How Sky Score works, in plain English

The full guide: what it is, what happens when you type a postcode, where the numbers come from, the Amazon pieces, the assistant, the checks, the money · 2026-09-21

What's in this guide

1 - What it is, and who it is for

A postcode goes in. A score from 0 to 10 comes out. The score says how liveable that spot is, built from eight things: aircraft noise, road noise, air quality, flood risk, house prices, crime, school results and transport.

Every one of the eight comes from an official dataset - the government departments and agencies that publish these things - and every number can be traced back to the file it came from. That is the whole idea: not opinion, but arithmetic over public data, with the working shown.

It has two front doors, because it has two kinds of user:

• People use the website: a map of a city's boroughs, coloured by score, with a search box. Type a postcode and a panel opens with the score, what's behind it, the nearest stations, recent sales, energy ratings and the closest GP. It covers ten city regions in England (London, Manchester, Birmingham, Leeds, Sheffield, Liverpool, Newcastle, Bristol, Leicester, Teesside) and New York.
• Businesses - a property portal, a lender, a relocation firm - use the API. An API is a way for one program to ask another for data, with no person and no screen in between. Their software sends a postcode; numbers come back; they show them wherever they like. The API covers two more cities than the website (Cardiff and Nottingham).

The website is the shop window; the API is the product. Both give the same answer for the same place, and an automatic check makes sure they keep agreeing.

2a - What happens when the page opens

Before you have typed anything, the page has already fetched everything it will need to draw the map and score a search. Here is that first second, measured on a fresh browser with nothing stored: thirteen requests, quiet after 1.3 seconds.

Two things that are not fetched at load, on purpose: the aircraft-noise readings and the station list. They are only needed once you search, so they are fetched the moment you start typing (2b, first line), not before. A page that fetches everything it might ever need is slow for everyone to serve the few who search.

2b - What happens when you type N1 7SX

Think of your browser as a person at a desk with a reference book. Here is the conversation it has, in the order it happens. The whole thing takes about a second.

your browser, on its own a free public service a small helper I run on Amazon

What ends up on the screen for N1 7SX

3 - Where the eight numbers come from

Each of the eight inputs is a real measurement, published by a named public body, turned into a number per borough by a script on my computer. Here is each one: what it measures, who publishes it, and how it becomes a mark out of ten.

4 - How eight numbers become one score

The eight inputs are first grouped into five components, each a mark out of ten. Then the five are blended with weights that depend on who is asking.

Weights: the same numbers, different people

A retired couple and a landlord care about different things, so the site offers buyer profiles. Each profile is just a different set of weights over the same five components. "Balanced" is the default. The weights always add up to 100%.

Notice that growth counts for nothing in every profile except Investor. A rising market is good news for a landlord and bad news for a first-time buyer, so it is not a "liveability" quality at all - it is only scored for the one person it helps.

A worked example: N1 7SX, Balanced

7.0 × 32% + 1.1 × 27% + 3.8 × 0% + 7.8 × 27% + 5.2 × 14% = 2.24 + 0.30 + 0 + 2.11 + 0.73 = 5.38, shown as 5.4. The sums are done with the unrounded components and rounded once at the end, so if you redo them from the rounded figures on the screen you can be out by up to a tenth. The methodology page says so, and a check makes sure that bound stays true.

Three design choices worth knowing

1. Affordability is national, on a curve. A borough's price is placed between the 5th and 95th percentile of every borough in the country, on a logarithmic scale - which is a way of saying the gap between £150,000 and £300,000 counts the same as the gap between £300,000 and £600,000. Before September 2026 it was compared only within its own city, which made the cheapest borough in a dear city look "affordable" (Barking scored 10) and the dearest in a cheap city look "unaffordable" (Stockton scored 0). Now Barking is 4.4 and Stockton 9.5, which is what the prices actually say.
2. Growth is after inflation, and against the whole country. Same reasoning: a city with four boroughs would always have one at the top and one at the bottom, whatever the numbers, so the comparison is national.
3. A missing input is left out, not faked. If a borough has no schools figure (Wales) or no flood data (New York), that component is dropped and the other weights are scaled up to fill the gap. The response says which weights were actually used. Nothing is ever filled in with an average and passed off as measured.

5 - What happens when a business calls

The website does the sums itself. A business cannot - its software has no table of numbers - so it asks the API, and the sums are done on Amazon. Same conversation format, different building.

