Skip to main content
Fit to CareCare into Action
← Resources

Did they answer? A public record of Prime Minister's Questions

Open codes, session scoreboard, and monthly party balance from January 2016

1. What this is

Every Wednesday the Prime Minister answers questions in the House of Commons. This page records, for every question the Prime Minister has answered in person since January 2016 (the 30 sessions answered by a deputy are left out), whether the question was answered: in full, in part, or not at all, with the cases where the two readers could not agree shown separately. It records the same thing for every question asked, by every MP, under the same rules.

Fit to Care, a public sector consultancy, publishes it as a public service. No party, person or client commissioned or paid for it. It does not say what anyone meant or intended. It counts and it quotes. Every quoted exchange in any article links to its Hansard speech. This page shows the totals; the view that shows every question with its own link follows in a later release.

Coverage: 329 sessions, 9,480 questions and answers, 6 January 2016 to 9 September 2026. 0 sessions are held back and listed in section 6 with the reason.

2. The scoreboard

Colours are codes for answers, never for people. Each Prime Minister is shown as the share of their answers under each code, per session, with the range the data supports.

Prime MinisterFullQ-FPartialQ-PPartial or noneNoneQ-NJustified challengeQ-JConstraintQ-CInterruptedQ-IUnclearQ-UNot judgedNC
Cameron48.3 (44.7 to 52)21.9 (18.8 to 25)5.3 (3.6 to 6.9)15.8 (13 to 18.8)0.7 (0 to 1.6)0 (0 to 0)1.2 (0.3 to 2.3)0 (0 to 0)6.7 (4.8 to 8.6)
May36.3 (34.6 to 38.1)23.5 (21.9 to 25.1)7.8 (6.8 to 8.8)24.7 (23.1 to 26.3)1.1 (0.8 to 1.5)0.3 (0 to 0.6)0.9 (0.5 to 1.3)0 (0 to 0)5.4 (4.6 to 6.3)
Johnson36.2 (34.3 to 38.2)23.1 (21.4 to 24.8)6.5 (5.4 to 7.6)27.9 (25.6 to 30.3)1.1 (0.7 to 1.5)0.3 (0.1 to 0.5)0.2 (0 to 0.4)0.1 (0 to 0.3)4.6 (3.8 to 5.5)
Trusstoo few sessions to compare30.3 (26.7 to 33.3)23.6 (16.7 to 27.6)4.4 (0 to 10)34.8 (33.3 to 36.7)1.1 (0 to 3.4)0 (0 to 0)0 (0 to 0)0 (0 to 0)5.6 (3.4 to 6.7)
Sunak32.3 (29.9 to 34.8)18.3 (16.5 to 20.3)7.4 (6 to 8.8)35.6 (33.2 to 38)0.9 (0.4 to 1.5)0.5 (0.1 to 1)0.3 (0.1 to 0.6)0.1 (0 to 0.2)4.5 (3.6 to 5.4)
Starmer33.3 (31 to 35.5)22.7 (20.2 to 25.3)8.3 (7 to 9.7)30.3 (27.7 to 33)0.9 (0.5 to 1.4)0.3 (0.1 to 0.5)0.2 (0.1 to 0.5)0 (0 to 0)4 (3 to 4.9)
Burnhamtoo few sessions to compare39.7 (33.3 to 46.2)32.1 (30.8 to 33.3)1.9 (0 to 3.8)18.1 (15.4 to 20.8)0 (0 to 0)0 (0 to 0)0 (0 to 0)0 (0 to 0)8.2 (3.8 to 12.5)
All PMs35.8 (34.9 to 36.9)22.4 (21.4 to 23.3)7.2 (6.7 to 7.8)27.8 (26.9 to 28.8)1 (0.8 to 1.3)0.3 (0.2 to 0.4)0.5 (0.3 to 0.6)0 (0 to 0.1)4.9 (4.4 to 5.3)

The columns are the answer codes of the rulebook the record was scored with (section 4). Each row adds to 100 across all nine columns. Two columns are easy to miss: "Partial or none" is where the two readers split between partial and none, and "Not judged" is where the answer could not be assessed, including every answer that one reader read as full and the other as no answer. Neither is ever resolved by hand. Shares are averages per session. The range in brackets is a 95 per cent bootstrap interval over sessions. Prime Ministers with fewer than five sessions are shown but not compared.

By year:

YearSessionsFull answerNo answer
20163243.4 (40.6 to 46.2)18.3 (16 to 20.5)
20172737.9 (34.4 to 41.4)22.6 (20 to 25.4)
20183437.3 (34.7 to 40.4)25 (22.2 to 27.7)
20192732.9 (30 to 36.1)28.7 (25.2 to 32.5)
20203737 (34.3 to 39.8)23.4 (20.5 to 26.5)
20213237.3 (34 to 40.9)29 (25.5 to 32.6)
20223132 (28.9 to 35.2)34.2 (30.5 to 37.8)
20232833.8 (30.7 to 37)34.7 (31.7 to 37.8)
20242732.5 (29.5 to 35.5)33.7 (30.7 to 36.8)
20253334.2 (31.2 to 37.3)29.2 (25.3 to 33.1)
20262132.8 (29 to 36.2)29.8 (26.2 to 33.2)

Selection rule for anything Fit to Care writes about from this board, in order: materiality, novelty, evidence readiness, timing. Heat is not a criterion. Fit to Care may decline to write about a finding; it may not remove it from the board.

