Most of the interesting work happens before the launch announcement. A rule is written. A test fails for the right reason. A beautiful plan meets a real system and needs to change.
Odin R&D is a place to show that work while it is happening.
The factory is the research subject
When agents can produce code, we still need to decide what a good system looks like. We need ways to check the result, understand the failures and improve the next attempt.
That is the thread connecting this workshop: intent, construction, verification and learning. The factory-floor drawing gives those ideas a place you can explore. It is a conceptual assembly drawing, not a live map of our infrastructure.
Build the tool. Open the experiment. Keep the record.
Start with something you can run
The first machine on the public floor is the Blueprint Conformance Engine. Its released package checks code against authored architectural boundaries. Humans own the intent; agents can use the same check while they work.
The test bench runs three packaged experiments: TypeScript module boundaries, Python module boundaries and configuration widening. Each compares a conforming fixture with a deliberately drifted fixture. Both go through the same authored rule.
The website publishes the emitted output, engine version, execution time and source revision. GitHub Actions reruns the experiments on changes, on request and weekly. If the expected discrimination fails, the workflow retains the result and does not publish a replacement site. The previous site keeps its original observation timestamp.
These are small, author-controlled mechanism experiments. They show whether a specific check tells the two fixtures apart. They do not establish that agents become faster, cheaper or more successful. That is a different research question and deserves its own study.
Build with us, privately
Factory Intelligence asks whether we can understand the quality and cost of an outcome well enough to improve the next attempt. Completion alone is too crude a measure: work can be finished without being useful.
Odin Gym asks whether real work can become an isolated learning environment, with a clear task and a verifiable outcome. Practice is most useful when the result means something.
Factory Intelligence and Odin Gym are developed for Odin’s internal use, private customers and early adopter partners. Bring a workflow, a quality challenge or an agent-learning use case: together we can discuss the scope, evaluation criteria and early access that fit your team. Talk to Odin about a private collaboration.
An invitation to make the work better
You can read the code, repeat an experiment, report a different result or propose a new question. The site itself is an open-source project, including the factory drawing and the experiment runner.
Our field notes will connect the technical detail to its meaning: what changed, who it helps, what we learned and what remains open. Over time, that collection should earn a useful introduction to the wider Odin concept.
Bring a question to the workbench.
This site develops Odin’s August 2026 dark-factory and blueprint-systems proposals into a public workshop. Historical proposals supply the direction; current public releases and dated experiment records supply the claims. The first site edition was issued on 7 September 2026.