The Outage Map — Branch & ATM Network Operations Center

A utility can tell you
how many customers are dark.
A bank cannot.

Every fault console knows a machine is down. None of them knows who is stranded by it. This one does — and it prints the subtraction that makes the figure honest: how many working devices were netted off before the headline was written.

California reference implementation · 5,384 measured offices · 595 assertions green On-prem · keyless basemaps · no runtime call to any public host

© 2026 Tabaqat · Built on Strata — sovereign geospatial applications. The estate and the population are measured and published. Device telemetry, cash, engineers and the whole ATM class are GENERATED (seed 20260818) and never mixed into a measured figure.

The decision this drives

Remote-fix, or roll an engineer — in the next hour.

That decision is made many times a day by one desk, against a service clock that is a contract rather than a rule. It is made today on a device list that knows the machine is down and knows nothing about the people it served.

So the app ranks the estate by who is cut off, not by how many pins are red. Two failures with identical fault codes are not the same incident: one strands nobody, because every tract around it keeps an alternative inside the same band; the other is the only cash access forty-seven thousand people have.

Buyer: head of the retail / self-service network Daily user: the operations desk Reads over the shoulder: the vendor-oversight owner

Not a compliance product. Nothing here is aimed at a risk, credit, ESG or CRA function — availability is the buyer's number, engineer visits are their cost, and the customer at a dead machine is their exposure.

170of California's 5,384 full-service physical offices — 3.2 % — are the sole cash access inside the published band for at least one census tract

1,486,347 people across 351 tracts depend on exactly one office. The largest single one holds 47,881.

Measured from the published federal branch register and the published tract population — not from telemetry. Only which device is offline at this second is generated.

The finding that shaped the app

Two industries answer half of this question each, and neither one crosses over.

Consoles: telemetry without geography

Excellent fault systems with a pin map bolted on. The map renders the device list; it never asks who lives around the device. The nearest thing to an impact figure counts failed transactions at the machine — by construction, only people already standing in front of it.

Location tools: geography without today

Site selection, trade areas, consolidation modelling. One public dashboard performs this app's exact tract-and-radius arithmetic — annually, on branches only, with ATMs explicitly excluded, and only for permanent presence.

Electricity solved it decades ago

An outage management system fuses SCADA, GIS and the customer register to report the number of customers impacted and to prioritise restoration. That is the shape. This is the banking version of that instrument.

One jurisdiction has already made reachable-population-after-a-closure a regulated calculation — the UK, on a twelve-week clock, for announced or sustained changes. Nobody, anywhere, computes it for a machine that failed at 06:40 this morning.

What actually ships — read page by page, 2026-08-18

Everyone holds half the instrument. Nobody holds both halves.

Category	What the product pages actually claim	Cadence	Turns an outage into people?
ATM / self-service consoles	availability, fault diagnosis, remote fix, dispatch — and a street-level state-of-health map	real time	No — status pins only
Transaction observability	in-flight capture at every network hop; failure and slowdown diagnosis in milliseconds	real time	No map at all
Location intelligence for banking	site selection, trade areas, whitespace, branch consolidation and relocation modelling	strategic	No — a planned network
The published desert dashboard	the identical tract-and-radius arithmetic on 2 / 5 / 10-mile bands	annual, branches only	Yes — but never for a failure
The UK cash-access regime	percentage of people within 1 mi urban / 3 mi rural of the remaining services; remedies inside 3 months	12 weeks	Yes — and not in the US
This application	people out of reach right now, per incident, net of working same-class devices in the same band	the refresh cycle	Yes — for a device that failed this morning

The sharpest finding of the study fits in one sentence: the systems that know the machine is down cannot draw the population, and the systems that can draw the population run once a year over a network nobody has broken yet.

The real incumbent, as always, is a spreadsheet assembled after the fact — which is the wrong artefact for a one-hour decision.

There is no availability regulator — but three rules shape the screen

What binds is not the estate. It is the vocabulary.

No standard definition of availability

Three incompatible conventions with three different expectations — network (99 %+), any-function (98–99 %), essential functions (96–98 %). A bare "97.4 %" on a wallboard is not defensible. Every availability figure here carries the convention that produced it, on screen and in every export.

Permanent removals are governed; failures are not

A branch closing needs 90 days' notice to the agency and to customers. So removed and planned-closure are first-class device states carrying a statutory date — and are never counted as an outage. 144 measured statutory removals ship as their own class.

