Why we count instead of scoring

The rule, as we wrote it

The phrase patterns behind the answer-patterns page live in a reviewed registry in our codebase. Its stated design rule, quoted verbatim:

“These are COUNTED PHRASES, NOT JUDGEMENTS. Every pattern is a deterministic, case-insensitive, whole-word regex over the text of ministerial answers. No model reads an answer, no verdict is attached to a named person, and a reader can re-run every count from the published pattern and the official record.”

The page itself repeats the rule to its readers: whether an answer that uses one of these phrases is an adequate answer is for the reader, and the Member who asked, to judge. The same discipline governs the unanswered board and the answer-time league: days elapsed against Parliament’s own expected dates, medians and percentiles over the record, nothing a model inferred.

A count is a fact. A score is an opinion wearing a number.

“This department used the phrase ‘not held centrally’ in 214 of 3,100 answers to questions tabled in the last year” is a fact. The pattern is published, the record is public, and anyone can re-run the count and get the same number. It carries its own proof.

“This minister scores 7.2 out of 10 for evasion” is a judgement. It cannot be re-run: the number depends on which model produced it, on which day, with which prompt, over which slice of text. Two runs can disagree with each other and neither can show its working. Attaching such a number to a named person does not make the judgement more rigorous. It makes it look like a measurement while removing the reasoning a reader would need to challenge it.

So we publish the counts and the verbatim, linked examples, and we leave the judging to the professional reading them. That is not a limitation of these pages. It is the design.

The liability a repeated score creates

Public affairs professionals repeat what their tools tell them: in client briefings, in board papers, in notes that get forwarded. Repeat a machine’s stance or evasion score about a named parliamentarian and it becomes your claim. You cannot reproduce it, the vendor cannot fully explain it, and the person it names can dispute it, to you, to your client, or in public. The professional inherits the liability for a judgement no human ever made.

A dated phrase count with the pattern published inherits nothing. If challenged, you point at the record and the regex, and the challenge resolves itself. That is the standard a claim should meet before it goes in front of a client with your name on it.

Inferred political opinions are special-category data

UK data-protection law (UK GDPR, Article 9) treats personal data revealing political opinions as special-category data, with a higher bar for processing it. The ICO’s position is that inferred data is still personal data: an opinion about a person’s politics generated by a machine and stored against their name is special-category data just as a stated opinion would be. A stance score about a named parliamentarian is exactly that kind of inference, manufactured at scale.

Counting published phrases in official public answers makes no inference about anyone’s opinions. It states what the public record contains, with the date it was measured. That is the side of the line we build on, and it is a policy, not an accident: we do not manufacture inferred opinions about named people, for these pages or anywhere else in the product.

How to re-run every number we publish

• Every phrase pattern is published in full on the answer-patterns page as the exact regular expression we run, currently 7 patterns. Matching is case-insensitive and whole-word, and text inside double quotation marks is removed first, so an answer restating the question is never counted as the department using the phrase.
• Every figure carries two dates: the day it was measured and the day our copy of the record of answers is complete to. Claims of “unanswered” are made as of the record-complete date, never as of today.
• “Overdue” always means past Parliament’s own recorded expected answer date, computed in sitting days by Parliament itself. We never invent a deadline.
• Every example is quoted verbatim, never paraphrased, and links to the question on questions-statements.parliament.uk, where the full exchange can be read in context.
• When a live measurement is unavailable, the page says so. It never shows a cached or estimated figure in its place.

This policy applies to every accountability surface we publish, and to the answers Emily gives her subscribers: she cites the dated record and lets the professional judge.