Two things the business gets that the website does not: a machine-readable answer it can store or display anywhere, and a way to ask about many postcodes at once (on the paid plans). One thing it does not get: the four helpers - stations, sales, energy, NHS - which are for the website's panel.

6 - The assistant, and the forty lines

There is one place AI is used: an assistant on the API that explains a score in a sentence. Here is one question going through it - and the part of Sky Score most worth understanding.

And when the model goes beyond the data

It just discarded a true answer. That is the cost of a rule that cannot tell a number the model worked out from a number the model made up - and on a service about someone's home, the cost is worth paying. A warning next to a made-up number would still ship the number, and people read the number.

Why it is built this way

1. An instruction is not a control. The written instructions to the model say, in capitals, never invent a number. The third real question the assistant ever got was "what will prices be in 2030?" - and the model gave a figure. Instructions are requests. The forty lines are the control.
2. It is deliberately dumb. It does not understand the sentence; it only checks that every number in it was handed over. That is why it works with any model and needs no second AI to judge the first. A judge would be another instruction.
3. It had three holes, and nobody noticed for five weeks. For a while it ignored every whole number from 0 to 10 - the very scale it exists to protect. Each hole was closed by writing a test that failed first, then fixing the code. Which is the subject of Part 9.

7 - The Amazon pieces, and why each one is there

Sky Score runs on Amazon Web Services (AWS) - Amazon's business of renting out computers and storage by the minute. We rent nine kinds of thing. None of them is a computer that sits switched on all day. Each is described the same way: what it is, in everyday terms; what we use it for; and why this and not something else.

Two things we deliberately do not rent

1. No always-on computer. There is no server to log into. Every piece above either stores files or wakes up for a job.
2. No map service. See Part 1: our own outlines, drawn in the browser.

What got taken out, and why

In March there were five AI features on the website: a chat panel, an automatic AI summary on every postcode, photo analysis, document analysis and a generated report. In May all five were removed in one go. The main reason: an AI summary on top of a carefully checked score adds an answer that varies and is sometimes wrong. The second reason: they were costing about $80–115 a month in credits. What came back in August is the one narrow assistant in Part 6, for about 25p per thousand questions.

8 - How new data gets in

The numbers on the site are not fetched live from the government. They are computed on my computer, checked, and then published - on a rhythm set by whoever publishes each dataset.

The four steps of every roll

1. Download the official file from the publisher.
2. Compute each borough's number from it, with a script that is the same every time.
3. Check: a second mode of the same script re-derives the published numbers from the source and fails loudly if any has drifted. Several of these checks run before every change is saved, so a stale number cannot slip through.
4. Deploy, in a fixed order: the API first (because the site's copy of the numbers must never be ahead of the API's), then the site, then confirm from the live address that what is served matches what was written.

9 - The 57 checks

Before any change to Sky Score is saved, a checklist of 57 automatic checks runs. 45 must pass or the change is refused; 12 only warn. I call it "preflight". It takes about a quarter of an hour.

The rule that makes the checks worth having

A check that has never gone red has not been tested. Before a new check is trusted, the thing it guards is broken on purpose, the check is watched turning red, the break is undone, and the check is watched turning green. Only then does it count.

That rule exists because twelve checks in this project's history turned out to be green for months while checking nothing. Nobody switched them off. One piped its results into a command that always succeeds, so the whole checklist said PASS while running nothing. One scanned only the home page and only for the most severe faults, so every real fault was invisible. One had a minimum of "at least 17 files" the day after the 18th file was added - so it could no longer notice that one file going missing. Each decision was sensible on its own; together they made a check that could not fail. Asking "what would this check miss?" is how they were found.

10 - How it was built with a coding assistant

Most of the code was written over six months with a coding assistant called Claude Code - an AI that reads and writes the project's files under direction. Here is the working arrangement, because it is different from "the AI wrote it".

1. I own the judgement; it owns the typing. What to score, what to throw away, what counts as grounded, which datasets are trustworthy - mine. The code that carries those decisions out - mostly its.
2. The check comes first. I describe what must be true when the work is done; the assistant makes it true; the check decides whether it is. If we disagree about whether something works, neither of us settles it - the check does.
3. It cannot go live on its own. Putting a new version live needs my explicit go-ahead. On one occasion its own safety layer refused a production deploy, and I ran the step myself.
4. It keeps notes between sessions, and the notes must record checks, not facts. "The tests pass" is a fact, and it was false the day it was written. "Prove a check can fail before you trust it" is a check, and it is still true. About eighty such notes exist for this project; the useful ones are all of the second kind.

