Expound/Platform/Accepted Work

Accepted Work, and what it cost

Accepted Work is work that actually counts: evidence-cleared, and accepted by someone with standing. A disposition is what happened to one piece of work (accepted, refused, escalated or abandoned), and cost-to-disposition is everything it took to get there, including the attempts that were refused. Divide one by the other and you have the measure that still means something when producing output costs almost nothing.

The unit

Count what was accepted, not what was produced.

Accepted Work identifies evidence-cleared business outcomes delivered by AI systems. Accepted Work per Dollar relates their governed value to the full cost of all attempts; refused or incomplete attempts remain in the denominator and add no accepted value.

Failed or incomplete attempts still cost money. A route that produces more output which is then refused is a more expensive route, and a measure counting tokens cannot see that.

Produced≠Accepted

Throughput≠Accepted Work per Dollar

Neither half of the Model Router creates Accepted Work: not Governed Multi-Route Selection (GMRS), which selects routes, and not Constrained Policy Reinforcement Learning (CP-RL), which ranks them. Only evidence-cleared work can become Accepted Work, and that determination is the kernel’s.

Refusal

A legitimate outcome, not a failure to deliver.

An early, well-argued no closes the item as a real outcome. The cheapest disposition is often the refusal you got early, and accepted work that later has to be unwound is the most expensive one.

This is also why the discipline publishes its friction. On a legitimate-work battery, 90 of 90 correct-but-unevidenced answers were refused. Those answers were correct. They were refused because the evidence required to rely on them was not there. That is the designed cost of the discipline, published alongside the benefit.

Correct≠Relied on

Why cloud-spend numbers are not enough

Spend tells you what was consumed.

It does not tell you what was accepted. Two routes can consume identical resources and produce entirely different quantities of work that is permitted to count.

A measure built on inputs cannot distinguish them, which is why the denominator has to carry every attempt and the numerator has to be gated by a verdict.

The denominator

Everything it took to reach any disposition at all: fully loaded, including the attempts that were refused.

The numerator

Only evidence-cleared outcomes. Gated by the verdict, not by whether something was produced.

What this changes

Spend follows what the decision actually requires, and no score buys its way around a hard constraint.