Microsoft-Decision-1 vs Clef: which gate sits in front of Sume?

Pick Microsoft-Decision-1 when the thing you want to judge is text or JSON (see its Foundry page for the input limit). Pick Clef when the input can include images, or when you need a 65,536-token window. Both return typed answers (a yes/no probability, one chosen option, or an ordered score) instead of prose, and neither one is a spend control. In front of a paid Sume call, the decision model only decides whether to ask; Sume's own gates decide whether the call can cost anything.

Both vendors describe the same three question shapes, noul, choice and score, so a gate you write for one is easy to port to the other.

What each vendor page says

Numbers below come from the vendor pages named in the caption. Where a page did not state a number, the cell says so rather than borrowing one from a news site.

What neither one can do for you

A decision model returns a probability. It does not hold credits, and it cannot refuse a request on Sume's side. On hosted MCP, the controls that actually bound spend are idempotency_key, dry_run, and the optional max_spend_credits, and over REST a balance that cannot cover the request gets `402 insufficient_credits` before any provider work starts.

So treat the decision as a cheap filter that saves you a dry run on briefs that were never going to render. Keep the Sume gates on regardless.

• Use the decision model to answer: is this brief renderable, which of three formats fits, how urgent is this ticket.
• Use Sume to answer: what will it cost (dry_run, usage.credits), what is the ceiling (max_spend_credits), and is this a retry (Idempotency-Key).
• Log both. Sume returns a request id on errors; the decision model returns no rationale, so keep the state you sent.

A simple rule for choosing

If your gate reads the brief text or the JSON arguments of a pending tool call, Decision-1 is enough, since a normal brief is short. If the gate must look at a reference image or a frame, use Clef, whose page documents up to 4 images of 4 MiB each. If both fit, run the same ten labelled briefs through each and keep the one whose probabilities separate your good and bad briefs better. Microsoft's own page tells you to validate on labelled examples before you automate a gate.