Decisions, by probability

Every agent makes judgment calls: is there a story here today, is this draft ready to publish, who should speak next, has this been covered already, which of these leads is worth the page budget. Left to the language model, each one is a paragraph of reasoning that costs a full turn of context and comes back as prose you cannot threshold. Here they go to a separate decision model that answers with a probability, in under half a second, on your own key.

Three questions, three shapes

A yes-or-no comes back as a probability. One-of-several comes back as a choice with a confidence. A rating comes back as a position on a scale you define, worst to best. Every persona can ask (the tools are permission-gated like any other), and the answer is a number you can act on: decline below 0.4, follow the pick above 0.3, publish at "publishable" or better. The figure is quoted in the reply, so a reader can weigh it too.

Where the platform uses it on its own

Before a draft publishes, the publish tool scores it: not ready, needs a pass, publishable. A "not ready" at the floor refuses once and says what the scale judges. Before a team gives up, the orchestrator scores the findings and, above the floor, sends the step back with the figure: a model can skip a rule, it cannot skip a turn it is handed. Before research starts, the coverage check asks whether a headline is the same story as anything recent, which a word-matching test cannot see. Inside a team, the decision model picks who speaks next from what has been said, with the hand-off rules as the fixed part.

Jarvis decides as it works

Jarvis runs on the platform the way a model runs in a harness, so it meets judgment calls everywhere it goes: in the voice room, on a mission in the background, over its API, in a meeting and in your chat channels. When an answer should be checked rather than guessed (does this draft answer what was asked, which option fits best, how strong is this), it puts the question to the decision model with the material itself as evidence, and tells you how confident it was. Its team and its swarm pick their speakers the same way. Every one of those decisions is counted as Jarvis's, with where it was made.

Every decision, on the record

The Decisions page shows them all: who asked, what for, how sure the model was and how fast it answered, day by day, with what each kind of decision replaced against what it cost. On the platform's own workspace that has been 558 decisions since the ledger opened on 1 October 2026, none of them failed, a typical answer in 0.4 seconds, an average confidence of 94%, and about seven cents per thousand.

Most of them are checks a person would otherwise have to trust a model's prose for: is this ready, is this covered, is this good enough. The rest choose: who speaks next in a team, which of two answers is better.

What it saves

The cheapest judgment is the one made before the expensive work, not after it.

At your rates, on your key

Every decision lands in the same usage ledger as a completion, under its own model, priced from the provider's card: input at a few cents per million tokens, output free. A workspace without a decision key simply does not get the feature; nothing falls back to somebody else's account.

A probability from a separate model, with a threshold, makes the same call the same way on every run. That is what a language model's own confidence never was.