11 - What it costs, and what it charges

What it costs to run

Because nothing is switched on until asked, a quiet month costs very little: file storage is pennies, the workers and the index cards are paid per job and per lookup, and the copies-in-every-city are paid per visit. The one figure worth remembering is the assistant: about 25p per thousand questions. For contrast, the five AI features removed in May were costing $80–115 a month at modest traffic - most of it the bigger model they called.

What it charges

The data underneath is all published under the Open Government Licence, which allows commercial use as long as the source is credited - which is why every API answer carries its sources, and why the terms require customers to pass them on.

12 - Privacy, safety and the phone app

What leaves your device, and where it goes

• On the website: the postcode you typed goes to postcodes.io (to find it) and to the four helpers (for the extras). Nothing else. No account, no cookies; the visit counter is a cookieless one that keeps no personal data.
• Saved places ("favourites") are stored against a random token your device makes up, not a name or an email.
• Through the API: the postcode and the caller's pass. Through the assistant: the question and the score numbers go to the AI model in Virginia - the only thing that leaves the UK - and the answer comes back. Nothing personal is sent, and the model is not allowed to remember anything.
• Logs are kept 30 days, and a check confirms that is really what Amazon is doing.
• Sign-ups for an API pass keep an email address. A stricter "confirm by email first" path is built and switched off until the sending address is verified.

Keeping the door shut

• Every business request needs a pass; every pass has an allowance; every route has a speed limit, sized from measured traffic.
• The badge that other websites can embed is an image, not a script, so a portal's own security settings cannot block it and it cannot run anything on their page.
• Anything typed by a user is checked before it is used - a postcode must look like a postcode, a question is capped at 500 characters.

The phone app

It is the same page, three ways. On the web, as-is. As an "installable" website on Android and iPhone (Add to Home Screen), which also works offline once loaded. And as a native iPhone app on the App Store, which wraps the same page and adds one native feature: a "score where I am" button that uses the phone's location. Web changes go live at once; the App Store version needs a new build and Apple's review, so it is refreshed every few weeks rather than every day. An Android version is built but not yet in the Play Store.

13 - How to show Part 2b live in 30 seconds

1. Open skyscore.co.uk in Chrome. Press F12. A panel opens: click Network at the top of it, then the Fetch/XHR button so only data requests are shown. Click the small 🗑 icon to empty the list.
2. Type N1 7SX in the search box and press Enter.
3. Watch the list fill: three files, postcodes.io, then the four helpers landing together. Read them out. Then: "and notice there is no request for the score."
4. Tip: drag the panel to the right-hand side (the three dots in its corner → Dock side) so the map and the list are visible together. Zoom the page to 150% at most on a 1366-wide screen, or it switches to its phone layout.

Appendix - the nine requests as the browser writes them

Captured live on 2026-09-21 while searching SW11 1AA; the same nine appear for any UK postcode.

Words used above

Browser

Chrome, Safari, Edge - the program you view websites in. It does more than display pages; it can do sums and ask questions of other computers.

Request

One question the browser sends over the internet, and the one answer that comes back.

API

A way for one program to ask another for data, without a person or a screen in between. Sky Score's API: send a postcode, get numbers back.

API key / pass

A long code that identifies who is calling, so their usage can be counted.

Component

One of the five marks out of ten (Quiet Skies, Affordability, Growth, Liveability, Environment) that are blended into the score.

Weights / profile

How much each component counts. A profile is a named set of weights: Balanced, Family, Investor, and so on.

Percentile

Where a value sits in a line-up. The 95th percentile price is dearer than 95 out of every 100 boroughs.

Deploy

Put a new version live.

Roll

Bringing in a new edition of a dataset and recomputing everything that depends on it.

Check / gate

One of the 57 automatic tests that run before a change is saved. "Red" means it failed; "green" means it passed.

Region

One of Amazon's clusters of buildings. Sky Score uses London; the AI model is only in Virginia.

Open Government Licence

The licence most UK public data is published under: free to use, including commercially, if you say where it came from.

Sky Score · requests captured live and panel values for N1 7SX read from the live site, 2026-09-21. Weights from the scoring engine; prices from the pricing page; removal figures from the commit of 7 May 2026; data-roll rhythms from the operations runbook. Methodology version 5.2.