Coverage level

NIKOLAI proposes Coverage level as a graded controlled-value axis stating the scope of content or behavior a safeguard is designed to catch, kept independent of how robust the safeguard is against circumvention. This is an unsourced NIKOLAI editorial synthesis. It must not be conflated with an evaluation's own test-category coverage (see divergence_note) or with a coverage *date* (a distinct property).

This is CASRAI's own proposed definition, not a definition any named organisation has agreed to. See what NIKOLAI is and is not.

Source of record

Where this definition comes from

Anthropic Risk Report, August 2026, §4.5, §4.5.2.2
“Coverage — the domain of usage a classifier is trained to catch”
https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf
Gemini 3.7 Flash FSF Report, p.29
“Coverage: 'assuming threat actors don't try to circumvent these safeguards, how much assistance would they get from the model'”
https://storage.googleapis.com/deepmind-media/gemini/gemini_3-7_flash_fsf_report.pdf

Crosswalk

How named organisations use this concept

Every row below is a shadow mapping. It is CASRAI's own reading of a published document. No lab, evaluator or regulator named here has declared, endorsed, or been consulted on this mapping. That will change only when an organisation files its own Mapping Declaration — see the non-endorsement policy.

Organisation	Their term, as published	Match	Source
Anthropic Anthropic Risk Report, August 2026	“"Coverage — the domain of usage a classifier is trained to catch"; Fable 5's classifier covers "the vast majority of potentially dual use usage relevant to research biology"; "100% of a generated research-biology probe set versus 17% for the narrower classifier"”	exact confidence: high	Anthropic Risk Report, August 2026
Google DeepMind Gemini 3.7 Flash FSF Report	“"Coverage: 'assuming threat actors don't try to circumvent these safeguards, how much assistance would they get from the model'"”	close confidence: medium	Gemini 3.7 Flash FSF Report
Frontier Model Forum FMF Third-Party Assessments	“"Evaluation coverage: 'identifying test categories, particular workflows, or user skill levels.'" This is coverage of an evaluation, not a safeguard.” False friend: FMF's "coverage" here names what an evaluation tests, not what a safeguard blocks. See divergence_note.	none confidence: medium	Frontier Model Forum, Third-Party Assessments

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.

OpenAI
Tier-one "fast, topical classifier model that determines whether or not the content is related to one of the two areas where we have deployed Preparedness Safeguards" — related, not a mapping (RL).
OpenAI GPT-5.6 System Card / Deployment Safety
xAI
Filters for "CRBN [sic] risks" — related, not a mapping (RL).
xAI Frontier AI Framework, 30 June 2026 (draft-labeled PDF metadata)

Divergence

Where sources materially disagree

FMF's "Evaluation coverage" is a false friend for this element: it describes the scope of an *evaluation* (which test categories, workflows, or user skill levels are probed), not the scope of a *safeguard* (what content or behavior a classifier is trained to catch). NIKOLAI's coverage-level element tracks safeguard scope only; evaluation coverage is a distinct concept that belongs under the evaluation record-type (N5), and the two must not be crosswalked to each other. See ALIGNMENT-MATRIX.md §3 E4, FMF row {FMF32}.

Gap

Anthropic and GDM independently separate coverage from robustness, which supports making them two elements. Anthropic's disclosed year-long gap (human-feedback vendor traffic "ran without blocking biological classifiers") was a coverage failure by deployment surface, not a robustness failure {RR}.