System One presets

A preset is a built-in System One question: the application signals, the judges and the checks each ask one. The wording and options below are the canonical, zero-shot definition; a family with a trained binding is asked the trained form (see the support matrix).

Preset catalog

Preset	Type	Options	Stage and input	Reads	Asked by
domain	choice	14 subject areas	request	The current turn, truncated	domains
fact_check	choice	FACT_CHECK_NEEDED, NO_FACT_CHECK_NEEDED	request	The current turn, truncated	fact_check
feedback	choice	SAT, NEED_CLARIFICATION, WRONG_ANSWER, WANT_DIFFERENT, NO_FEEDBACK	request	The current turn, truncated, and the answer it replies to	user_feedbacks
jailbreak	choice	benign, jailbreak	request or response	The whole input; the answer at the response stage	jailbreak
harm	choice	safe, unsafe	request or response	The whole input; the answer at the response stage	safety without labels
harm_categories	set	The 34 harm categories	request or response	The whole input	safety with labels
pii	span	11 personal data labels	request	The whole input, in windows	pii, masking
hallucination	span	unsupported	response	The answer and its grounding context	hallucination
refusal	choice	compliance, refusal	response	The request and the answer	refusal, escalate accept.refusal
turn_relation	choice	continue, reask, shift, return	request	The current turn and the last 4 turns	turn_relation
complexity	choice	easy, medium, hard	request	The current turn, truncated	complexity
modality	choice	text, image, text_image, image_edit, speech, audio, video, video_edit	request	The current turn, truncated	modality
relevance	noul		request	The request and a candidate	context sources and compression
same_request	noul		request	The request and a cached request	response_cache
tool_rank	choice	The candidate tools	request	The request	tools
tool_need	noul		request	The request	tools
answer_check	noul		response	The request and the answer, whole	escalate accept.systemone

The domains are biology, business, chemistry, computer science, economics, engineering, health, history, law, math, other, philosophy, physics and psychology.

Harm categories

A safety rule's labels name some of these 34 categories, the labels of the trained vela3 hazard question:

• AEGIS 2.0 (21): Controlled/Regulated Substances, Copyright/Trademark/Plagiarism, Criminal Planning/Confessions, Fraud/Deception, Guns and Illegal Weapons, Harassment, Hate/Identity Hate, High Risk Gov Decision Making, Illegal Activity, Immoral/Unethical, Needs Caution, Other, PII/Privacy, Political/Misinformation/Conspiracy, Profanity, Sexual, Sexual (minor), Suicide and Self Harm, Threat, Unauthorized Advice, Violence.
• MLCommons hazard codes (13): S1 (violent crimes), S2 (non-violent crimes), S3 (sex-related crimes), S4 (child sexual exploitation), S5 (defamation), S6 (specialised advice), S7 (privacy), S8 (intellectual property), S9 (indiscriminate weapons), S10 (hate), S11 (suicide and self-harm), S12 (sexual content) and S14 (code interpreter abuse).

How each family is asked them:

• vela3-fast asks its trained 34-label set once.
• At the request stage, the Vela 3.0 d3 sizes are asked one hazard:<label> yes/no question per category, in the form vela3-lite's hazard adapter was trained on; vela3-mini, -flash and -turbo ship no hazard adapter and answer those questions zero-shot.
• Every other model answers the same labels zero-shot, in the same call.

Personal data labels

PERSON, EMAIL_ADDRESS, PHONE_NUMBER, US_SSN, CREDIT_CARD, LOCATION, IP_ADDRESS, DATE_TIME, US_DRIVER_LICENSE, US_PASSPORT and IBAN_CODE. A pii rule's ignore_labels and a masking plugin's entity_types name some of them.