This page records two things: figures that were published and turned out to be wrong, and analyses that were completed, checked and then rejected before they were published. Both lists are below. Neither is empty.
Every one of the following was a finished result with a headline attached. Each was killed by its own data. They are published because a corrections page that is empty on day one tells you nothing about the standard being applied.
Rejected 2026-08-18
A first pass measured heat-pump share against free-text heating permits and found Lexington at 100% and Cambridge at 87%. Both figures were artifacts of record-keeping, not heating markets: those towns file fossil-fuel heating work under a fixed category label rather than a description, so a text search finds every heat pump and no boiler. Lexington's 84 gas permits carry exactly one distinct description string between them. Rather than drop the story we added an eligibility test and published the exclusions as part of the finding -- see the heat-pump article.
Rejected 2026-08-18
The 7x spread was a permit-mix artifact. In two of the towns, around 61% of the valuation-bearing 'kitchen' rows were electrical or plumbing sub-permits, versus 9% in the town at the other end. Comparing building permits to building permits, the spread collapses to 1.9x -- against 1.6x for roofing, which was supposed to be the control. Kitchens vary about as much as roofs do, which is the opposite of the headline.
Rejected 2026-08-18
Dropped on privacy, permanently. 88.5% of contractor-bearing permits name an individual rather than a firm, and the largest single filer in the solar data is a private person. No level of aggregation makes that safe to publish, so the contractor column is never article subject matter here.
Rejected 2026-08-18
The apparent 20% downward revision was measured across a window in which our own corpus went from 44,296 rows to 23,577 to 34,382 to 37,426, while sources were being lost and rebuilt. Every 'revision' was ours. Worth revisiting after roughly 90 days of uninterrupted snapshots on a stable source list.
Three claims went out wrong on the first day of this tier and were corrected the same day. Each is listed with what it said, what the data actually shows, and why the error was possible. The articles kept their URLs.
Corrected 2026-08-18 · /research/heat-pump-share
The page said the excluded sources carried more heating permits than the whole eligible sample. They carried fewer - 462 against 499 on the day it published - and both numbers were printed on the same page and in its own downloadable evidence file. The cause: a comparative hardcoded in the prose while the two numbers either side of it were computed, so nothing could ever check it. The relation is computed now, and the sentence prints the eligible total beside it.
Corrected 2026-08-18 · /research/adu-permits
The page inferred from the word 'garage' that ADU work is usually a conversion of existing structure. Matching conversion and new-construction language separately shows the opposite: new construction leads 53 to 25, and within the garage-mentioning permits it is too close to call. A third of the permits use neither vocabulary and are now reported as unclassified instead of being read as conversions. The section has been rewritten around what was actually measured.
Corrected 2026-08-18 · /research/friday-permit-gap
That was false: 21 of the 28 sources behind the article run the same permit vendor, so the sentence asserting vendor independence was the claim most worth checking and the least checked. The article now splits the data on the vendor and publishes both halves. The finding survives - the seven sources on entirely separate systems, including Boston, Worcester and Cambridge, all sit below their own weekday average - and the shared-vendor towns dip harder, which is stated rather than hidden.
All three were found by an adversarial audit of this tier, not by a reader. That is what the audit is for, and publishing the result is what this page is for.
Every article links the JSON file containing its own figures, so a disagreement can be made specific. Corrections go to [email protected].