Choosing the Endpoint

LM Studio serves decision models through two endpoints, and the gem has a protocol for each:

	:decisions (default)	:system_one
Endpoint	POST /v1/decisions	POST /v1/systemone
Origin	OpenAI's Decisions API	Jev (TypeSafe) System One API
Input field	input	state (required)
Questions	an array: {type: predicate \| choice \| score, name, instructions, choices \| levels}	a hash keyed by name: {type: noul \| choice \| score, instructions, criteria}
Probability criteria	folded into the instruction text as "Yes: … / No: …" lines	sent as structured criteria
Images	input_image data URLs inside a content array	data URLs in a top-level images array
Answers	an array with a name on each	a hash keyed by question name

Pick one with lms_judgment_protocol:

RubyLLM.configure { |config| config.lms_judgment_protocol = :system_one }  # process-wide

context = RubyLLM.context { |config| config.lms_judgment_protocol = :system_one }
RubyLLM.judge(text, questions: { ... }, model: 'd1-omni-600m', provider: :lms, context:)

Judgments go to a decision endpoint whatever lms_protocol says; a chat protocol never handles them.

Do the answers differ?

For a large model, barely. For a small one, sometimes a lot. The two endpoints phrase the same questions to the model differently, and a 600M-parameter model notices. From example 18_decision_models.rb, one billing ticket:

Model	Endpoint	Urgent	Team	Confidence	Mood (0..2)
Clef Flash	decisions	0.96	billing	0.96	1.21
Clef Flash	system_one	0.95	billing	0.97	1.15
D1 Omni 600M	decisions	0.91	billing	0.51	0.53
D1 Omni 600M	system_one	0.91	billing	0.94	1.08
Kev 0.8B	decisions	0.64	billing	0.66	1.03
Kev 0.8B	system_one	0.64	billing	0.70	1.21

The biggest gap is probability criteria — the optional yes/no descriptions on a probability question. Decisions has no field for them, so RubyLLM folds them into the instruction text; System One sends them as data:

recommend: { type: :probability,
             instructions: 'Would this customer recommend the product to a friend?',
             criteria: { yes: 'Ends satisfied with the product itself', no: 'Would steer others away' } }

Model	decisions	system_one
Clef Flash	0.758	0.910
D1 Omni 600M	0.012	0.803
Kev 0.8B	0.602	0.788

Same model, same question, opposite conclusions. Images show the same pattern for D1 Omni 600M.

Recommendation

1. Pick the endpoint first. :decisions is the default and fine for most work; with D1 Omni 600M and either probability criteria or images, :system_one gave clearly better answers on the author's machine.
2. Tune on that endpoint. Thresholds and wording tuned on one endpoint don't carry over to the other unchanged.
3. Prefer guidance in the question text over criteria when you need to switch endpoints: the instruction text reaches the model the same way on both.

Gem-specific details

• Answers are matched by name. RubyLLM 2.1.1's stock Decisions protocol pairs answers with questions by position; LM Studio returns them in question order today, but the gem matches on each answer's name so a reordering can't silently swap answers. A missing answer raises RubyLLM::Error.
• The System One path is fixed for an lms_api_base that already ends in /v1 (the stock protocol would request /v1/v1/systemone).
• System One images go in the top-level images array, where LM Studio reads them; the stock protocol refuses images entirely. See Judging Images.