Working group NIKOLAI · AI safety
NIKOLAI: Safety cases and evaluations
The NIKOLAI safety cases and evaluations working group reviews the proposed vocabulary connecting evaluation evidence to a go/no-go decision: the claims a safety case makes, the risk level and residual-risk determination it reaches, and the evaluations, elicitation methods and validity threats its evidence rests on. It reviews the elements in NIKOLAI tracks N4–N5, tests CASRAI's shadow crosswalks against the published frameworks and evaluation reports they cite, and proposes corrections, gaps and new elements for the next NIKOLAI release.
Charter and scope
What this working group covers
The NIKOLAI safety cases and evaluations working group reviews the proposed vocabulary connecting evaluation evidence to a go/no-go decision: the claims a safety case makes, the risk level and residual-risk determination it reaches, and the evaluations, elicitation methods and validity threats its evidence rests on. It reviews the elements in NIKOLAI tracks N4–N5, tests CASRAI's shadow crosswalks against the published frameworks and evaluation reports they cite, and proposes corrections, gaps and new elements for the next NIKOLAI release.
- N4 Claims and argument: claim, safety case, likelihood term, risk level, confidence adjustment, residual risk and risk acceptance determination, marginal versus absolute risk basis, limitation
- N5 Evidence and evaluations: evaluation, evaluation run, elicitation method, saturation status, evaluation-validity threat
- How a safety case's claims reference the evaluation evidence defined in N5
- Accuracy of CASRAI's shadow crosswalk rows for these elements against the documents they cite
- Concepts used in published safety-case and evaluation methodology that NIKOLAI does not yet define
Because members may work for competing AI developers, CASRAI's antitrust guidelines (POL-02) and conflict-of-interest disclosure apply to every meeting and deliverable, and the group discusses published documents and vocabulary only — never pricing, products, release plans or other competitively sensitive information.
NIKOLAI is a CASRAI proposal (nikolai-v0.2). Every crosswalk row is CASRAI's own shadow mapping unless the organisation concerned files a Mapping Declaration, and no AI developer, evaluator or regulator has endorsed or joined NIKOLAI or this working group.
Composition
Current composition
Chair
Applications under review
Candidates are being assessed for the 2026 forming round. Two-year term, renewable once.
Submit your candidacy →Co-chair
Accepting candidates
Optional second chair, prioritised for regional or sector balance.
Submit your candidacy →Community seats
Accepting candidates for 12 seats
Seats are allocated against a published rubric covering domain expertise, institutional diversity, regional balance, and an explicit slot for early-career researchers.
Apply for a community seat →Deliverables
Planned and delivered deliverables
- Planned: a reviewed set of N4–N5 element definitions, with corrections recorded in the NIKOLAI changelog
- Planned: a worked example showing how N4 claims reference N5 evaluation evidence
- Planned: a check of every N4–N5 shadow crosswalk row against the document it cites
- Planned: recommendations for the next NIKOLAI minor release
Recent activity
Working-group cadence and milestones
Illustrative working-group cadence for the 2026 forming round. Substantive deliverables and meeting minutes will be linked here as the group convenes.
2026-10 · Milestone
Working group forming — applications open
Open call for chair, co-chair and community seats on the NIKOLAI: Safety cases and evaluations working group. Applications are under review.
2026-09-19 · Release
NIKOLAI nikolai-v0.2 published
Track N10 (assurance roles) and the Mapping Declaration process added. See the NIKOLAI changelog.
2026-09-18 · Added
NIKOLAI proposal published
CASRAI published nikolai-v0.1, its proposed frontier-AI-safety vocabulary, with CASRAI's own shadow crosswalks.
Open consultations
Currently in public comment
Currently none open. Comments on any NIKOLAI element are welcome at any time via /get-involved/comment.
External bodies
Published documents the group reads
Bodies whose published documents this group reads and cites. Listing is not a partnership: none of them has endorsed, joined or reviewed NIKOLAI or this working group.
Going deeper on CASRAI
Related CASRAI guidance
Adjacent working groups
Related working groups
How to join
Apply to a seat on this working group
This working group is receiving candidates for the 2026 forming round. The application is short — name, institution, ORCID iD (optional), the seat you are applying for, and a paragraph on why this group. Decisions are returned within four weeks of the close of the open-call window.
FAQ
Frequently asked
- What does this working group do?
The NIKOLAI: Safety cases and evaluations working group reviews the proposed NIKOLAI elements in tracks N4, N5: it checks their definitions and CASRAI's shadow crosswalk rows against the published documents they cite, and proposes corrections, gaps and new elements for the next NIKOLAI release. NIKOLAI is a CASRAI proposal (nikolai-v0.2); no AI developer, evaluator or regulator has endorsed or joined it.
- Who can apply to join?
Membership is open to qualified practitioners in a personal capacity — AI safety and evaluation researchers, policy and standards staff, and people working at AI developers, evaluators, regulators, universities or civil-society organisations. Members may work for competing AI developers, so CASRAI's antitrust guidelines (POL-02) and conflict-of-interest disclosure apply. Apply via the working-group application form.
- What is the time commitment?
Cadence: Monthly virtual sessions while forming; vocabulary work released through NIKOLAI minor releases under the published versioning policy.. Time commitment averages four hours per quarter for an active member, more for a chair. Asynchronous review happens between meetings via email. Chairs are recognised on the editorial masthead and serve a two-year term, renewable once.
- How are seats allocated?
Each working group has 12 community seats plus a chair and (optionally) a co-chair. Seats are filled by an open call reviewed against a published rubric: domain expertise, institutional diversity, regional balance, and an explicit slot for early-career researchers. Applications for the 2026 forming round are open and candidates are currently under review — see how to apply.
- Can my institution sponsor a seat?
Institutional sponsorship is welcome but is not a route to a guaranteed seat. Sponsors support meeting infrastructure, public-comment publication and contributor honoraria; they receive named acknowledgement on the working-group page and in release notes. Sponsorship enquiries go to [email protected]. Seat selection remains on merit.
