Open research, controlled release

Core principle

DEVIM is not intended to be a closed lab. We want people to see what we tried, what worked, what failed, and which questions remain open.

Openness does not mean publishing every research asset at once.

Open the evidence; do not blindly open the entire recipe.

The research narrative needed to understand and critique a result should be public whenever practical. Final blind evaluation sets, rights-limited corpus material, strategically material recipes, and unreleased novel mechanisms pass through a separate release decision.

Four release classes

• Public research: Findings, method summaries, limitations, negative results, and research notes.
• Public reproducible: Code or artifacts sufficient to reproduce a bounded result without exposing protected core assets.
• Gated release: Checkpoints, data, or tools distributed behind explicit access or license terms.
• Private core: Assets kept non-public for measurement integrity, rights, security, or strategic research reasons.

How others may use the work

Our default intent for public research text is CC BY 4.0 unless a publication states otherwise. People may share, adapt, teach from, and build on the material, including commercially, provided DEVIM and the relevant publication are credited.

That license applies to the published research text. Source code, model weights, datasets, evaluator artifacts, and other files use their own explicit release terms.

What remains controlled

Some research assets remain non-public when disclosure would weaken a blind evaluation, conflict with data rights, or disclose an experimental mechanism before a release decision has been made.

The purpose is not to hide the scientific result. It is to protect measurement quality, data rights, and original work while it is still being tested.

When a model is released

Model publication will be staged: research results and a model card can come first, followed by a gated research checkpoint where useful, and later an open-weight release if the evidence and release review justify it.

If access is restricted, we will not call the release open source. The label should accurately describe the access that actually exists.