3. Party balance

This table covers September 2026 only, the latest month in the record: two sittings, 46 answers judged. With so few answers, a party's figure can swing a lot from one month to the next, and the ranges shown cannot capture that, so read this as a snapshot of one month, not a comparison between parties. Labour was the party of government throughout, so Labour MPs are asking from the Prime Minister's own side and every other party is in opposition.

Published monthly whatever it shows. For each party, the table counts the questions asked by that party's MPs and shows the share that received a full answer.

Party of the MP askingAnswers judgedFull answers (%)
Alliance1100 (100 to 100)too few to compare
Conservative1533.3 (28.6 to 37.5)
Labour1855.6 (37.5 to 70)
Liberal Democrat742.9 (33.3 to 50)too few to compare
Plaid Cymru20 (0 to 0)too few to compare
Scottish National Party333.3 (0 to 50)too few to compare

Last updated 7 October 2026.

4. How a question is judged

Seven codes for the answer and five for any factual claim, adapted from published academic work on question avoidance (Bull and Waddle, among others). The codes and their definitions are open. The engine that applies them at scale and the ranking behind any story are Fit to Care's own work and are not published.

Answer codes

  • Q-F. Full answer. Supplies the requested information or clearly states its absence where that is the answer. Factual accuracy is assessed separately.
  • Q-P. Partial answer. Addresses only part of the information request. Identify the missing part.
  • Q-N. Non-answer. Substitutes another issue, or gives no requested information despite a usable opportunity.
  • Q-J. Justified challenge. Challenges a materially false or ambiguous premise and explains the problem. Does not automatically settle any separable valid question.
  • Q-C. Recognised constraint. Declines to give the requested information and names, as its reason, a constraint from the closed list in the rules below that applies to what was asked.

    The closed list of constraints is part of the scoring method and is not published.

  • Q-I. Opportunity interrupted. Cannot fairly assess completion because the speaker was cut off or the available clip is incomplete.
  • Q-U. Unclear. Ambiguous question, inaudible content or uncertain reference prevents assessment.

Factual codes

  • F-S. Supported. Relevant reliable evidence supports the proposition within its stated scope and time.
  • F-R. Refuted. Relevant reliable evidence contradicts the proposition within the same scope and time.
  • F-C. Conflicting evidence. Material credible evidence supports competing conclusions and the conflict is unresolved.
  • F-U. Insufficient evidence. Available evidence cannot establish support or refutation.
  • F-N. Not a verifiable factual claim. Value judgement, preference, prediction or rhetorical utterance; extract embedded factual claims separately.

Derived codes

  • NC. Could not be judged. The two readers could not agree a code, or the answer could not be assessed.
  • Partial or none. Partial or no answer. The two readers disagreed between a partial answer and no answer; counted as not a full answer and not resolved by hand.
  • No clear ask. Could not be judged at the question gate: no identifiable ask in the question.
  • Points elsewhere. Could not be judged at the question gate: the answer points to a document, report or earlier statement the reader was not given.

Two separate AI passes. Every session is read in two separate AI passes, one by each of two AI readers. The two never see each other's work, and each is told to err on the strict side. They are two different Google Gemini models, so some of their agreement may come from a shared make. Where they disagree on whether an answer was full or partial, the record shows partial. Where they disagree between partial and none, it shows partial or no answer. Where one reads a full answer and the other none, it shows could not be judged, and the count is in section 6. Nobody at Fit to Care settles a disputed answer by hand. The two passes agreed on 82.2 per cent of the 9,480 answers.

Humans check the machine, not the other way round. Agreement between the two passes beyond chance (Cohen's kappa, where 0 is chance and 1 is perfect) is 0.74, which the usual scale calls substantial. Three people are marking the same 50 answers, each choosing before seeing the readers' verdict or anyone else's marks: Anthony Lawton, who built the record; a consultant who is bidding for work jointly with Fit to Care (no work or payment has changed hands yet); and a PR professional who knows Anthony Lawton personally but has no business connection to Fit to Care. Anthony Lawton's first choices matched the readers on 29 of 49 judged answers (59 per cent). On the last 30 answers the page then showed him the readers' reasons: his first choice matched on 16 of 29 judged (55 per cent), and after reading the reasons he agreed on 23 (79 per cent). As of 14 October 2026 the second marker has marked all 50, matching the readers on 26 (52 per cent), and the third has marked 27, matching on 13 (48 per cent). On the 27 answers all three have marked, all three agree on 6, and two of the three pairs agree no better than chance. The readers agree with each other more often than any of the three people agrees with the readers, which is why the published numbers come from the readers under fixed rules and not from anyone's marking. Consistency is not correctness: this shows the readers apply the rules the same way, not that the rules are right.

