How These Numbers Are Made

This site says Elmira's property taxes fall unfairly. That claim is only worth something if the numbers behind it are right.

Until now you had no way to test that. The explanations lived in code comments that nobody outside the project reads. So this page sets out how a figure is built, from the state record it starts as to the sentence you read. It also says plainly which figures are checked automatically and which are not, because that difference matters more than anything else here. Where the method itself is a judgment call, that is the other page: known limits and judgment calls.

A green outline means a script re-checks that page's figures on every change. An amber one means nothing does. Click any page to jump to its recipe below. This map is drawn from the same file the recipes come from, so it cannot drift from them.

Nearly every page here mixes both. So the question worth asking about any figure on this site is not "where did it come from" — that is in the footer of every page. It is "is anything still watching it?" For most of the site, the answer is no. The rest of this page is about that split.

For these, a change is checked by a script that actively tries to prove the page wrong. Here is what it does.

1. Rebuilds the analysis from the cleaned data. It does not read the published answer first, so it cannot be led toward agreeing.
2. Compares its answer against the published data file. Catches the common failure: someone edits a script and forgets to re-run it, so the site keeps serving last month's answer.
3. Reads the page and finds each figure by the sentence around it. Not by hunting for a bare number, which could match anything. It looks for the claim being made, then checks the number inside it.
4. Fails if the sentence it expects has gone. If prose is reworded so an anchor no longer matches, that is an error, not a pass. A check that goes quiet when it stops understanding a page is worse than no check, because it reads as approval.
5. Fails on any figure it does not recognise. Add a new number to a guarded page without declaring where it came from, and the script refuses it. That is how the guarantee stays complete over time.
6. Confirms that retracted figures have not come back. Old numbers we have withdrawn are searched for across the site, and allowed only where the surrounding text is correcting the record.
7. Exits with an error, which blocks publishing. This is a gate before a change ships, not a suggestion.

This is not a lower standard of sourcing. Every figure on these pages was worked out from the same records and is cited in the page footer. The difference is what happens afterwards.

1. A script or a document produces the figure once. Same data, same scripts, same care as anywhere else.
2. Someone reads it and types it into the page. Checked by hand at the time, against the file it came from.
3. The source goes in the footer. So a reader can go back to the original and work it out again.
4. Nothing checks it again. Not when the data is refreshed. Not when the prose is edited. Not when a related figure elsewhere is corrected. The number stays until a person happens to look at it.

So the honest description of a figure on one of these pages is: correct when written, sourced, and unwatched since. That is a good deal weaker than the guarded pages, and you should read it that way.

Each one says where that page's data starts, which script works out its figures, what gets filtered out on the way, and which choices were judgment rather than arithmetic. They carry rules and never figures — “drop any sale under $10,000” stays true when the data is refreshed, where a count would quietly go stale. Where a rule is enforced by a line of code, the recipe names that line, and a check fails the build if the code stops saying what the recipe claims.

Every source is downloadable twice over: the publisher's own copy, and a gzipped copy of the exact file this site used. Both matter, because the publisher's version moves on and one of them is a search form rather than a fixed file.

Throughout this, the ratio means one thing: a home's assessed value divided by the price it actually sold for. A ratio of 0.60 means the assessor valued the house at 60% of what a buyer paid. The higher the ratio, the more tax you pay per dollar your home is really worth.

1. Start with every recorded sale

   New York publishes the price of every property sale through its SalesWeb service. We keep a copy at data/raw/SaleswebExtract.csv, covering all of Chemung County.

2. Narrow it to sales that can be compared

   Keep only the City of Elmira, only ordinary single-family houses, only arm's-length sales, only 2018 to 2025. Drop anything under $10,000 and any ratio above 5.0. Those are almost always data errors or transfers between relatives.

   What is left: 1,689 sales

3. Work out each home's ratio

   Assessed value divided by sale price, using the assessment that was on the roll at the time of that sale, not a later one.

   The middle ratio of all 1,689: 0.631. The typical Elmira house was assessed at about 63% of what it sold for.

4. Sort by price and compare the two ends

   Group the sales into price bands. Homes that sold under $50,000 had a middle ratio of 1.175. Homes at $150,000 and up had 0.457.

   Cheap homes were assessed above their sale price. Expensive ones at less than half. The raw gap: 2.573×

5. Stop. That gap is too big, and part of it is our own doing

   Sorting by sale price puts sale price in two places at once: along the bottom of the chart, and underneath the ratio. A house that happened to sell cheap lands in the low band and gets a high ratio, for the same reason. The curve bends before any unfairness is involved.

   Sorting by assessed value instead does not fix this. It tilts the error the other way, because assessed value is the top of the ratio. So neither version can be measured against a flat line.

6. Measure how much of the bend is our doing

   Find every house that sold twice within three years — 510 pairs. How far apart those two prices are tells you how much a single sale price strays from what a house is really worth.

   Then simulate a city with exactly that much randomness in its prices and no unfairness at all. Run the same method on it. It still produces a gap of 1.271×. That is the part we caused, now measured.

7. Take it out

   2.573 ÷ 1.271 = 2.02×

   That is what the page publishes, and what it rounds to "about twice as much". The bigger, better-sounding 2.573× is not claimed anywhere, because part of it was never real.

8. Write it out and draw it

   All of it goes into jcurve.json. The assessment page downloads that file and draws both curves, raw and corrected.

9. Check it, including against our past selves

   audit_jcurve_figures.py redoes every step above from the raw sales file each time it runs. It also refuses to let a price band be drawn if too few sales fall into it, which is the mistake that produced the first version of this analysis.

   That first version pooled all eleven towns in the county together. Those towns assess on very different scales, so differences between towns were being read as unfairness within Elmira. It was wrong. We withdrew it, and the check now makes sure those old figures cannot quietly reappear anywhere on the site.

1. There is no dataset behind it

   The 2.0× came from 1,689 rows in a file. This came from an agreement signed in 1994 between the city and a college. There is no table to recompute, no roll to filter, nothing for a script to rebuild.

2. What we have instead

   The City Chamberlain's office confirmed the amount and said it has not changed since the 1994 agreement.

   The payment appears in the 2026 adopted budget under account 412890, "Other General Department Income". Notably it is not booked with the PILOTs at all, which is itself part of the story.

3. What we do not have

   The executed 1994 agreement itself. A records request for it is open, and the PILOT page says so in plain sight rather than implying we have read it.

4. So checking it means something entirely different

   No script can confirm this figure, and adding one would not help. Verifying it means a person reading a budget PDF, or obtaining the agreement.

   If the city renegotiated the payment tomorrow, nothing in our pipeline would notice. The sentence would sit there being wrong until someone read the next budget.

5. What holds it up, then

   Naming the source precisely enough that you can go and check it yourself, and saying out loud which part we have not seen. That is a weaker guarantee than the 2.0× has, and it should be read as weaker.

There are several places where a careful person could do this differently and reach a different answer. The state's published equalization rate for the whole roll says 56%, while the ratio measured from house sales lands anywhere from 46% to 63% depending on which sales and whose dollars. The 3D maps join parcel shapes recorded in 2021 to assessments from 2025. Sorting homes by sale price bends the assessment curve before any unfairness is involved.

Those have a page of their own: Known limits and judgment calls. Each entry says what we chose, why, what the choice costs, and what would change if you chose otherwise. It also lists the figures we have withdrawn, and what this site does not claim. If you are here because we asked you to check our work, start there.

Still to come: a figure-by-figure list covering the whole site, built from the checking scripts themselves.