36 hours is the one hard federal deadline

A notification incident that materially disrupts operations must reach the primary federal regulator within 36 hours. A switch or vendor failure taking out a region is squarely in scope. So the outage record has to be evidence-grade, not a wallboard that repaints and forgets.

in servicedispenser, card reader and network all up

degradedrunning, but not on the printed convention

out of servicethe only state that produces an impact figure

removed (statutory)measured and dated — never an outage

ATM class — GENERATEDits own hue, named in the legend, never summed with the measured class

Colour is never the only channel. Every state is carried by colour and marker size and a legend row naming it in words, and every popup prints the state as text — because the style compiler renders every marker as a circle, so shape cannot carry meaning here.

The application, first paint

A wall-capable board that states its own terms before it states a number.

The Branch and ATM Network Operations Center at first paint: a five-KPI board over a boxed centre map of California with flanking columns, the availability gauge and fault chart at left, the incident ticker and device detail at right, and a persistent notice bar naming every convention in force.

Shipped build, 2026-08-20 — a browser screenshot taken by the automated driver, not a mock-up.

The silhouette is deliberate. A full-width KPI board over a boxed centre map with flanking columns — not a map with one side rail, and not a table over a demoted map. Both of those are other products' silhouettes.
The notice bar is permanent, not a toast. The measured register and its index date, the generated telemetry and its seed, the band basis, the availability convention, the classifier assumption, the 124 weakly-positioned offices.
The headline never appears alone. 22,375 people out of reach — and beside it, 41,907 working devices netted off, 7 tracts, 67 devices offline.
Two classes, two figures, no total. Branch reach is measured; ATM reach is generated. Both baselines are printed separately and no combined number is offered.
Three pages, one map instance — Pulse · Dispatch · Cash & Plan — each with its own five-KPI strip and left column.

The number that is new

People out of reach — net of the alternative.

For every device offline at this moment: the population of the tracts whose only cash-access point of that class inside the published band is that device, minus every working same-class device inside the same band. Recomputed on the refresh cycle, attributed per incident, summable across an area outage.

band from the tract's own flag — 2 mi urban / 5 mi mixed / 10 mi rural
straight-line, labelled as such — not driven distance
tracts whose sole access is the offline device → 10 tracts
working same-class devices inside the same band → 96 netted off
= 47,881 people out of reach, from one office

The subtraction is the integrity of the number. Without it, this is a population map with an outage painted over it. So the count of devices netted off sits beside the headline, always — a design that reported people inside the band of an offline device would have printed 157,019 for the first row below and roughly 33 million for the last. Both would be wrong.

Verified scenario tableone fixed convention

Palisades + Altadena — 8 offices dark0

38 tracts touched · every one keeps a working alternative inside its band — a mean of 42.9 of them · 157,019 people covered

Spring Valley, San Diego — 1 office47,881

10 tracts stranded · 10 more touched but covered, 42,271 people · the largest single point of failure in the state

Imperial County — all 9 offices157,230

34 tracts newly out · nothing covered — the county has no alternative to fall back on

One institution statewide — 754 offices211,767

only 49 tracts newly out, because 33,156,461 people across 7,586 touched tracts keep an alternative

Baseline before any outage: 317 tracts / 1,149,470 people already beyond the band — reproduced exactly against an independent reading of the same published convention. The pipeline is verified, not merely run.

Eight offices dark strand nobody; one office strands forty-seven thousand. Fault-code severity and customer impact are not the same ordering — and only one of them is the buyer's.

The signature interaction

A reading that changes brings its own view.

The Spring Valley drill in force: the headline reads 47,881 people out of reach with 96 working devices netted off, the map has flown to San Diego County with the stranded tracts painted red, and an amber DRILL IN FORCE notice states that the live feed is excluded.

A row in the incident ticker → the map flies to that device, flashes it, opens its popup and fills the right-hand detail. A second click on the same row releases: selection cleared, popup closed, detail back to its empty text.

Choosing a drill once moved the headline to another county and left the map where it was — so the consequence the number described was off screen. The map now fits to the drill's devices and to the tracts they strand.

A drill is computed in isolation from the live feed — and says so in the KPI label, in a chip and in a persistent amber notice. Unioning them mixed generated failures into the answer, and Palisades read 22,375 instead of 0.
The legend filters the layer, it does not fade it — and every row keeps its denominator, n of N.
Esc unwinds one thing at a time — modal, drawer, isolate, hidden classes, selection, chart filter, scope.
Every active filter is a chip with an ×, and the deep link carries page · class · county · drill · device back into a cold load.

53 behaviours specified and tested

The second output — and it costs one export

The record an examiner can read 36 hours later.

