NIKOLAI
NIKOLAI is the common language for frontier-AI safety reporting: one definition and one identifier for every element a safety framework, risk report, or evaluation can contain, crosswalked to how major labs, evaluators and governments each use these concepts — so commitments can be compared, not just read.
What NIKOLAI is
An open, versioned, machine-readable Dictionary of Elements for frontier-AI safety documentation, published by CASRAI under CC-BY 4.0. Each element has a working definition, a stable identifier, and — where the research corpus supports it — a crosswalk showing how named organisations' own published documents use the concept.
What NIKOLAI is not
NIKOLAI is not an evaluator, an auditor, a regulator, or a lab consortium. It runs no evaluations, holds no model weights, and receives no pre-release access. It does not certify that any lab is safe, and it does not issue “compliant” or “frontier-class” designations.
Every crosswalk here is a shadow mapping
No lab, evaluator or regulator has endorsed NIKOLAI, reviewed a mapping, or been consulted on this dictionary. Where an element page shows how “Anthropic” or “the EU AI Act” uses a term, that reading is CASRAI's own — drawn from that organisation's published documents, never from a conversation with them. This is labelled on every single crosswalk row, not only here. It stops being a shadow mapping for a given organisation only when that organisation files its own Mapping Declaration — a mechanism this dictionary does not yet expose publicly.
Where the vocabulary comes from
The published elements are drawn from a research pass reading primary-source safety frameworks and government instruments directly — Anthropic's Responsible Scaling Policy and Risk Report, OpenAI's Preparedness Framework, Google DeepMind's Frontier Safety Framework, xAI's Frontier AI Framework, Meta's Advanced AI Scaling Framework, California SB 53, the EU GPAI Code of Practice, and evaluator and standards-body material from METR and the Frontier Model Forum. Every citation on an element page resolves to a real published document.
Look up a specific concept
Common entry points into the dictionary:
- Safety case — the structured argument that a model's risk is acceptably low
- Capability threshold — the point at which a model requires new safeguards
- AI incident reporting — deadlines, recipients, and discovery method
- AI model cards & scope — developer, model identifier, deployment surface
- Evaluator independence — conflict-of-interest terms for third-party review
- Redaction — what a developer may withhold from a published safety report
The tracks
Elements are being published across nine active tracks.
- N1 · Actors, models and scope
Developer, model identifier, deployment surface, coverage date, safety framework version, coverage scope threshold, severity threshold, risk domain.
- N2 · Threat models and risk framing
Threat model, risk pathway, threat-actor profile.
- N3 · Thresholds and checkpoints
Capability threshold, threshold status, alert threshold, checkpoint rule, halt condition, reassessment trigger.
- 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.
- N6 · Mitigations and security
Monitor, safeguard, robustness level, coverage level, exemption, security level, security control, mitigation change, internal deployment and internal-use risk.
- N7 · Incidents
Incident, incident type, discovery method, incident reporting deadline and recipient.
- N8 · Transparency and review
Redaction, redaction reason, external review, evaluator access attestation, publication rights clause, AI-model review, evaluator independence and conflict of interest.
- N9 · Commitments and governance
Commitment, framework update protocol, roadmap item, industry-wide recommendation, pre-release sharing window, mapping declaration, accountable decision-maker and sign-off, noncompliance and whistleblower reporting.
- N10 · Assurance roles
The CRediT-pattern accountability layer for NIKOLAI claims. Registered as a track; deliberately unpopulated in this release.
Registered, not yet populated
