See where the work stops. Before it starts.

What you get

Five things the diagnostics report.

Every answer on this page appears in a full example forecast — an invented customer, a real engine run.

One run, against the workflow and the information you give us.

Things we can tell you

1. Where the design itself is leaking — if it is. Some workflows are expensive because of the way they are built, not because of what they cost per token. Where yours is one of them, we name what is doing it and say what closes it. Often there is no leak at all.
2. What a spending limit costs you. A ladder: for each level of budget reduction, the cap per task that gets you there, how much delivered work it destroys, and the step where the work stops.
3. Which step costs the most. Where spend concentrates once repetition is counted — including how much of each step's bill is work being re-done.
4. How much review it takes. How often a person is pulled in, and how much of that is work coming back to them a second time.
5. What stops a forecast. Your workflow is compiled before anything is forecast, and two things stop it there: a step that can start unlimited work of its own, or a trigger that fires without a person. Either one means we cannot price it, and we will work it out with you.

Things to consider

1. Who checks the work is a downstream-impact decision, not a cost one. People and models behave differently at a review step. An AI validator is itself a model call, so it can be confidently wrong about work that was fine — where a human reviewer produces no artifact and can only miss or catch. We price the checkpoint you describe, and we do not assume the two are interchangeable.
2. Length compounds in both directions.
   • Outward, into the sweep. Every step you add multiplies the whole board. Every model is scored in every position, and each configuration is simulated 64 times over.
   • Downstream, into your bill. Every step inherits whatever came before it — a silent failure at step two that nobody inspects until step nine has contaminated seven steps of work, and you pay for every one of them.

   AI steps	Configurations	Runs (×64)
   2	196	13,000
   3	2,700	180,000
   4	38,000	2.5 million
   5	540,000	34 million
   6	7.5 million	480 million
   7	105 million	6.7 billion
   8	1.5 billion	94 billion

   Rounded, at the current catalog of fourteen models. Six is the limit for one forecast — larger flows are split into linked workflows and scoped separately.