Prove · measurement
What can be compared, and what cannot
Every figure elsewhere on this platform is a figure as at one instant. This page asks the narrower question of which of them has an earlier figure stored behind it. 7 of the 9 series examined do. The rest either hold nothing earlier at all, or hold something that looks comparable and is not, and both are shown below rather than left to be found.
5
Quarters of stored history
on the two ledgers that keep a calendar, ending in 2026-Q3, still open
363
Rows carrying their own before
of 478 examined across two tables, each a single earlier reading
$0
Verified in the open quarter
against $191.1m claimed — verification happens after a quarter closes
2
Places the record disagrees with itself
two fields in the same table answering the same question differently
Has a stored before
5 quarters, on the only calendar the platform keeps
The two ledgers that keep a calendar keep the same one, five quarters ending in the quarter now open. The open quarter is not finished, so its figures are lower than a closed quarter for a reason that has nothing to do with performance, and the page marks it rather than hiding it.
Benefit claimed
What each of the 131 lines in a quarter says it saved, before anybody looked at it.
2025-Q3
$154.3m
131 lines
2025-Q4
$224.1m
131 lines
2026-Q1
$286.6m
131 lines
2026-Q2
$341.7m
131 lines
2026-Q3 · open
$191.1m
131 lines
Benefit realized
What the same lines report actually landed, still on the say-so of the line owner.
2025-Q3
$128.2m
131 lines
2025-Q4
$186.2m
131 lines
2026-Q1
$235.0m
131 lines
2026-Q2
$282.7m
131 lines
2026-Q3 · open
$156.2m
131 lines
Benefit verified
The part of realized that carries a verified amount on the line.
2025-Q3
$50.3m
131 lines
2025-Q4
$83.2m
131 lines
2026-Q1
$94.6m
131 lines
2026-Q2
$124.1m
131 lines
2026-Q3 · open
$0
131 lines
Attributed spend
What the cost centers were charged in the quarter, across 54 records a quarter.
2025-Q3
$63.0m
54 lines
2025-Q4
$74.6m
54 lines
2026-Q1
$84.8m
54 lines
2026-Q2
$92.8m
54 lines
2026-Q3 · open
$69.1m
54 lines
Capacity plans against what arrived
8 plans across 8 named periods around a peak. 48 of the 64 rows hold both a forecast and what actually arrived, which makes this the one place on the platform where a prediction can be scored against an outcome.
W-5
63,775
forecast
70,559
actual
W-4
67,635
forecast
63,326
actual
W-3
75,940
forecast
84,092
actual
W-2
88,437
forecast
82,772
actual
W-1
117,861
forecast
130,383
actual
Peak week
227,116
forecast
213,246
actual
W+1
140,530
forecast
not written
actual
W+2
72,035
forecast
not written
actual
Across the 48 scored rows the plans called 640,764 units and 644,378 arrived, a miss of 0.6% in aggregate. That figure flatters the plans: 21 of the 48 rows came in over forecast and 21 under, so the errors cancel in the total while only 6 of the scored rows landed on the forecast exactly. An aggregate accuracy figure is the wrong way to read this table and it is the figure a summary would print.
Has a before, but no date
363 rows that carry their own earlier reading
Two tables store a before against each row instead of against a date. That is enough to say a row moved and not enough to say when, how many times, or whether it moved back in between. A row-level prior is a single earlier reading, not a trend, and treating it as one would be a mistake this page will not make for you.
Evaluation runs against their own last score
Each run stores the score the same suite returned last time it was run against the same agent.
121
moved up
102
moved down
122
did not move
345 of 460 rows carry a prior · 83.8 → 86.9 mean score +3.7%
The three directions land within a few runs of each other, which is what a set of scores drawn around a mean looks like rather than a program improving. The mean has barely moved. Nothing here says a suite got better; it says the same suite returned a different number.
Exception classes against their own prior ninety days
Each class stores what it counted in the last ninety days and what it counted in the ninety before that.
11
moved up
7
moved down
0
did not move
18 of 18 rows carry a prior · 9,133 → 9,389 cases +2.8%
+86.4% Output distribution drifted from the evaluated baseline · 22 → 41
+44.3% Work completed outside the platform entirely · 61 → 88
-31.9% Escalation with no one on the other side · 141 → 96
This is the only table on the platform where every row carries its own before, and the totals move by a few percent while individual classes move by twenty. An estate-level exception count that looks steady is hiding classes going in both directions underneath it.
Looks comparable and is not
2 series that would produce a confident wrong answer
The two series below carry a field called period and cannot be compared across it. They are shown because leaving them out would let somebody find them later and compare them anyway.
