Methodology

From poll to score

1. 01

   Collect

   Polls and replies are imported from the source X account. Each poll is linked to one or more subjects, people, organisations, or themes.

2. 02

   Classify

   Every poll option is labelled trust, distrust, or neither for each subject on that poll. Replies are scored the same way, per subject.

3. 03

   Score the poll

   A per-poll trust percentage is calculated (see below), then blended with reply sentiment when replies exist.

4. 04

   Aggregate

   The headline score is a weighted average across that subject's polls, larger and newer polls count for more.

Two kinds of poll

How a poll turns into a percentage depends on its shape.

Likert polls

Classic trust questions, options that express trust or distrust toward a subject (for example “Trust” / “Distrust”, or graded scales). Neutral options are ignored in the percentage.

trust% = trust votes ÷ (trust votes + distrust votes) × 100

Comparative polls

Forced-choice questions such as “Who do you trust more?” where each named option is a subject. There is no distrust option for that subject, support is shared across the field. We use Laplace smoothing so small samples do not swing wildly.

score_i = 100 × (v_i + α/k) ÷ (V + α)

v_i = votes for this subject · V = votes in the competitive field · k = field size · α = 4 (prior strength)

Reply sentiment

Replies under a poll are classified for each tagged subject. Reply trust% uses the same trust ÷ (trust + distrust) rule as Likert polls. When replies exist, they are blended into the poll score:

blended = 80% × poll% + 20% × reply%

Poll weight 80% · reply weight 20%. If there are no classifiable replies, the poll percentage stands alone.

Headline score

A subject's published score is a weighted mean of its per-poll blended percentages:

weight = ln(1 + votes) × ½^{age / half-life}

The log term favours larger polls without letting a single huge poll dominate. The half-life is 90 days: a poll that age contributes half the weight of an identical poll today.

How to read a score

Scores run from 0 to 100. Labels are interpretive bands, not statistical confidence intervals.

• Strongly trusted70–100
• Trusted55–69
• Contested45–54
• Distrusted30–44
• Strongly distrusted0–29

What this is not

• Not a national sample. Respondents are whoever engages with polls on one X account.
• Not voting intention. The question is trust, not who you would vote for, though some polls are comparative and look similar.
• Not immune to thin data. Small vote counts and few polls make scores swing. Profiles flag thin samples when the evidence is light.
• Classification is assisted. Option stances and reply sentiment are labelled with language models, then used in the formulae above. Mistakes in labelling flow into the score.