Examples
Examples from published documents
- Google DeepMind sets the ML R&D automation cut-off at 90% of benchmark challenges rather than the theoretical 100% (Gemini 3.7 Flash FSF report, p.4)
- The EU Code requires the margin to take into account uncertainties of "systemic risk sources (e.g. capability improvements after the time of assessment)" as well as of assessments and mitigations (EU GPAI Code of Practice, Safety and Security chapter, Measure 4.1)
- Anthropic's earlier policy specified "creating a 6× buffer" in its testing methodology before RSP v2.0 replaced it with more general requirements (Anthropic RSP v3.4, changelog)
Counter-examples
Not to be confused with
- Elicitation effort, meaning how hard evaluators try to draw a capability out of a model, reduces under-estimation directly; a safety margin is the allowance kept for whatever under-estimation remains.
Source of record
Where this definition comes from
Gemini 3.7 Flash FSF report, p.4
“"a model theoretically needs to complete 100% of the benchmark challenges to reach the CCL, but we conservatively set our safety margin at 90%"”
https://storage.googleapis.com/deepmind-media/gemini/gemini_3-7_flash_fsf_report.pdfEU GPAI Code of Practice, Safety and Security chapter, Measure 4.1
“"The safety margin will: (1) be appropriate for the systemic risk; and (2) take into account potential limitations, changes, and uncertainties of: (a) systemic risk sources (e.g. capability improvements after the time of assessment)"”
https://ec.europa.eu/newsroom/dae/redirection/document/118119
Crosswalk
How named organisations use this concept
| Organisation | Their term, as published | Match & verification | Source |
|---|---|---|---|
| Google DeepMindShadow mapping Gemini 3.7 Flash FSF report / Gemini 3.1 Pro model card | “Safety margin: "a model theoretically needs to complete 100% of the benchmark challenges to reach the CCL, but we conservatively set our safety margin at 90%" (report p.4). Safety buffer: "Our frontier safety strategy is based on a 'safety buffer' to prevent models from reaching critical capability levels (CCLs), i.e. if a frontier model does not reach the alert threshold for a CCL, we can assume models developed before the next regular testing interval will not reach that CCL." (model card, Frontier Safety)” | exactEQ confidence: high | Gemini 3.7 Flash FSF report |
| EUShadow mapping EU GPAI Code of Practice, Safety and Security chapter | “Measure 4.1: Signatories apply acceptance criteria "incorporating a safety margin (as specified in the following paragraph)". "The safety margin will: (1) be appropriate for the systemic risk; and (2) take into account potential limitations, changes, and uncertainties of: (a) systemic risk sources (e.g. capability improvements after the time of assessment)". Measure 7.2: the Model Report includes "details of the safety margins incorporated".” | exactEQ confidence: high | EU GPAI Code of Practice, Safety and Security chapter |
| xAIShadow mapping xAI Frontier AI Framework (30 Jun 2026) | “"xAI applies a systemic risk acceptance criteria to each identified risk, incorporating a margin of security, to determine whether each identified systemic risk and the overall systemic risk are acceptable" (s.2.3). "In assessing whether to accept residual risk, we review the risk tiers for each systemic risk category and incorporate appropriate safety margins." (s.2.3)” Margin named but not quantified. PDF metadata /Title reads "Privileged/Confidential DRAFT working FRAMEWORK DOC"; no xAI statement disambiguating draft vs. final was found. | exactEQ confidence: medium | xAI Frontier AI Framework (30 Jun 2026, draft-marked) |
| AnthropicShadow mapping Responsible Scaling Policy v3.4 (changelog entry for RSP v2.0) | “"We have replaced some specifics in our previous testing methodology (e.g., using 1% of compute for elicitation or creating a 6× buffer)"; "the science of evaluations is not currently mature enough to make confident predictions about the precise buffer we should require between current models and a Capability Threshold." (Changelog, RSP v2.0)” Historical: describes the RSP v2.0 change away from a fixed numeric buffer; RSP v3.0 was a comprehensive rewrite. | closeCL confidence: high | Anthropic Responsible Scaling Policy v3.4 |
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.
- Microsoft
A heading in the framework (p.8) refers to a "safety buffer", but the text under it describes re-evaluating models after mitigation rather than a defined margin.
Microsoft Frontier Governance Framework (Feb 2026) - OpenAI
The Frontier Governance Framework treats models as having crossed a threshold when it is "unable to rule out" that they have (§2.3); that precautionary rule is recorded under threshold-status.
OpenAI Frontier Governance Framework
Divergence
Where sources materially disagree
"Safety buffer" and "safety margin" are used for different allowances. Google DeepMind's safety buffer is a gap in time and capability between an alert threshold and the next scheduled test; its safety margin, like the EU Code's, is a lower cut-off or allowance applied at the decision itself. Anthropic once used a numeric buffer (6× effective compute) and replaced it with more general requirements because the science of evaluations was not mature enough to set a precise buffer.
Gap
Among the current frameworks reviewed, CASRAI found one published numeric margin, Google DeepMind's 90% figure for ML R&D automation; the EU Code requires margins to be explained in Model Reports, which are not public, and xAI names margins without values.
