Skip to content

Hosted decision providers and JevBench

Contract review: 2026-10-16. This inventory covers the hosted offerings and author demos on Benchmark Heaven's API board and the five services in Celeris's launch comparison. It also records newly documented compatible services found during that review. It is a dated inventory, not an exhaustive claim about future providers.

Choose a factory

All named factories return TypeSafeClient and accept an optional caller-owned HttpClient and configuration callback. Credentials must be issued by the chosen service. Base URLs below omit /v1; the generated transport preserves any prefix. The Perplexity and Inception profiles change the final operation to /v1/decisions.

Factory Base URL Default model Managed images Model discovery
Celeris https://inference.celeris.ai/celeris-1-decision/ celeris-1-decision Named x_celeris.images Supported
Perplexity https://api.perplexity.ai/ pplx-decider-v1.1-27b Inline image_url parts in state Not documented for this profile
Inception https://api.inceptionlabs.ai/ mercury-decide Inline image_url parts in state Supported
Liquid https://api.liquid.ai/decisions/ d1 Root images data URLs Not documented for this profile
Wity https://api.wity.alphanimble.com/ wity-1 One root image data URL Not documented
SubQ https://api.subq.ai/ subq-snap-preview Root images objects with mediaType and base64 Supported
Sage https://sage.levanto.ai/ sage-latest Separate native API required Supported
Vansa https://api.vansa.org/ vansa-3.4 Text profile Supported
Milliseconds https://api.milliseconds.ai/ decision-machine-1 Separate native API required Not documented
Instinct https://api.zoowork.ai/ instinct Text profile Not documented
Classifier https://classifier.dev/ jev Root images data URLs with model imajev-4b Supported
Bespoke https://api.bespokelabs.ai/ nimble-latest Text profile Not documented for this profile
Surd https://api.simplexai.net/ spx-cd-flash Text profile Supported
Fastino https://api.fastino.ai/ fastino/GLiDE Text profile Uses a different catalog API
Amos https://api.amos.scira.ai/ amos-1-large This adapter implements its text subset Not documented for this profile
MicrosoftDecision Caller-supplied Foundry resource + /providers/microsoft/ Caller-supplied deployment name Text profile Deployment management is separate 
1
2
3
4
5
using var client = CustomProviders.Inception("INCEPTION_API_KEY");
var questions = new QuestionSet();
var refund = questions.AddNoul("refund", new NoulQuestion("Is a refund requested?"));
var result = await client.SystemOneAsync("I was charged twice; please refund one payment.", questions);
Console.WriteLine(result.Get(refund).Noul);

Choose alternate models per call or through DefaultModel, for example instinct-dual-4b, nimble-v4-max, spx-cd-pro, or amos-1-medium. Model names and availability belong to each provider; discovery returns the provider's catalog, which can include models unsupported by its decision endpoint.

Instinct's paid ZooData endpoint uses https://api.zoodata.ai/ and separate keys; change BaseUri explicitly. Classifier accepts its own workspace key or the literal "unused" for documented anonymous access. Amos's default "unused" bearer value is a placeholder for its currently public endpoint. MicrosoftDecision accepts a caller-managed Entra bearer token, the resource endpoint, and deployment name; refresh tokens via per-call RequestOptions.ApiKey. Azure API-key auth and automatic Entra refresh are not implemented by this helper.

Images and response semantics

RequestOptions.Images accepts PNG/JPEG ImageAsset instances. The adapter always uses the provider's inline JSON shape. On Perplexity/Inception it preserves the existing state, appends a text label Image named <name>:, then an image_url part for each image. Existing array items keep their order. Root image arrays use dictionary insertion order; their names do not become provider-side asset handles. Raw inline content and other documented native options can be passed as structured state or ExtraBody. Typed/raw image-field collisions throw. Raw fields remain caller-owned and server-validated.

The SDK enforces selected documented question, criterion, image-count, byte, and serialized-body limits. These are early checks, not full provider validation; pixel dimensions, rendered token limits, account tiers, and remaining schemas are checked by the server. SubQ allows four images, 4 MiB each and 8 MiB total; Inception permits 16 images of up to 10 MiB each; Wity permits one image. The smaller whole-body limit can bind before decoded image limits. Box, Celeris-only options, TypeSafe experimental headers, and gzip are rejected by these profiles. The Amos adapter currently exposes Noul/choice/score, not its multi/span/region primitives. Images for Sage and milliseconds.ai require their separate native APIs.

