Data releases

Changed in this release

• Refusal-reason corpus repaired. A regeneration on 15 July 2026 had cut the coded refusal corpus to roughly a third of its size. It has been rebuilt to 13,823 coded reason paragraphs across 4,707 refused schemes, of which 12,571 across 4,532 schemes are refusal reasons; the remainder are appeal decision letters, section headings and fragments picked up by the harvester, and are excluded from the statistics. The repair was a merge rather than a rollback, so applications harvested since June are retained.
• Coverage note added for Haringey. Its refusal-reason extraction sits outside the repaired set and is not directly comparable with the other 32 authorities.

Known limits in this release

• The most recent months are incomplete. Decisions reach the London Datahub with a lag, so monthly counts fall away after March 2026. Recent quarters will fill in rather than stay as published.
• The 33 borough dashboards now sit on the same cohort as the rest of the site. They were built from each authority’s own coded workbook, which disagreed with the pan-London table on 29 of 33 boroughs by at least ten decided cases or two percentage points. Lambeth was the worst: 440 decided at 56.4% approval on the workbook against 286 at 41.6% once non-development applications are removed. The dashboards now sum to exactly the headline figure. Several borough approval rates moved as a result, Lambeth most of all.
• The headline cohort now excludes non-development applications. The dataset carries 932 rows that are condition discharges, non-material amendments, lawful-development certificates and details submitted pursuant to condition. They are 83.5% approved, because discharges nearly always are, and they are not proposals for homes. They have been excluded from the borough differential since June 2026, but the headline figures still counted them. Applications for one to nine homes now read 14,600 rather than 15,534, decided 11,694 rather than 12,493, and the overall refusal rate 45.4% rather than 43.5%. The raw counts are unchanged and remain the reconciling figures; what changed is which of them the site puts forward as the headline.
• Refusal-reason classification has been rebuilt, and the design share moved. The June 2026 classification did not survive the July regeneration. All 13,823 reasons were re-coded in August 2026 and the new coding measures 91.8% accurate against a stratified sample of 143 decision-notice reasons adjudicated one at a time, against 46.3% for the keyword coding it replaces. Design as a first-listed reason now reads 24.2%, against the 35.5% previously published. Space standards roughly double, to 22.1%: the old coding was routing substandard-accommodation reasons to design because their text mentions design and layout.
• Rebuild reconciliation closed. The rebuild and the raw dataset agree at 15,534 rows and 12,493 decided. An earlier version of this note described a discrepancy between them; it was resolved at a re-stamp and the sentence outlived the problem it described. The headline figures on this site are now the development cohort, which is smaller — see the cohort entry above. Rather than restate figures piecemeal, the site, the white paper and the claims register are re-stamped together at a refresh.

• Refusal reasons were reclassified, twice. The original classifier defaulted to design, which inflated it as a refusal category. A June 2026 pass moved design from roughly 55% of primary reasons to 35.5%. A second pass in August 2026, after the corpus was repaired, moved it again to 24.2% and roughly doubled space standards. The same pass found that 11.7% of the corpus was not refusal reasons at all, chiefly appeal decision letters harvested in place of council decision notices; those rows are now excluded. Both passes found the same fault in the same direction; the second measured itself against an adjudicated sample, which the first did not.
• Non-development applications were removed from the cohort. Condition discharges, non-material amendments and lawful development certificates had leaked into the small-site cohort. Excluding them changed several borough approval rates.

The dataset·Download the data·Method·Report a problem