The AI space declares 33 types. Five of them hold nothing at all.
Each of these was searched for by name inside the space, in long and short form. None of them exists there under any type.
Accuracy — the most common way benchmarks report a result, and the only metric name written anywhere in the space.
Task success rate — how agents are scored, and the space holds 161 agents.
F1 score — the standard metric wherever precision and recall are traded off.
Pass@1 — the reporting standard for code benchmarks. SWE-bench and HumanEval are both already here.
Exact match — the standard metric for question answering. SQuAD and SQuAD 2.0 are both already here.
BLEU — the long-standing measure of machine-translation quality.
Win rate — what arena-style leaderboards report. Elo Rating exists here only as a tag.
Preparedness Framework — OpenAI's published risk framework, the counterpart to Anthropic's Responsible Scaling Policy, which is already here.
Model Spec — OpenAI's published specification of how its models should behave, the counterpart to Constitutional AI, which is already here.
Deliberative alignment — a published safety training method, and the safety type exists to hold exactly these.
Proposed to the AI space, each typed Safety approach with a description, an official source, and relations to models and topics already in the space. Together they take the type from four members representing one lab to seven representing two.
Preparedness Framework — OpenAI's framework for grading frontier capabilities and setting the safeguards required before deployment.
Model Spec — OpenAI's specification of how its models are intended to behave.
Deliberative alignment — a training method in which a model reasons over a written safety specification before answering.