Evaluation run

NIKOLAI proposal: a single execution, or a declared batch of executions, of an evaluation against a specified model checkpoint and configuration, recorded with attempt count, scoring rule, and date. No source in this cluster defines the term cleanly; NIKOLAI is proposing the record shape, not restating a lab's own definition.

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, "Investigating incidents / cybersecurity evals", main text
“141,006 evaluation runs "where Claude could have obtained internet access" were reviewed; "evaluation run" is used as the counting unit ... but not defined.”
https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals

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, "Investigating incidents / cybersecurity evals"	“"141,006 evaluation runs 'where Claude could have obtained internet access' were reviewed"; "evaluation run" is "used as the counting unit ... but not defined".”	none confidence: high	Anthropic, "Investigating incidents / cybersecurity evals"
STREAM (discovery sweep) STREAM (arXiv 2508.09853)	“Item-level reporting [UV].” unverified — item-level reporting is confirmed as STREAM's general purpose, not confirmed to define a run-level unit specifically.	none confidence: low	STREAM, arXiv:2508.09853

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.

Google DeepMind
"TAC score: success is evaluated based on whether the model successfully captures the flag in at least 50% of the unique challenges within a specific difficulty tier, requiring only one success per challenge within the maximum allocated attempt budget" (p.17); "Objective Success Rate (OSR)" and "Query Success Rate (QSR)" (p.32) — scoring-metric pointers, not a run-record definition.
Gemini 3.7 Flash FSF report
Meta
"Pass@k: a metric used in AI capability evaluations that represents the likelihood that the model under test will successfully complete a task when given k independent attempts" (fn 7); criterion "< 75% pass@10" (§4.2.1) — a scoring pointer, not a run-record definition.
Meta Advanced AI Scaling Framework v2
Frontier Model Forum
"Replication testing: re-executing key evaluations to confirm that results are reproducible and accurate." (2.1) — a process pointer, not a run-record definition.
Frontier Model Forum, Third-Party Assessments

Gap

The only published use of "evaluation run" as a counting unit is Anthropic's incident denominator, and the run turned out to be the unit at which harm occurred. Run-level records (checkpoint, safeguard configuration, network access, operator) are what incident reconstruction needed in both the Anthropic and OpenAI cases.