A decision model's confidence is not a spend cap: Sume's real gates

No. A confidence score says how sure a model is about an answer. It does not limit what the next call can cost. The Strands Decider model card describes calibrated confidence scores, and it also lists limits: calibration is one temperature per primitive, and the model inherits label noise from public training data. Treat the number as advice and let the Sume gates do the limiting.

Which gate stops what

Each Sume gate stops a different failure. The decision model is the only row in the table that is advisory.

Why the cap is yours to set

On hosted MCP, max_spend_usd is optional, and Sume enforces it only when you provide it. A flow that trusts a high confidence score and omits it has no ceiling left except the wallet and admission. Send it on every paid call, and size it from the dry run, not from the model's certainty.

On Agent Completions the cap is mandatory because an unattended agent has tools and access to your generation wallet. On a schedule, the API clamps a per-run number to min(request, schedule cap): a request can lower the cap and can never raise it.

Use the score for triage, not for limits

Log the decision next to the receipt so that you can compare the two later. A decision with a high score on a run that hit its cap is worth a look. A decision with a low score on a run that finished cheaply is worth a look as well.