How accurate is this?

Every figure on this site is read by software from a decision the tribunal published on GOV.UK, and every page links to the original. Software misreads things. This page says how often, how we know, and what is done about it. The table is counted from the database on 31 Oct 2026.

How a decision is read

Each decision is read twice. First by fixed patterns: the header block (property, applicant, respondent, type of application), the operative "Decision" section, and fixed vocabularies for the kinds of charge, section 20C and section 20. Then by a language model that must quote, word for word, the passage holding each value; a value whose quote is not in the decision is dropped, and a company name is kept only where the decision names it. Where the two readings disagree on the outcome, the section 20C result or a total, the decision is listed with "figures not yet checked" and left out of every count, unless the model's quote sits in the operative decision section, where its reading stands. A company the two readings disagree on is stored blank and the decision counts as "not stated". People are never named.

Hand checks

Twice, 40 service charge decisions were read by readers given only the text, with no access to the software or its answers, and their reading compared field by field with what the software stored. The first set (29 October 2026) was drawn from the sample the software was tuned on; the second (30 October 2026) from the full run, none of them in the first set. Every disagreement was then read in the decision and settled.

Because the readers' own reading of the outcome was under 85%, every share that rests on the outcome is shown as "about X%" until a later check clears it. The section 20C share is shown as a plain percentage.

The second check, field by field

The readers' settled reading of the 40 decisions in the second set against what the database holds today (40 of 40 are in it with figures). A field agrees when the two give the same value; for the lists of years and kinds of charge, one item's difference either way still counts as agreeing; a company name agrees when it is the same company under another spelling.

The weak fields

Under 85% in the second check:

• The freeholder or landlord: 30 of 40. Every miss was a blank where the decision does name the landlord, half of them a council or housing association. Where the two readings disagree nothing is stored, so a company page shows no decision on a guess; a decision can still be missing from the page it belongs on.
• The managing agent: 32 of 40. Mostly a blank where the decision names the agent; twice a company the decision names in another role, once a spelling. Where the two readings disagree nothing is stored, so a company page shows no decision on a guess; a decision can still be missing from the page it belongs on.
• One flat, a block or an estate: 31 of 40. A block read as one flat or an estate, or the other way round: decisions rarely say it in so many words. Shown on the decision page, never counted.
• Number of flats: 30 of 40. A blank, or the number of flats in the estate read where the decision counts the block, or the other way round. Shown on the decision page, never counted.
• The service charge years in dispute: 18 of 40. The usual miss is a run of years read as its first and last year only, or only the year of the demand in dispute. The years are shown on the decision page, never counted.
• The kinds of charge in dispute: 26 of 40. The usual miss is a kind of charge mentioned in passing and listed as in dispute, or a named kind filed under "other". The "in dispute" column is context for the "disallowed or reduced" column, which is checked separately.
• The total found payable, where stated as one figure: 33 of 40. A blank where the decision states a total, a total where the reader found none stated as one figure, once a figure for one item read as the total. Only decisions stating both totals as single figures are in the typical reduction.
• Costs and fees ordered: 33 of 40. The check matches on the first amount only, so the same order in other words counts as a miss (four of the seven). Shown on the decision page in the decision's own words, never counted.

What is held back or left out

• Held back today: 1,387 of the 7,007 decisions shown are marked "figures not yet checked" (the two readings disagreed on the outcome, the section 20C result or a total) and are left out of every count except the number of decisions read.
• Companies: 371 decisions count as "not stated" because the two readings gave different names.
• Dispensation applications (case code LDC): their own line, never in a share-reduced figure, because they are landlord-driven (granted in 2,684 of 2,839 counted).
• Later costs rulings on the same case: linked to the main decision, not counted as a decision of their own.
• Image-only PDFs and anonymised cases: left out. Re-published cases: the main decision counts once.
• Shares from fewer than 10 decisions are given as counts, not percentages.

Every figure carries its n. If you find an error, write to hello@rentrisecheck.uk: it will be checked against the original and corrected in the open. The figures themselves are on the figures page and in press/findings.txt.