Service level breaches by period
looks — Every breach carries a period, there are 6 of them, and they sort into a sensible-looking order.
is — They sit on two different calendars. 3 of the labels are quarters and 3 are months, and no field records which calendar a breach was filed against. A quarter cannot be set against a month, three months cannot be assumed to make one of these quarters, and adding the two together would double-count any breach that appears on both. The page shows the labels and refuses the arithmetic.
Break conditions and their last value
looks — 209 of 627 break conditions store a last value, which reads like a reading you could watch move.
is — It is one reading. The field is overwritten each time the condition is evaluated and nothing keeps what it held before, so there is no earlier value to compare it against and no record of how many times it has changed. A condition that has crossed its threshold and come back looks identical to one that has never moved.
Where the record disagrees with itself
Two fields, one table, different answers
These are not defects in the page. They are places where two fields in the same record answer the same question differently, and the platform has never had anything that put the two answers next to each other. Nothing here has been corrected in the data.
An exception class whose label and its own counts point different ways
Exception TaxonomyThe stored direction reads up on 10 classes, down on 7 and flat on 1.
The two counts on the same rows read up on 11, down on 7 and flat on 0.
1 of the 18 classes disagree with themselves. The direction is a stored word, written once; the counts are two stored numbers. Nothing recomputes the word when the numbers change, and no check anywhere on the platform compares the two, so the label is what gets read on the exception page and the numbers are what get read here.
Model confidence below the acting threshold is labeled flat while its counts move from 1,198 to 1,247, a change of 4.1 percent.
Benefit lines that are verified by one field and not by the other
Benefit Ledger190 of the 655 lines carry a verified amount above zero.
23 lines carry the status verified.
Every one of the 23 sits inside the 190, so this is not two populations but one inside the other. The 167 in between have had an amount written against them by somebody and have not been signed off, which is exactly the state that gets read as verified by anyone summing the column. The platform publishes the stricter of the two figures everywhere it can, and this is the only page that shows both.
A sum of the verified column returns money for 190 lines. A count of the verified status returns 23. Both are true of the same table.
Actions
What is waiting on a person
Nothing on this page can be acted on by the platform. Every item below is a question about the record itself, and each one is answered by a person going and looking.
evaluation runs with no earlier score to sit against
345 of 460 runs store what the same suite returned last time. The remainder are first runs or lost their prior, and a run with no before cannot say whether the change that triggered it helped.
planned periods that never had an outcome written back
48 of the 64 rows across 8 capacity plans hold both a forecast and what actually arrived. The rest hold a forecast and nothing, so the plan cannot be scored on them.
places where two fields in the same table give different answers
A stored direction that nothing recomputes when the counts beneath it change, and a verified amount written on far more benefit lines than carry the verified status. Neither field is wrong on its own. They are only wrong together, and no job anywhere on this platform compares them.
series that carry a period and cannot be compared across it
Breaches filed against two different calendars, and a last value that is overwritten each time it is read. Both are one query away from producing a comparison that looks correct and is not.
Operations
What this desk is allowed to start
A surface that only reports is not operable. This is the work this page can set in motion, and the bound it runs into.
Trigger and bound
This page reads seven tables and writes to none of them. It does not snapshot, it does not schedule, and it holds no history of its own: where a comparison is possible it is possible because a seed wrote two figures, and where it is not the page says so instead of interpolating one. Nothing here can be set as a target and nothing raises an alert.
Live observability
What the record shows right now
The 478 rows examined across the two tables that store a prior against each row, by which way the row moved. Direction only: none of these carries a date, so none of them is a trend.
Current distribution
478 rows
Is policy and strategy coming to fruition
Whether the written intent is holding here
Not yet. The platform can say what a number is now. On most of its surfaces it cannot say what the same number was.
Not holding on the record
The intent behind every count on this platform is that somebody watches it move. That is possible in 7 places out of the 9 examined here, and only because the data that seeded them happened to include an earlier figure. There is no snapshot table, no scheduled capture and no record of what any tile read yesterday, so a page can go from green to red and back with nothing anywhere noticing. The 5 quarters of benefit and spend are the only genuine calendar in the platform, and the newest of them, 2026-Q3, holds no verified value at all. Until something writes down what the estate read each week, comparison stays a property of six tables rather than of the platform.
Why most of this platform is not on this page
A comparison is only possible where the platform already stored an earlier figure. Nothing on this page is recomputed from history, because no history is kept: there is no snapshot table, no nightly capture and no audit of what any count read yesterday. Every series below exists because the seed that wrote it happened to write more than one period, and the ones that are missing are missing for the same reason.