Priority, Outcomes & Calibration

What This Covers

Priority is how you tell the platform what matters more on this work: keeping cost down, raising quality, or going faster. You express it in plain terms; the platform translates that into concrete routing and verification policy, then shows you what actually happened.

Three surfaces, in the order you meet them:

Surface What it is for
/platform/ai/assignmentsPriority & Models Set the everyday priority, plus advanced per-coworker guardrails
/platform/ai/priority/outcomesPriority — Outcomes See what each recent run actually did against what you asked for
The suggestion banner on that page A better-fitted default, proposed from your own run history

/platform/ai/priority now redirects to /platform/ai/assignments. The everyday priority and the advanced guardrails were merged onto one surface rather than living apart.

What It Is — and What It Deliberately Is Not

The priority control is a preference-to-policy compiler. You choose a posture; it produces explicit policy adjustments against the routing and decision contracts that already exist.

It is not a model picker, not a second router, and not a separate model registry. That distinction is load-bearing: a control that quietly ran its own routing beside the real one would let the screen and the system disagree. Instead the compiler feeds the existing routing call as defaults — every explicit setting, and the local-only sovereignty switch, still wins over it.

Presets are the primary control. The triangle itself is a fine-tune and visualization layer, colour-coded by balance, not the thing you must drag to get work done.

Two properties that make it safe

The Work Itself Can Raise the Bar

The newer half of this: priority is no longer only about your preference. The kind of work now sets floors the priority cannot trade away.

When work happens in a Workroom, the room’s collaboration shape is passed to the compiler, and three shapes carry a floor:

Kind of work Minimum quality tier Verification
Outward review — the action leaves the business under its own name Strong Deep
Approval sign-off — an accountable approver signs off on prepared evidence Strong Shallow
Consequential change — confirmed before it executes Strong

Two things about this table are deliberate:

So a marketing send inside an outward-review room runs at a strong tier with deep verification even if the platform priority leans toward cost. You did not have to remember to raise it.

How Governed Work Actually Runs explains where the shape comes from; My Work and Workrooms shows the per-room panel.

Reading the Outcomes View

/platform/ai/priority/outcomes answers the question a preference control usually dodges: did the platform actually do what I asked?

Each recent run is compared against the priority currently in force and marked:

The distinction that matters most is why a run deviated, and the view separates two causes that look identical in a log:

Collapsing those two into one “degraded” state is what makes most such dashboards useless. Here they are named apart.

The Calibration Suggestion

When enough runs accumulate, a banner at the top of the outcomes view proposes a better-fitted default. It suggests and never auto-applies — the change stays your decision.

It is conservative on purpose. Below five runs under the current priority it says so and stays quiet. Above that, the most actionable signal wins:

What it says What triggered it What it means
Runs are failing over often 30% or more fell back to a backup provider The tier you asked for may not be reliably available. Add capacity, or ease the quality floor.
This priority may be under-provisioned 30% or more landed below the requested tier, or failed verification Nudge the triangle toward Quality.
You may have room to save A quality-leaning priority where every run was clean A more balanced priority would likely still pass, while spending less.
This priority looks well-fitted None of the above Nothing stands out — delivering what you asked without obvious waste.

The third one is the interesting case, and the one most platforms never offer: the system telling you that you are over-buying. A control that only ever suggests spending more is a sales funnel, not a calibration loop.

Honest Limits