Until providers serve Open System One themselves, a client or gateway translates. These are the mappings for today’s System One APIs. Each adapter also checks every answer against the answer guarantees and rejects what a model cannot do with the standard errors.
The TypeSafe and Cloudflare mappings were verified against their live APIs on 2026-11-07. The OpenAI mapping follows OpenAI’s published documentation.
TypeSafe
POST https://api.typesafe.ai/v1/systemone
Open System One is a superset of TypeSafe’s format, so requests and responses pass through unchanged, except:
Question types:noul, choice, score, and box. Reject extraction types with unsupported_question_type.
noul without instructions: TypeSafe rejects it, although its schema allows it. Send the question name as the instructions.
Images:png, jpg, and jpeg only; reject webp with unsupported_parameter. Requests with images need the header TypeSafe-Danger-Zone-Images: true.
Cloudflare Clef
POST https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/run/@cf/cloudflare/clef
Open System One
Clef
model
model, such as clef
state
state
questions
questions. Every question needs instructions: send the name when they are missing.
assets
images: a list of data URLs, read before the state. At most 4, up to 4 MiB each and 8 MiB in total.
Question types:noul, choice, and score.
Limits: at most 64 questions, names of up to 100 letters, digits, _, ., or -, at least 2 choice options, and 2 to 10 score levels. Reject anything beyond them with unsupported_parameter.
Response: the answers are in result of Cloudflare’s { success, result } envelope. Clef does not report truncation.
Cloudflare Clairic
POST https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/run/@cf/cloudflare/clairic
Open System One
Clairic
model
model, such as clairic
state
state as text: send structured state as compact JSON
questions
fields, with names of up to 64 characters, starting with a letter or _
assets
images, as for Clef
reference_date
reference_date
quantiles on an estimate
options.include_quantiles, for the whole request
—
options.return_state_text: true, so spans always have the text they index
Question types: the extraction types. Reject decision types with unsupported_question_type, and forecasts beyond 64 steps with unsupported_parameter.
Response: read result from the envelope. Return state_text when it differs from the state, and truncated.
Quantiles: keep them only on estimates that asked for them, never on forecast steps. When a forecast comes back as parallel arrays, rebuild its steps.
Derived fields: Clairic can return a value that contradicts its own p_null, a guess of null, a span that does not cover its text, or a text list with repeated items collapsed into one. Recompute value from p_null and guess, treat a null guess as no guess, locate the text again or set the span to null, and restore repeated items from the spans.
OpenAI Decisions
POST https://api.openai.com/v1/decisions
Open System One
OpenAI Decisions
state as a string
input as a string
state as an object or array
input: a user message with an input_json part
questions: { name: {…} }
questions: [{ name, … }]
noul
predicate, with criteria.true and criteria.false written into instructions
choice with criteria: { value: description }
choice with choices: [{ value, description }]
score with criteria: [label, …]
score with levels: [{ label }], 2 to 10
assets
input_image parts with data URLs
OpenAI Decisions answer
Open System One answer
answers: [{ name, … }]
answers: { name: {…} }
predicate with probability
noul with noul
choice with probabilities: [{ value, probability }]
choice with probabilities: { value: probability }
score with probabilities: [{ value, label, probability }]
score with probabilities: { index: probability } and legend: { index: label }
refusal
No equivalent yet: the request fails
Question types:noul, choice, and score. OpenAI documents no box type and no extraction types; reject them with unsupported_question_type.
Open System One