CleanZone Method
From 26 Million Data Points to a Shortlist of Five
Somewhere out there is a 25 km square of the planet that suits you better than the one you're standing on. The problem is not finding it — the problem is that there are roughly two million other squares competing for your attention, and a map you can pan and zoom is a terrible way to choose between them. The fix isn't a better map. It's a funnel.
At a glance
- Browsing a map rewards whatever's visually salient. Filtering rewards what actually matters.
- Hard disqualifiers go first — non-negotiables like excessive night noise, high radon potential, or floodplain membership remove the most ground for the least effort.
- Cheap, binary filters should run before expensive, gradient ones — order changes how much you compute, not just what survives.
- Only after disqualification do you rank the survivors on weighted priorities — that's a taste problem, not a safety problem.
- The shortlist is a start, not an answer: export to CSV and verify every candidate on the ground.
Why browsing doesn't scale
Pan-and-zoom is how most people evaluate a place, because it's how most maps are built to be used: one viewport, one story at a time. It works when you already know roughly where you want to live and you're comparing three suburbs. It fails completely when the honest answer to "where should I live" is "anywhere on a 25 km grid," because that grid — at global scale — runs into millions of cells. Nobody pans through two million anything. You need a mechanism that discards the overwhelming majority of the search space before you ever look at it, and does so on criteria you chose in advance rather than on whatever happens to be centred on the screen.
That mechanism is a funnel: a sequence of filters, each one strictly narrowing, applied in a deliberate order. It is the same logic a recruiter uses on a stack of CVs, or a radiologist uses on a screening population — reject the disqualified fast and cheap, then spend real judgement only on the survivors.
Browsing a map
Rewards the cell that's visible, colourful, or nearest the last place you clicked. No memory of what you ruled out an hour ago. Fatigue sets in around the twentieth candidate and standards quietly slip.
Filtering a grid
Every cell is judged against the same fixed thresholds, in the same order, every time. Fatigue doesn't affect the fifth candidate differently to the five-millionth. The shortlist is reproducible.
Stage one: hard disqualifiers
A disqualifier is a threshold you will not negotiate on, regardless of how attractive everything else about a cell looks. Good disqualifiers share three properties: they're binary (pass/fail, not a sliding scale), they're backed by a published reference value, and they're cheap to evaluate — usually a single field lookup against a fixed number. Three examples that hold up across most relocation briefs:
- Night noise. The World Health Organization's environmental noise guidelines recommend an average night-time level (
L_night) below 40 dB to avoid adverse health effects; anything mapped consistently above that on thenoisefield is an automatic exclusion for a light sleeper, not a point deduction. - Radon potential. EU Directive 2013/59/Euratom sets a national reference level of 300 Bq/m³ for indoor radon. A cell flagged high on
radondoesn't mean every house in it is dangerous — it means every house in it needs a long-term test before you'd sign anything, which for most people is disqualifying on its own. - Floodplain membership. A cell sitting inside a mapped 1-in-100-year floodplain carries insurance, resale, and structural consequences that no amount of nice weather offsets. It's a fixed-year hydrological delineation, not a spectrum — either the parcel is in the zone or it isn't.
Run your cheapest, most binary disqualifiers first — a single boolean field beats a computed composite every time. Filtering two million rows on in_floodplain = true is near-instant; filtering the same two million on a five-factor weighted score is not. Disqualify fast, rank slow, on a much smaller table.
Stage two: stack the layers, don't average them
A common mistake is to blend disqualifiers and preferences into one composite score from the start — say, averaging noise, radon, and "nice climate" into a single number per cell. That buries hard failures inside soft ones: a cell can score well overall while sitting on a floodplain, because three good numbers outweighed one bad one in the average. Layering avoids this by treating each disqualifier as an independent mask applied to the map, not a term in an equation. A cell has to clear every mask, not average across them.
Averaging is for preferences. Masking is for non-negotiables. If a threshold is the kind of thing you'd refuse to compromise on in a face-to-face viewing — a floodplain, a factory upwind, a motorway on the doorstep — it belongs in the mask stage, not in anyone's weighted score.
Stage three: rank what's left
Once the hard failures are gone, what remains is a genuine taste problem, and taste is where weights belong. Two buyers looking at the same 340,000 surviving cells can land on completely different shortlists because one weights commute distance heavily and the other weights climate stability heavily — and that's correct, not a bug. This is the stage to attach numeric weights to whichever public-standard fields matter to you (PM2.5 annual mean against the WHO 5 µg/m³ guideline, water hardness class, G-Tower/EMF density, aircraft overflight frequency) and sort.
The map doesn't need to get smaller. Your search does.
Stage four: export, then trust nothing until you've stood on it
A shortlist of five is a research output, not a decision. Export the surviving rows — the fields you filtered on, the fields you ranked on, and the raw source values behind each — to CSV, and treat that file as a to-do list, not a verdict. Public reference maps operate at a coarse resolution: a national radon-potential classification or a floodplain delineation describes the geological unit or the parcel, not the specific building on it. The grid narrows the search from continental to conversational. It doesn't replace the conversation.
That means, for each of the five: a site visit at a different time of day than your first look, a long-term measurement where the disqualifier was a modelled potential rather than a direct reading (radon above all — a three-month alpha-track test, not a two-day canister), and a conversation with someone who already lives there. The funnel's job was to stop you wasting that effort on 1,929,995 cells that were never going to work. It was never going to do the last mile for you.
Funnel thinking beats browsing because it forces the non-negotiables to be decided once, in advance, rather than re-litigated every time a pretty listing shows up. Disqualify hard and cheap first (noise, radon, floodplain membership), stack the masks rather than averaging them, rank the survivors on the weights that are actually yours, then export and verify on the ground.