How to read a number on any surface
Every figure travels with chips saying what it came from and what it excludes. Learn them once and every screen becomes readable.
On this page
A dashboard figure is only useful if you know what it is a figure of. $61.25 over how many
rows, priced from which table, excluding what? On these screens that context travels with the
number rather than living in a footnote, and it uses the same six devices everywhere.
Learn them once and every surface in the product becomes readable, including ones you have never opened. This page is the vocabulary; the sections that own each concept are linked from each part.
The row-count chip
Beside almost every panel sits a small chip reporting the population the figure was computed over. On the demo ledger it reads:
46 rows 2 est 2 unpricedThree separate facts. 46 rows were evaluated. 2 of them carried token counts the provider did not report, so their cost is an estimate. 2 carried a model with no entry in the price table, so they contribute nothing to any dollar total and are counted here instead.
An empty panel shows the same chip reading 0 rows, which is why an empty dashboard is
readable at all. A chip that reports truncated is telling you the read did not finish, and
the figure beside it is over less than it should be.
EST on a dollar
A dollar rendered with an EST flag was computed from token counts that were estimated rather
than reported by the provider. The flag travels with the value through every aggregation, so a
total containing one estimated row is itself flagged.
That is a different statement from an approximation in the arithmetic. The price table and the multiplication are exact; what is uncertain is the input. Where a cost number comes from covers when tokens are estimated and how to reduce it.
UNPRICED instead of a dollar
When no price covers a row, the surface renders UNPRICED where a figure would go. It does
not render $0.00, and the row does not silently join a total.
↑ Excludes 6 unpriced rows from all totals — spend the price catalogue cannot cover is counted and disclosed, never quietly folded into the total as a zero. The total itself carries EST because some of its rows are estimated.
This is the rule that shows up most often in practice, because a model the price catalogue has not caught up with is normal rather than exceptional. The Pricing surface exists to triage those rows, and override a price is how you resolve one. The reasoning behind refusing the zero is unpriced is not zero.
The pricing snapshot, and mixed ones
The header carries the pricing snapshot the instance is costing against:
litellm@2026-07-14#179affbSource, fetch date, and a short hash of the upstream table. Every priced row is stamped with the snapshot that priced it, which is what makes an old row explainable after the table moves on.
When a panel’s rows span more than one snapshot it says so, with a Mixed snapshot chip
reading totals span 2 pricing snapshots. That is not an error. It is the disclosure that a
total was assembled under two price tables, which matters when you are comparing periods.
Pricing keys and snapshots owns the mechanism.
Verdict chips
Anywhere the product reaches a conclusion about routing quality, it renders a chip carrying the server’s own verdict string. The vocabulary is closed and the chip does not translate it: the text is the exact value the API sent, and the component only picks a colour.
| Verdict | What it says |
|---|---|
BEAT |
the router beat the comparison on this evidence |
NOT_BEAT |
it did not |
ON_FRONTIER |
it sits on the cost-quality frontier rather than above or below it |
CANNOT_MEASURE |
there is not enough labelled evidence to judge |
CANNOT_MEASURE is the one people misread, so it is worth stating plainly: it is an answer,
not a failure. It arrives with the reason and the shortfall, as in insufficient labels (0 of 8) to judge routing, and with what to do about it. A dashboard that turned that into a zero
or a blank would be claiming something it cannot support.
Because the chip renders verbatim, do not paraphrase one in a ticket or a caption. Quote it. The two verdict vocabularies in this product are not the same thing, and paraphrasing is how they get conflated.
Approximation, stated with its bound
Where a figure is approximate, the surface says so and publishes what makes it approximate rather than hedging. A limit’s burn panel carries:
Enforcement is approximate by construction.
burn excludes in-flight reservations and any teed rows dropped under load
tee-lagged rows: 3
dropped rows: 0
in-flight reserved: 1,024
estimate skew p95: 0.03Four counters you can act on, rather than an adverb. The same treatment appears on budget overruns, where the disclosed bound is computed from real lag and drop counters. Enforcement and its bounds owns the mechanism.
Every chart has its rows
Beside each chart is a View as table control that shows the exact rows behind the picture, and most marks are one click from becoming a filter on the Requests explorer with the window already applied. That is the affordance that makes a chart checkable: you are never more than one step from the rows that produced a shape.
Two more markings you will meet
The demo watermark. An instance running in demo mode carries a DEMO label in its header
and a diagonal DEMO DATA watermark across spend tiles. It is a property of the instance, not
of a panel, so a watermark on a panel of a non-demo instance means that panel is showing a
fixed example. Four currently do; what a new dashboard
shows lists them.
The promote-gate badge. An instance in route mode that lacks fresh shadow evidence wears
a route without shadow evidence badge in the header, and every audit row of the move carries
it too. It is computed from the live mode against the evidence, so it cannot be dismissed by
clicking. Routing owns the gate; the badge is what it looks like when
somebody has gone around it.
What this page does not cover
It teaches the markings, not the subjects. What a cost figure is computed from is Cost & attribution; what licenses a savings claim is the honesty model; what a quality verdict measures is Calibration & Evals. If a marking on your screen is not in the list above, the surface index will point you at the section that owns it.