← Pauline Tully

Rhetoric and votes

Pauline Tully

Why the denominator is small

The measure only exists where this member actually spoke in the debate their vote belongs to, and where that vote can be tied to a specific question. Both cuts are steep.

The third row is the sharpest cut. A committee-stage sitting can hold fifteen divisions while its speeches discuss forty-two numbered amendments, and the record does not say which amendment each division was. Those votes are counted here and deliberately left unscored rather than guessed.

Alignment against the whip

The same scorable votes, split by whether the member went with their party. The top-right cell is the one worth reading: a stated position reversed while following the party line.

Not scored: 21 where the vote could not be tied to a question, 0 where no clear position was stated in the debate.

Where the words and the vote diverge

Every case the pipeline flagged, with the contribution it read and the vote as recorded. Judge them yourself — that is the point of showing them rather than only the rate.

Tone

Valence runs −1 (negative) to +1 (positive). This is a separate measure from alignment above and says nothing about how anyone voted — "the Minister has been asleep at the wheel" is strongly negative and carries no position on the question.

Which policies move them

Mean tone by topic, most negative first. Topics come from an LDA model fitted over debate contributions; the terms are the model's, the labels are hand-assigned.

Departures from their own baseline

Contributions whose tone sits more than two standard deviations from this member's own average. Baselines are per member on purpose — an absolute threshold would just rank people by temperament.

The whip, across the whole corpus

Alignment by the party a member sat for on the day, with 95% Wilson intervals. This is the test the measure could have failed.

Members of government parties align their stated position with their vote 9.4 points less often than Independents — bootstrap 95% CI [+2.7, +16.0], which excludes zero. Opposition parties sit 9.5 points above government too, CI [+3.1, +15.9]. Independents and opposition members are indistinguishable from each other (−0.0 points, CI [−6.3, +6.2]). So the gap does not come from having no party; it comes from being in government, which is what a whip effect should look like. Read this as the measure behaving sensibly, not as a finding about any individual.

How the scorer scores

Two automated methods were built and both failed. What survives is the rule-based layer, which is why the coverage above is narrow.

Zero-shot entailment was asked to judge contributions whose true position is known — the member moved the amendment, so the answer is support by construction — and got it right 11.5% of the time, returning "unclear" or "oppose" for nearly everything. A classifier trained on the rule-labelled cases over the corpus embeddings reached 71.1% against a 65.3% majority-class baseline: better than chance, not better enough. Both are excluded from every figure on this page. Only the rule-based cascade, which is exact where it fires and silent where it does not, is used.

Method

Position is read from the member's own contributions in the debate section the division belongs to, by a three-step cascade: an explicit move of the question ("I move amendment No. 5", true by construction); declarative stance phrases ordered so negations are tested first, because "I cannot support" and "I do not oppose" both defeat a naive keyword match; and, where neither fires, a zero-shot entailment model judging the contribution against the question put. Tone uses a separate model and is reported separately. Agreement against hand-labelled cases, with bootstrap intervals, is on the evidence page; no score here should be read without it.