GLiDE returns an integer winning score and a fractional expected_level. For responses whose model is glide, the SDK maps expected_level to ScoreAnswer.Score, including projections and replay, preserving the SDK's expected-score semantics. The original integer and provider metadata remain in RawJson. GLiDE adds bounded HTTP 425 warm-up retries; its default request timeout is 300 seconds and total retry budget six minutes. Caller configuration can replace those defaults. Other providers keep the existing retry policy.

SystemOneResponse.RawJson and individual answer RawJson preserve unknown metadata, including Wity reasoning and classifier abstention fields. A typed choice is not permission to ignore an abstained flag. Reported confidence and probabilities are retained without imposing one provider's formula on another. All tests use synthetic handlers: they establish serialization and SDK behavior, not model acceptance, uptime, accuracy, latency, calibration, or billed usage.

Remaining benchmark offerings

A benchmark row may use a custom adapter, a temporary demo, or ordinary chat inference. It does not establish a public System One service. The following entries complete the reviewed board inventory beyond the factories above.

Offering Integration status and source
OpenAI Decisions, GPT-6 Luna decision comparator Use the generated Decisions surface in tryAGI.OpenAI; its predicate/array schema differs from System One.
Mercury Decide on OpenRouter Direct Inception profile supported above; OpenRouter routing is a separate protocol/host contract.
Liquid d1 on OpenRouter/Vercel Direct Liquid profile supported above; gateway access is separate.
attest-31B-v1.0 Public page could not be retrieved during review. An older author submission describes a different 12B attest build; do not assume those IDs or contracts identify the current 31B service.
Gemma 4 (Autoloops) Provider is a benchmark lead; a supported public System One contract was not established.
SimpleJev Qwen3.6-35B-A3B / Qwen3.8-27B Author implementation, author-hosted demos rather than a stable credentialed factory.
system-one-open Author repository; use a verified deployment with CustomProviders.SystemOne.
openjev-sglang Author repository; verify its deployment/adapter contract before selecting a custom endpoint.
JevAct Author submission uses a different native protocol; requires a dedicated adapter.
Fastino GLiNER-2.5-Decide Native GLiNER schema, distinct from GLiDE's System One route.
BB-Qwen3.5-4B-LoRA Author submission describes a temporary authenticated tunnel; obtain current endpoint/key before using a custom profile.
GPT-6 Luna low/medium, GPT-5.6 Luna chat comparators Use the OpenAI SDK chat/Responses APIs, not the System One profile.
Gemini 3.1 Flash-Lite Use the Google Gemini SDK native contract.
DeepSeek V4.1 Flash / Qwen3.8 27B on Chutes General-purpose hosted comparators; use the provider's documented chat API, not an inferred System One route.
Wity auto/always/off variants Same Wity factory; choose documented reasoning through ExtraBody.
Classifier fast wrapper Classifier factory above; it routes to Jev and is not an independent model.
Sage 1.4 / Nimble-v4 and max, additional or dated listings Sage/Bespoke factories above; use current catalog or documented model IDs rather than assuming frozen benchmark versions remain served.

For a provider with a verified public System One contract and no named factory:

1
2
using var client = CustomProviders.SystemOne("PROVIDER_KEY",
    new Uri("https://your-verified-provider.example/prefix/"), "provider-model");

This sends Noul/choice/score over generated Bearer-authenticated transport to <base>/v1/systemone. It assumes neither image support nor model discovery. It does not translate incompatible APIs or certify an arbitrary endpoint.

Which JevBench?

Celeris's launch comparison evaluates a 3,210-item/22-configuration Jev-bench corpus and also reports typed-decisions results. Benchmark Heaven's API board uses its own versioned JevBench protocol, full and equated subsets, calibration, cost, speed, and composite/capability views. Decision Index and Decision Bench are additional evaluations referenced by vendors. Scores from these different corpora, methods, and dates are not interchangeable. This SDK embeds no ranking, price, or latency claim; evaluate representative requests with the exact hosted model and record response model IDs and raw usage.

Primary contract sources