One answer, two readings. 8 February 2023. Feryal Clark (Labour) asked when further aid for the Turkey and Syria earthquake would be announced, and what discussions the Prime Minister was having with international counterparts. The Prime Minister described his call with President Erdogan, the search and rescue teams on the ground, and the Foreign Secretary's contact with the United Nations.

Reader A: full answer. "Fully addresses the questions by detailing his own and the Foreign Secretary's discussions with international counterparts and explaining the current and ongoing nature of the UK's aid commitments."

Reader B: partial answer. "Addresses the second part of the question by outlining discussions. However, he does not state when an announcement on further aid commitments can be expected."

Published code: partial, because when the readers split between full and partial the record shows partial. All three human markers chose full answer.

Two rulebooks: the ten-year record stays fixed under the rules it was scored with, v0.5. New sessions are scored under the current rules, v0.7. Both are listed in section 7 with what changed and why, so a reader can see the system changing and can judge it.

5. Right of reply and corrections

Anyone named on this page, or their office, may reply. Before Fit to Care publishes any article drawn from this record, the people named in it are sent the relevant findings at least 48 hours in advance and their reply is published with the article.

Articles drawn from this record, with their sources, are published on The Record.

To ask for a correction, write to anthony.lawton@ffmi.co.uk. A correction that shows the record is wrong (for example, a question answered later in the same session) is made within 48 hours. The original wording stays visible, struck through, beside the correction. Every correction is listed below with its date, whatever it shows.

Corrections log

  • 14 October 2026: one sentence in the article Whoever governs (fittocare.co.uk/the-record/pmqs-personal-remarks) corrected. It said Anthony Lawton was the only marker shown the AI's reasons; all three markers were shown them, after making their choice. No figure changed.

6. Sessions held back, and the widest splits

No session is currently held back.

212 of 9,480 answers in the record (2.2 per cent) were read as a full answer by one reader and as no answer by the other. They are shown as could not be judged. The rule: Where one reader records a full answer and the other no answer, the answer is not classified and the count is published; from rubric v0.7 new sessions go to a third reader under the three-step rule.

Until 2 October 2026 a session with such a split was held back from the board pending a review that never changed a code. From rulebook v0.7 (section 7) a new session with such a split goes to a third reader, and no session is held back for this reason. A session with incomplete Hansard text is still held back and listed here with the reason.

7. The rules this page keeps

RuleText
1Scope: elected national politicians in Hansard text, and public sector documents. Never a named public sector officer, NHS or council employee, or private individual.
2Facts, never intent.
3Same rubric for everyone, every session scored in full.
4The whole scoreboard is public. Nothing is withheld.
5Two separate passes before anything leaves.
6Every claim carries a link.
7Right of reply and a public corrections log, on this record's own page.
8Party balance published monthly.
9Numbers and quotes, no adjectives about people.
10No claim of impact without a measurement.
11Licence and attribution on every output.
12Open codes, protected engine.
13The kill switch.

(The thirteen credibility rules: the rule column only, from Credibility Rules v0.2, with the project name replaced by "this record" and rule 5 reworded on 14 October 2026 to say separate passes, because both passes use models from one maker. The "in practice" and "example" columns are not published: they carry typed figures and quoted words that the page's own gates forbid.)

v0.5. What changed: Appendix 3 refusals and commitments, Q-C recognised constraint, tactics on Q-P and Q-N, and asked-again flag; readers A and B prompts v0_5. Why: Option A pilot and ten-year corpus scoring; fixed for the published record.

v0.7. What changed: Full against no answer splits use a third reader and the three-step rule instead of session holds; step 1 NC reasons no clear ask and answer points elsewhere. Why: October 2026 ruling: no human adjudication on full against no answer splits.

The three-step rule of rulebook v0.7, applied by a third reader shown the question and answer and neither earlier reading, when one reader records a full answer and the other no answer:

  1. Question gate. List every ask in the question. No identifiable ask: could not be judged, reason "no clear ask". An answer that only points to a document, report or earlier statement the reader has not been given: could not be judged, reason "answer points elsewhere".
  2. Ask by ask. For each ask: answered, addressed without the information, or not addressed (including answering a different question, or speaking about the questioner rather than the question).
  3. Result. All asks answered: full answer. Some answered or addressed: partial answer. None addressed: no answer.

No step records why the person answered as they did. The ten-year record stays as scored under rulebook v0.5; the rule applies to sessions scored from October 2026.

8. Licence and sources

Contains Parliamentary information licensed under the Open Parliament Licence v3.0. Hansard text via TheyWorkForYou (mySociety). Every quoted exchange on this page and in any article links to its Hansard speech id and date.

This record measures whether a question was answered. It does not measure whether what was said was true. Fact checking, where Fit to Care does it, is published inside an article with its sources, never as a list on this page.