lumen

Collaborate on Lumen

Lumen serves one problem family: unsafe and unaccountable AI. This page lists concrete, scoped ways an organization can work with the project. Every task here is real and completable. Nothing on this page implies an existing partnership; collaborations are invitations until they actually happen, and no partner is named publicly without their written agreement.

Who this is for

Any institution deploying AI under trust obligations (health, education, faith, and public sector bodies), plus AI vendors whose systems need refusal behavior they can demonstrate rather than assert.

Sized tasks

1. Adopt the guardrail policy layer. Take the refuser pattern (a runtime output check, independent of prompts) and the case set format, and apply them to your own domain’s non-negotiables: the outputs your system must never produce, whatever a user asks or a model drifts into. We want your deployment notes, honest ones, published as a case study with your approval.
2. Contribute refusal test cases. Add cases to packages/catechism-rag/demo/eval_cases.json from your domain: outputs that must be blocked, outputs that must be allowed, and the borderline quotations where your institution has to choose a tradeoff. Borderline cases are reported, not scored, and the reasoning is the contribution.
3. Publish your eval results against the shared refusal suite. Run the suite against your own guardrail implementation and publish the artifact next to ours. Suites improve when more than one implementation is measured against them.

What you get

A public, dated record in this repository: the case study, the contributed cases with the eval rerun, or your published results, with your organization’s role stated as you approve it.

Start

Open an issue in this repository describing the task you have in mind and the trust obligation behind it. Small and specific beats grand and vague.

This site is open source. Improve this page.