Examples
Examples from published documents
- Google DeepMind treats "checkpoints and versions the same model if their base capabilities stem primarily from the same foundational training run" (Google DeepMind Frontier Safety Framework v3.1, glossary, p.18)
- Meta presumes that non-material modifications "inherit the same risk threshold as the underlying model" (Meta Advanced AI Scaling Framework v2, §3.3.1, p.17)
- "Unless stated otherwise, evaluation results are on the final deployed checkpoint of Grok 4.6." (Grok 4.6 model card, §1.1, p.5)
Counter-examples
Not to be confused with
- A pseudonym such as "Internal Model 1 (IM1)", used by OpenAI for an internal-only research model, labels a model within one document but does not identify it across documents.
- A model family name such as Gemini or Grok groups related models; claims made about the family do not pin down which version was tested.
Related terms
Related AI safety terms
Terms defined under this element that are not NIKOLAI elements themselves, each as the cited document uses it. Every term is also listed in the AI safety glossary.
- Foundation model
- Foundation model, in California SB 53, means an AI model that is trained on a broad data set, designed for generality of output, and adaptable to a wide range of distinctive tasks. A frontier model is a foundation model above a compute threshold.
- Source: California SB 53
- General-purpose AI model
- A general-purpose AI model, under EU AI Act Article 3(63), is an AI model that displays significant generality, can competently perform a wide range of distinct tasks, and can be integrated into a variety of downstream systems or applications.
- Source: EU AI Act Article 3 (Definitions)
- General-purpose AI
- General-purpose AI, in the International AI Safety Report, refers to AI models and systems that can perform a wide variety of tasks, as opposed to systems specialised for a single task.
- Source: International AI Safety Report 2026
- Final deployed checkpoint
- A final deployed checkpoint is the exact model version that is released to users. xAI's Grok 4.6 card states that its evaluation results are on the final deployed checkpoint unless stated otherwise.
- Source: Grok 4.6 model card
Source of record
Where this definition comes from
Google DeepMind Frontier Safety Framework v3.1, Glossary, p.18
“"we consider checkpoints and versions the same model if their base capabilities stem primarily from the same foundational training run, but will conduct additional testing on new versions as specified in Section 1.3"”
https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/strengthening-our-frontier-safety-framework/frontier-safety-framework_3-1.pdfMeta Advanced AI Scaling Framework v2, §4.2, §3.3.1
“"Asset": "the version of the model that we will test" (§4.2, p.26); non-material modifications "are presumed to inherit the same risk threshold as the underlying model" (§3.3.1, p.17).”
https://ai.meta.com/static-resource/Meta_Advanced-AI-Scaling-Framework-v2
Crosswalk
How named organisations use this concept
| Organisation | Their term, as published | Match & verification | Source |
|---|---|---|---|
| AnthropicShadow mapping Risk Report | “"Covered models": the models an assessment section focuses on; in §2 these are "Claude Mythos 5 and Model 2" (para., §2.3). CASRAI's reading of the report is that it gives no identifiers for anything.” | narrowNR confidence: medium | Anthropic Risk Report, August 2026 |
| OpenAIShadow mapping Hugging Face incident writeup / GPT-5.6 model card | “Pseudonymous internal model: "an internal-only research model that we will call Internal Model 1 (IM1)". Product tiers named Sol, Terra and Luna in the GPT-5.6 card.” | narrowNR confidence: medium | OpenAI, "Hugging Face incident and the road ahead" |
| Google DeepMindShadow mapping Frontier Safety Framework v3.1 | “Identity rule: "we consider checkpoints and versions the same model if their base capabilities stem primarily from the same foundational training run, but will conduct additional testing on new versions as specified in Section 1.3" (FSF glossary, p.18).” Falls under the "checkpoint" false-friend label: here it denotes a lifecycle version, distinct from Amodei's policy-rule sense or OpenAI's training-checkpoint sense of the same word. | closeCL confidence: high | Google DeepMind Frontier Safety Framework v3.1 |
| xAIShadow mapping Grok 4.6 / Grok 4.20 model cards | “"Unless stated otherwise, evaluation results are on the final deployed checkpoint of Grok 4.6." (§1.1, p.5). Grok 4.20 modes: "single-agent (Grok 4.2 SA) or multi-agent (Grok 4.2 MA)" (S1).” | narrowNR confidence: medium | Grok 4.6 model card |
| MetaShadow mapping Advanced AI Scaling Framework v2 | “"Asset": "the version of the model that we will test" (§4.2, p.26). Inheritance rule: non-material modifications "are presumed to inherit the same risk threshold as the underlying model" (§3.3.1, p.17).” | closeCL confidence: high | Meta Advanced AI Scaling Framework v2 |
| California SB 53Shadow mapping SB 53 statute text | “"a new frontier model or a substantially modified version of an existing frontier model" (22757.12(c)(1)). This is a class definition, not an identifier, and "substantially modified" is left undefined.” | broadBR confidence: medium | California SB 53 |
| US Government (Executive Order 14409)Shadow mapping Executive Order 14409 | “"covered frontier model", designated "for the purposes of this order" by the Director of NSA (Sec. 3(a)). A designation class, not an identifier.” | broadBR confidence: medium | Executive Order 14409 |
What do these codes mean?
- exact
- The source term is equivalent to this element
- close
- The source term is close but not equivalent to this element
- broad
- The source term is broader than this element
- narrow
- The source term is narrower than this element
- none
- No mapping claim — used for false-friend and declared-but-undefined rows
- EQ
- Equivalent
- CL
- Close
- BR
- Source is broader than the element
- NR
- Source is narrower than the element
- FF
- False friend — same or similar label, different meaning
- DU
- Declared but undefined by the source
- UV
- Unverified
Related, not mapped
Pointers that are not crosswalk claims
These sources mention this concept but do not define or map it clearly enough to count as a crosswalk row — noted here so the research is visible without overstating it as a mapping.
- Frontier Model Forum
Access items include "representative model version" and "near-identical model version" -- a pointer to how third-party assessors talk about model identity, not a mapping to a NIKOLAI identifier scheme.
Frontier Model Forum, Third-Party Assessments technical report
Divergence
Where sources materially disagree
"Checkpoint" is a false-friend label across this corpus: a policy rule in Amodei's own essay, a training checkpoint in OpenAI's Preparedness Framework, a lifecycle checkpoint/version in GDM and xAI (as used in this element's rows), and the first stage of a two-stage Loss of Control evaluation in Meta's framework. NIKOLAI should not conflate these uses under one term.
Gap
No source publishes a machine-resolvable model identifier. GDM's "same foundational training run" rule and Meta's inheritance rule are the two existing identity rules and should be crosswalked explicitly.







