Support matrix and thresholds

Every model answers every preset it can in its canonical, zero-shot form. Some families were trained on a preset, and some have a calibrated threshold for it. This page shows both, so you can tell what changes when you switch the decision model.

Support matrix

TC trained and calibrated, T trained, B the base model with every adapter off, Z zero-shot, — cannot run.

Preset	vela3-fast	vela3-lite	vela3-mini, -flash, -turbo	d3 (local or hosted)	Decision 2.0	Vela 2.0
domain, fact_check, feedback	TC	T	T	Z	Z	TC
jailbreak	TC	B	B	Z	Z	TC
harm	TC	B at the request stage; T at the response stage	B at the request stage; T at the response stage	Z	Z	TC
harm categories	T	T at the request stage; Z at the response stage	Z	Z	Z	Z
pii	TC	T	T	—	—	TC
hallucination	T	T	T	—	—	T
refusal	T	T	T	Z	Z	Z
modality	TC for text, image, text_image; Z for the others	T for the three; Z	T for the three; Z	Z	Z	Z
turn_relation, complexity, relevance, same_request, tool_rank, tool_need, answer_check	Z	Z	Z	Z	Z	Z

• The Vela 3.0 d3 sizes turn trained adapters on by question ID: domain, factcheck, feedback, r_harm, r_refusal, modality, and, on vela3-lite only, the hazard: prefix of the harm categories; vela3-fast and Vela 2.0 by wording.
• The Vela 3.0 d3 sizes answer the jailbreak preset and the request-stage harm preset with their base, every adapter off. Their attack and p_harm adapters lose an attack that other text surrounds (on vela3-turbo, 0.975 against the base's 0.9992 for an injection after 8K tokens of prose) and rate benign text higher, and guard thresholds such as 0.995 were set on the base's answers.
• Image input is read by the d3 family and the Vela 3.0 d3 sizes. No current decision model reads audio or video for routing rules.

Calibrated thresholds

The default predicate of a rule that writes none, on the models measured on the router signal suite:

Model	jailbreak	domain	pii	fact_check	feedback	harm	modality
vllm-sr/vela3-fast	0.88	0.41	0.18	0.9999	0.47	0.52	0.63
vllm-sr/Vela-2.0-0.3B	0.75	0.28	0.01	0.93	0.37	0.46	0.51
vllm-sr/Vela-2.0-0.8B	0.71	0.38	0.07	0.994	0.34
vllm-sr/Vela-2.0-4B	0.63	0.45	0.05	0.9984	0.33
vllm-sr/Vela-2.0-9B	0.42	0.46	0.14	0.998	0.35

note

Every other model uses 0.5, with an uncalibrated_preset warning at load. The hallucination preset's spans keep the model's own thresholds. The records behind these values are in the model runtime records, and each model's page under Supported models lists its own.