Documentation

Technical pilot guide.

Singular is an OpenAI-compatible gateway offered through a controlled pilot. Credentials, model routes, traffic boundaries, and activation are provisioned during onboarding; there is no public or automatic API access.

OpenAI-compatibleText-only pilotNon-streamingFail-open verified before activation

Quickstart

Qualify, measure, activate.

A pilot is a gated evaluation of one customer-shaped workload, not a general account signup.

  1. 01

    Qualify one workload

    Agree on the route, provider/model configuration, context shape, success criteria, privacy boundary, and rollback owner.

  2. 02

    Run side by side

    Keep the unoptimized baseline while the pilot measures reduction, quality, latency, skips, and passthrough behavior.

  3. 03

    Activate deliberately

    Enable only the qualified traffic slice after evidence and operating gates pass. Roll back through the agreed kill switch if they stop holding.

Request contract

POST /v1/chat/completions

Use the pilot base URL and Singular API key issued during onboarding. Keep secrets in your environment; never commit them or paste them into the Portal.

Example requestplaceholders only
curl -X POST "$SINGULAR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer $SINGULAR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<qualified-model>",
    "messages": [
      {"role": "user", "content": "Question\n<singular_context type=\"rag\">...retrieved context...</singular_context>"}
    ]
  }'

Unsupported request shapes pass through or are rejected according to the route contract agreed for the pilot. Provider credentials remain server-side.

Evidence

Read the result in context.

Public benchmark numbers are release evidence, not a forecast for your traffic. Your pilot receives its own baseline, limits, activation decision, and rollback record.

Marker contract

Supported matrix and limits.

The controlled runtime is text-only and non-streaming with one response, at most 64 messages, a 100,000-token input envelope, and up to 4,096 output tokens. Tools, functions, images, and audio are not supported. Only content inside the exact S-Core RAG marker may be reduced; benchmark evidence for GLM 5.2 FP8 and Kimi K2.7 does not automatically enable a production route.

Eval and activation

Passthrough and rollback.

S-Core activates only after workload-specific evidence, privacy review, rollback readiness, and customer approval. Within the supported request contract, unsafe, inapplicable, low-value, disabled, or failed optimization uses the original request and is not billed as optimized usage. Unsupported API shapes are rejected.

Privacy and security defaults

Release notes and support.

Raw prompts, marked context, and provider responses are not persisted by default. Current release: S-Core 0.3.0, profile mvp.safe.v3. For qualification or support, contact sales@mattom.io without sending prompts, credentials, or session tokens.