Skip to content

Microsoft-Decision-1 — hosted decision-model intake (#1324)

Status: discovery / no live SchemaRouter qualification. Not installed, not a free local model, not a substitute for an execution authorization policy. Reviewed 2026-10-13.

What it is

Microsoft released Microsoft-Decision-1 on 2026-10-09, a hosted single-pass decision-scoring model post-trained from Qwen3.5-9B. It returns structured choice (finite-option selection), noul (binary probability), and score (ordered rating) results, not prose. Microsoft's advertised 36-benchmark accuracy/speed is vendor evidence, not a measured SchemaRouter comparison.

  • Announcement: https://commandline.microsoft.com/microsoft-decision-1-model-foundry/
  • Native deployment/API: https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/use-foundry-models-microsoft-decision
  • OpenRouter: https://openrouter.ai/microsoft/microsoft-decision-1/

Microsoft documents an input price of USD $0.042 per million tokens, $0 for output as of 2026-10-13. Foundry requires a billed Azure account, deployment and authentication. OpenRouter likewise requires an account/API key. The Microsoft-adapted weights have not been publicly released; an open-weight Qwen3.5 base model is not the Microsoft decision checkpoint.

Transport/API: not a proven SDK drop-in

Microsoft Foundry documents:

POST <foundry-resource-endpoint>/providers/microsoft/v1/systemone
Content-Type: application/json
api-key: <foundry-key>
# OR: Authorization: Bearer <Entra access token>

The body has the same conceptual System One request fields:

{
  "model": "<the-created-Foundry-deployment-name>",
  "state": {
    "query": "Cancel a customer's order."
  },
  "questions": {
    "route": {
      "type": "choice",
      "instructions": "Select a supported route. Choose none_of_above if no registered route performs the requested action.",
      "criteria": {
        "orders_lookup": "Look up an order; cannot cancel it",
        "none_of_above": "No offered route supports the requested operation"
      }
    }
  }
}

The Foundry model field is a deployment name, not the microsoft/microsoft-decision-1 OpenRouter model ID. That OpenRouter ID is used on OpenRouter's Decisions API, whose alpha route is separately versioned; never interchange credentials or paths.

SchemaRouter currently has SystemOneDecisionBackend, but it wraps the TypeSafe SDK, with SDK client system_one() calls and parsed response.choices["selection"]. Microsoft's documented Foundry HTTP response instead uses answers and distinct Azure authentication. Therefore neither a plain SystemOneDecisionBackend(base_url=...) nor --system-one-base-url has been qualified to work against Foundry. Likewise, OpenRouter model-ID swaps require response and SDK transport verification. A small independent HTTP/CallableDecisionBackend adapter is the appropriate experimental integration boundary, with strict local-option-ID and numeric probability validation; do not change planner defaults merely to register a provider.

No-fit detection and safety (#1291)

The model does not infer an implicit rejection option when a choice is constrained to registered routes. Explicitly include a reserved none_of_above candidate and use an independent, preregistered supported/unsupported/near-domain/forbidden test cohort to measure:

  • exact supported operation routing and supported false-refusal;
  • unsupported false-match and correct no-fit rejection;
  • probability calibration, order/wording perturbations and rate/transport errors;
  • real request latency, input tokens, exact provider/deployment revision availability and dated costs.

A provider decision remains advice. Deny the reserved no-fit choice before any tool call; on supported candidates, still require the host authorization, schema/argument contract and execution no-match gate. Never turn an noul score or arbitrary confidence into access authority. Do not fit a threshold using previously published ClicShopping DEV unsupported requests or rerun/retune frozen 0.14 #432/#424.

Evidence and go/no-go

The Jev team's small 2026-10-11 OpenRouter comparison (one pass; 200 supported BFCL items and 240 no-fit items) reports no-fit rejection of 92.5% Microsoft vs 88.8% Jev and median round trips of 688ms vs 623ms respectively. It is useful directional evidence but is a provider-competitor-authored result, not a SchemaRouter reproduction. Microsoft's mixed-provider latency headline is also not a matched-serving comparison. Independent leaderboards have disagreed on the overall model ordering.

This intake changes neither core execution code nor promotion targets. The corresponding issue #1324 remains open until the concrete transport is verified against an authentic provider response, the model identity is recorded, and genuinely disjoint supported/unsupported results are obtained without retrospective threshold tuning. This doc does not trigger network calls, billing or an API deployment.