Decision API Reference
NVIDIA Dynamo exposes two non-streaming text-evaluation contracts. For runnable curl and SDK examples, see the Decision Model API guide.
Contracts
The public contracts are the supported text-evaluation subset of OpenAI Decisions and Jev/TypeSafe System One. Do not send SGLang-native /v1/decisions request fields to Dynamo. Neither route accepts nvext, including {} or {"format":"oai"}.
Use the exact model alias returned by GET /v1/models. Supported inputs, candidate counts, and accounting depend on the model/executor profile. Dynamo currently admits at most 128 questions per request, subject to independent aggregate resource limits. These APIs accept text inputs; image, audio, and video inputs are not supported by the documented profiles.
Usage
Input usage counts rendered prompts across question executions, not only the shared input text. SGLang scores without decoding output tokens. OpenAI total_tokens is the sum of input_tokens and output_tokens. Cache-read values come from backend counters or a profile verified to perform no cache reads; Dynamo returns 500 rather than reporting an unmeasured read count as zero. Questions in one request can reuse cached prefixes, but separate requests do not share them.
Dynamo currently always returns usage.input_tokens_details.cache_write_tokens: 0 because cache-write usage is not measured. This value does not mean that caching is disabled or that no KV blocks were written. The current Decisions response schema requires an integer, so omission, null, and "N/A" are not supported. Use cached_tokens for measured cache reads; it is separate from the unmeasured write count.
Answer Statistics
x_label_mass is a Dynamo extension on Jev answers reporting candidate vocabulary mass, not another conditional choice probability. It is not added to OpenAI answer objects. confidence is an executor-derived statistic, not a calibrated probability of correctness.
Errors and Headers
An OpenAI refusal is an answer variant, not an HTTP error. If an executor refuses any Jev question, Dynamo returns decision_refused for the request rather than a partial answer object.
Capacity responses have code rate_limit_exceeded and Retry-After: 1; status alone does not identify overload. OpenAI errors use a nested error object; Jev uses top-level error fields. Responses expose x-request-id and x-typesafe-request-id.
For enablement, admission limits, and cache isolation, see Decision API Configuration. For backend-specific requirements, see the SGLang guide.