A wallboard repaints and forgets. The federal notification rule does not: what went down, when, for how long, over how much of the estate — and whom it reached. That last column is the one no fault console can produce, and it is why this artefact is worth more than the dispatch it sits beside.

Every printed caption travels with the figures — the availability convention, the band basis, the population vintage, the classifier assumption, the straight-line disclaimer.
The zero is printed as a zero. This incident stranded nobody: 576 working same-class devices were netted off, and 42 touched tracts kept an alternative. An impact tool that only ever reports impact is a tool nobody believes the second time.
The stranded-tract list downloads as CSV, carrying the same captions.
Dispatch is propose-and-acknowledge — it writes nothing server-side, and raises a toast that says so.

The evidence-grade outage record: a monospaced document naming what went down, when, how much of the estate, whom it reached, the reach standard and its assumption, the nearest engineer's straight-line distance, and the notices that qualify every figure.

The outage record for a single branch incident, as shipped. No route line is drawn on a road network anywhere in this app — not on the engineer leg, not on the cash leg. Sending engineer-to-site pairs to a public router publishes where a cash machine has failed and who is travelling to it. That is a security position, not a rate-limit convenience.

Proof, not promises

Built, driven and measured — 2026-08-20.

595assertions green — 224 live · 232 offline · 139 in real headless Chrome

53behaviours specified, each mapped to the suite that exercises it — the floor is 3

6.10worst informational contrast ratio, in both light and dark — the floor is 4.5

5,384measured offices · 9,129 tracts · 133,627 in-band tract-to-office pairs, all computed at build time

Contrast is measured, not parsed

The browser suite repaints nine live surfaces in both themes into a 1×1 canvas over the ancestor stack actually composited behind them. Parsing a computed colour string is wrong the moment a colour is computed — and a contrast check that is itself wrong is worse than none.

The classifier moves the baseline 3×

The same arithmetic under a county-level metro classifier gives 3,568,904 people beyond the band against 1,149,470. Both are defensible readings of one published convention. The app fixes one, prints it, and offers no control that changes it — a desk that could move the standard could move its own impact number.

The recipe was corrected by the data

Its step B.3 could not be followed: none of the 152 year-over-year removals remains in the current register. They ship instead as a separate measured removed class from their own source, never counted as an outage. The alternative was to invent a state the data cannot support.

Reading the figures: the branch estate, the census geography and the population are real, published and probed. Device status, uptime, module faults, cash levels, engineer positions and the entire ATM class are GENERATED (seed 20260818) — there is no public register of US bank ATMs — and are labelled as such on every screen, in every popup and in every export.

Stated as boundaries, not caveats

What this application will not do.

An analytics layer, never a system of record. The incident of record stays in the bank's fleet and ticketing systems. A persistent footer says so on every page.
Near-real-time by refresh, not streaming. A 120-second cycle with an as-of stamp that turns amber past two intervals. No transaction stream, and no claim of one.
No write path. Dispatch proposes and acknowledges locally. A real write-back needs a writable, authenticated backend, and none is named.
Distance, not drive-time. The band is a straight-line radius labelled as such on screen, in the popup, in the legend and in every export. A drive-time claim needs a self-hosted isochrone endpoint inside the perimeter.

Not a banking-desert count. The published desert definition counts branches only and excludes ATMs. This counts cash-access devices — a different and deliberately stated thing.
124 offices carry a position too weak to draw a band around — 63 unmatched, 58 unstated in the register's own vocabulary, 3 postal. Flagged in the popup, never silently treated as located.
Population estimates people, not customers, on a 2020 count and 2020 tract geography. The app says people, and prints the vintage.
English only — one locale, no RTL surface. Recorded rather than designed, because none of the US data substrate serves the market that would need it.

And one more, said plainly: California is where the data is verifiable. It is not where the buyer is. Reading a Californian figure here as a statement about another estate is a misreading — and the app is built to be hard to misread that way.

Point it at your estate

One device register.
One population layer.

Your estate with coordinates and device class, your status feed, your service-contract clock, and a population layer for your market. Two weeks, and this board reads your network instead of a generated one.

Swappable by configuration, not by code: the reporting geography, the population source, the band convention and the availability definition. No distance, label, currency or threshold is hardcoded to a US rule — the reach arithmetic is a method the app renders, not a list of US numbers it ships.

tabaqat.net → Solutions → Financial Services info@tabaqat.net

© 2026 Tabaqat · Built on Strata. Reference implementation over the California branch estate; telemetry, cash, engineers and the ATM class GENERATED (seed 20260818). Analytics only — not a system of record, and not an incident-management system.

1 / 12