Methodology Changes and Corrections

Our findings change when the data improves. A better reading of a county's records can move a published number substantially without anything about anyone's property having changed. When that happens we date it here, say what moved, and state plainly how far the correction reached — whether it touched only a figure we publish, the conclusions shown to individual owners, or a document a customer had already received.

Entries are added for material changes to our data processing, our methodology, and our published aggregate figures, newest first.

2026-08-07 — portability counts corrected upward: one county's records had been unreadable to us

What moved. Two published statewide figures. Parcels whose record is consistent with a transferred benefit not yet claimed: 776,941 → 780,376. Of those, the ones our checker shows a visitor: 301,645 → 303,184. Both increases are one county — Indian River — appearing at its true size for the first time. No other county moved, and the homestead figure (617,235) did not move at all.

What caused it. Not a change in anyone's property or in our method. Indian River's prior-year tax roll was stored under a misspelled filename, so the program that reads prior-year sales silently found nothing for that county and counted no portability there. The file was renamed on 2026-08-05 and the figures were regenerated from the county records they are supposed to derive from. The published numbers had been produced while the source was unreadable; they are now produced from the source.

How far it reached. Published aggregate figures only. No owner was shown a different answer, no conclusion changed, and no delivered document was affected — Indian River is not a county we publish an assessment conclusion in. What was wrong was a statewide total on our method page.

One thing we were still checking — now resolved (2026-08-08). The 2026-08-02 entry below records a three-parcel disagreement in Lake County as reconciled. That was premature: the two counts agreed for one rebuild rather than because the cause was fixed, and the same three parcels differed again. We traced it. One of our two calculations was built with one source file unreachable — the state sales file whose current codes override stale ones — so three sales an older roll marked qualified were counted even though the current record disqualifies them. The published figures were right; our internal count was wrong by twelve records, three of which could have reached a visitor's card, no visitor-facing answer was affected, and the build now refuses to run at all when that source is missing rather than proceeding without it. The two counts agree again — this time because the cause is fixed, which is what the 2026-08-02 entry should have waited for.

2026-08-05 — conclusions withdrawn where a comparison group contained a nominal sale

What moved. Thirteen properties across two counties are no longer shown a comparable-sales conclusion. Where we previously stated that an assessment looked high, looked fair, or was close to what nearby sales indicated, those properties now show that we could not run the analysis and why.

What caused it. A sale recorded at a nominal price — a transfer at a round figure far below what the property is worth, rather than an open-market sale — was admitted into those comparison groups by a guard that excluded prices below a threshold but not at it. A second check that would normally catch an implausible sale only runs on larger comparison groups, so in small groups nothing stopped it. Both checks missed the same sale.

How far it reached. Two of the thirteen were being shown a stated finding that their assessment was substantially above what sales indicated; the rest were shown that their assessment held up or sat close to it. A conclusion drawn from a group containing a nominal sale is unfounded in either direction, so all thirteen are withdrawn rather than only the two. No customer document was affected, because none has been delivered. The withheld state is the honest one: with the nominal sale removed, too few sales remain to support an analysis at all.

The underlying defect is being fixed rather than worked around, and these properties will be re-analyzed once that lands. We are recording this the same day it was confirmed.

2026-08-05 — comparable sales: square footage and year built were missing for newly built comparables

What moved. No figure and no conclusion. The comparable-sales exhibit lists each sale the analysis relied on with a price per square foot, and for some of those sales the square footage and year built columns were blank — 841,399 comparable rows, 3.6% of those in service, across all ten counties the analysis covered.

What caused it. A comparable that sold after the previous roll was certified appears on that roll as land: the parcel existed, the building did not. Comparable attributes were read from that earlier roll, so for those sales there was no square footage to show. The price per square foot itself was computed from the current roll and was correct throughout — what was missing was the column that lets a reader check it.

How far it reached. Display completeness of the evidence exhibit only. No estimate changed, no conclusion changed, and no assessment changed. No document delivered to a customer was affected, because no customer document had been delivered. Comparable attributes for those sales now come from the roll that knows the building, and the report will not print a price per square foot it cannot show the square footage for. A small number of comparables carry no year of construction on the roll at all; those columns remain blank, and that is the roll's silence rather than ours.

2026-08-02 — portability count: a three-parcel disagreement between two implementations

What moved. Our published count of parcels with no ported benefit on record stands at 301,645. Before reconciliation, the statewide screen reported 301,645 and the address checker 301,648 — a difference of three parcels in Lake County.

What caused it. The guard that holds the two implementations to the same answer compared only one of the two figures both of them publish, so a disagreement in the other was invisible to it. The guard now covers both.

How far it reached. A published aggregate figure only, by three parcels. No owner was shown a different answer, and no delivered document was affected. We record it because the size is the point rather than an excuse: it is the same shape as a much larger divergence found elsewhere, caught here at three parcels instead of years later.

As of 2026-08-08: the figures in this entry were true when written (2026-08-02) and have since moved — the count now stands at 303,184 after a source correction recorded in the 2026-08-07 entry above. The reconciliation this entry reports also proved premature: the same three parcels diverged again in a later rebuild, were traced to a missing source file rather than to the guard, and are resolved in that entry. This text is left as written; the annotation is the correction.

2026-08-01 — millage: two sources disagreed, and one county is a consolidated city-county

What moved. Two corrections to the tax rates our estimates are computed from. Duval's rate was read as 6.55 mills when the correct figure is 17.56 — Jacksonville is a consolidated city-county and files its levies differently from every other county. Separately, four of our internal sources carried one non-school rate for that county while a fifth carried a lower one; the Department of Revenue's own table confirmed the lower figure, and the rate now has a single source rather than five.

How far it reached. Further than the two entries below. Millage is what turns an exemption or an assessment difference into an estimated dollar figure, so this moved per-parcel estimates shown to owners, not only a published total. No document delivered to a customer was affected, because no customer document had been delivered when these were corrected.

A third millage defect, and how we know it is closed. Our scoring and our statewide screen once applied a single county's rate to all sixty-seven when estimating annual savings. Both now look the rate up per county, and we checked that rather than assuming it: the scored county databases carry a distinct rate for each county, and the published statewide table carries sixty-four distinct school rates and sixty-six distinct non-school rates across sixty-seven counties. An earlier estimate of this defect's effect — roughly −3.1% in aggregate, with five counties diverging by ten percent or more — describes the period before the per-county routing landed, and does not describe what we publish now.

2026-07-31 — statewide homestead screen: nine counties were read with the wrong address convention

What moved. Our published statewide count of parcels with no homestead exemption on record moved from 579,853 to 617,235 — an increase of 37,805.

What caused it. The statewide screen carried its own copy of the comparison that decides whether a parcel's mailing address matches the property address. That copy was labelled as taken verbatim from the address checker's implementation. It was not, and had not been since the checker learned per-county address conventions. Nine counties write addresses in ways the stale copy could not read — numeric ordinals, street-type spellings — so their rates were suppressed toward zero. There is one implementation now, and the screen calls it.

How far it reached. A published aggregate figure only. The per-parcel findings shown to owners come from the address checker, which was already using the per-county conventions; no owner was shown a different answer before and after this correction, and no document already delivered to a customer was affected.

The movement by county: Pasco +16,629, Charlotte +10,462, Santa Rosa +6,854, Flagler +2,928, Levy +637, Gilchrist +144, Dixie +70, Hamilton +41, Liberty +40.