largest independent code preservation archive. The repo could vanish from GitHub tomorrow. The code doesn’t.
The record was indexed in OpenAIRE, making it discoverable by academic and research search engines worldwide.
That’s a citation-ready, independently verified, permanently preserved record of who built what and when. Not a star. Not a tag. A DOI.
The 5-Minute Tax Every AI Builder Should Pay
The setup is trivial.
Connect your GitHub repo to Zenodo at zenodo.org. Toggle the integration on for the repo you want tracked. Publish a GitHub Release as you normally would. Zenodo archives it automatically, mints a versioned DOI, and handles everything else.
No configuration. No cost. No ongoing maintenance.
The only thing it requires is actually tagging your releases — which you should be doing anyway.
In return: every release gets a permanent, citable, independently archived record. Researchers who want to build on your work can cite it properly. Teams adopting the platform can pin to a DOI instead of a mutable tag. Anyone who wants to know when a particular design decision appeared in the codebase has a verifiable answer.
The Broader Point
The AI tooling space rewards speed. Ship first, iterate fast, build in public. That’s the right approach.
But speed without a paper trail is how good work gets misattributed, rediscovered, and absorbed without credit. The projects that last — that get cited, built on, and remembered — are the ones with infrastructure behind them, not just a README and a star count.
A DOI takes five minutes. It protects work that took months.
That’s not bureaucracy. That’s just good engineering.
GitHub → github.com/aiqualitylab/ai-natural-language-tests
DOI → 10.5281/zenodo.20002379
Blog → aiqualityengineer.com
Top comments (0)