> Source: https://vitna.ai/docs/routing/outcome
> Raw Markdown: https://vitna.ai/docs/routing/outcome.md
> Corpus index: https://vitna.ai/llms.txt

# Did the decision do what it said

A decision predicts. An [imported usage export](usage-import.md) observes. This
is the two of them side by side, which the record could always answer and never
had to.

The comparison is read at `/console`, which counts the checks and the
shortfalls, and in full on a workload's own page.

## Where the prediction comes from

Not from a stored copy. A decision keeps the exact `witness.price_period` and
`aperture.evidence` ids it read, in `inputs`, and those rows are append-only, so
the figures it was taken on are recovered by running the same derivation over
the same rows. That is the property
[a decision](../concepts/decision.md) is built to have, and this is the first
thing to spend it: a stored figure could drift from the rows it came from, and a
regenerated one cannot.

Two consequences follow, and both are reported rather than hidden.

A decision somebody **entered by hand** names no prices or measurements. Most of
the entered history is like this. Its own figures cannot be recovered, so it has
no comparison, and the console says that rather than showing an empty panel.

A decision whose derivation **no longer runs**, because the profile it recorded
spends a dimension the pinned tape does not carry, is a refusal with the reason
on it rather than a figure.

## The three things it is careful about

**It compares raw cost per unit, not the stored headline.** A decision's
headline saving is quality-adjusted: it divides each cost by a success rate a
measurement supplied, so it is a saving per SUCCESSFUL unit. An export carries
no success rate. Comparing the stored headline against an observed cost per unit
would be comparing two different quantities, and the difference would read as a
shortfall that nothing about the deployment caused. So both sides are raw cost
per unit, derived from the same two figures the decision stored, and the
quality-adjusted headline is reported beside them as what it was.

**Both sides are priced at what the decision read.** The pinned periods do not
move, so the mix is the only thing that can change the observed figure, which is
the question: whether the traffic went where it was told. It follows that this
says nothing about what the work cost at the market's later prices. That is a
different question and it needs the window's own periods.

**A figure says how much of the window it covers.** Rows need not sum to
`window_units`: a model an export names that the registry does not carry gets no
row and its units still count in the denominator, and a sku outside the two the
decision priced has no cost here either. Both are counted as unpriced, excluded
from the blend rather than counted at zero, and disclosed. A blended cost over
60 percent of a window presented as the cost of the window is the kind of figure
this record exists not to print.

## What it does not answer

**Quality.** Whether it held after the change is a measurement somebody enters
with `aperture:evidence:add`, not something an export can observe. Vitna runs no
evaluation and that boundary does not move.

**Not cost only, since the evidence loop.** A reading pushed by a runtime now
also carries how the units were reached, what was refused and what failed, the
tokens the provider reported, and latency at the middle and the tail. Those come
back per workload and are printed beside the confirmation, so "the decision was
applied" and "the work went well" stopped being the same sentence. A reading an
operator typed from an export carries none of them, and the page says so rather
than printing a zero.

And it is a reading of one window, the newest one imported. A workload with no
export has no comparison, which is not the same as a decision that failed.

## Reading a shortfall

A shortfall is usually the mix. A window where seven units in ten went to the
decided sku realises seven tenths of the predicted saving, and the sentence
under the figures says so with the share that caused it. That is a deployment
that is partly done rather than a decision that was wrong, and the two want
different responses.

A window that comes in **better** than predicted has one explanation with two
priced skus: the decision moved to the dearer of them, which a workload run for
quality does, and the cheaper incumbent is still serving part of the window. The
saving is not a win, it is the change not having finished.
