Methods

This is AI-generated grading, not a human review. We're currently in an alpha-testing phase, actively working on the quality and consistency of the process described below. If a particular verdict looks wrong, or a question and answer seem mismatched, please tell us — there's a feedback form on every question's own page, or see Contact. We read every submission.

How a verdict is reached

Each question is broken down into the distinct, separately-answerable items it actually asks for. Each reply is separately broken down into its individual statements. Every item is then checked against the reply's statements and marked:

• Addressed — the reply substantively answers this item.
• Partially addressed — the reply touches on it, but is incomplete or vague.
• Deferred — the reply acknowledges the item but defers the substantive answer to a future date.
• Data not held — the reply gives everything the department actually holds on the topic, and is honest that a more specific breakdown isn't separately tracked. Treated as a complete, honest answer, not an evasion.
• Not addressed — the reply doesn't engage with the item at all.

Each item is also rated on how specifically the reply engages with it — a reply can be technically "addressed" while still being vague, generic, or evasive in how it gets there. That full item-by-item breakdown, with a short rationale for each, is shown on every question's own page under "How we got there".

The score and the colour

Those item-level judgments are rolled up into a single score out of 100, which maps onto the Red/Amber/Green colour you see throughout the site: below 25 is Red, 75 and above is Green, and everything in between is Amber — deliberately the largest band, since most real answers are neither a clean pass nor a clean fail.

What gets judged, and what doesn't

Not every line in Hansard is a question with an answer to score. A ministerial Statement's own opening announcement, and an Urgent Question's own opening statement in reply, are shown as context on the relevant page but aren't themselves scored — there's nothing in a procedural "will the Minister make a statement" request to meaningfully judge. Scoring starts with the first genuine question or supplementary that follows.

Who does the judging

The judgment itself is made by a large language model (currently Anthropic's Claude), prompted with the question, the reply, and the criteria above. It is a careful, consistent reading of the text — not a formal, legal, or fact-checked assessment, and not a view on whether the underlying policy itself is good or bad.

Rhetorical style

Alongside the coverage judgment, some question items and reply statements are also tagged with a rhetorical style — a separate, purely descriptive label for how something is said, not whether it was answered. These tags never affect the score or the RAG colour; they're there for a reader who wants to notice patterns, such as whether a particular kind of question tends to get a particular kind of reply.

• Point-scoring — a jibe or attack on the other side's record, character, or competence, with no real policy content behind it. A pointed or critical question isn't automatically point-scoring: "will the Government reverse this damaging policy" is a real ask with a real possible answer. Only items and statements with no answerable content beyond an attack or a reaction to one are tagged.
• Vibes — emotionally resonant or aspirational language — hope, unity, pride, tributes, "a brighter future" — with nothing checkable behind it: no action, figure, mechanism, or date, just sentiment. This is distinct from a genuine question or commitment that's simply vague or open-ended (that's an honest "hedged" or "open" ask, and is rated as part of Specificity, which does affect the score); vibes has no ask or commitment behind the sentiment at all.

Where a question's items or a reply's statements carry one of these tags, a percentage badge is shown alongside the relevant text — for example "100% vibes" — showing what share of that question's items, or that reply's statements, carry the tag. The badge is only shown when at least one item or statement is tagged.

How search works

Search matches your words two ways at once. The first is an exact match against each question's own text — the same literal search anyone would expect. The second is a semantic match: your search is compared, by meaning rather than wording, against every question using an AI embedding model, which can surface a relevant question even when it's phrased quite differently from your search (for example, a search for "hospital closures" finding a question about an A&E department shutting). Exact matches are shown first, in date order; a handful of semantic-only matches — ones that don't contain your actual words — are appended after them and marked SIMILAR, so it's clear why they're there.

Limitations

This approach is continually being refined as real cases turn up edge cases it doesn't yet handle well — and some are still being worked through. If a particular verdict looks wrong to you, use the feedback form on that question's own page, or see Contact.