The words behind a reading.

Population twin

A simulated member of a place’s population. Each twin is modeled on the real make-up of that place — age, household, work, housing, and for businesses their industry and size — and answers a question the way someone in that position plausibly would.

Also: twin, synthetic respondent.

Synthetic respondent

The general term for a simulated answerer in research. In Helm Express every synthetic respondent is a population twin of a specific place.

Place

The geography a population is built for: a city, county, metro area or state. Every reading is about one place, or about several set side by side.

Area (neighbourhood)

A smaller part of a place, used to draw the map and to cut a reading by where twins live. Positions on the map are for display; nothing scored depends on a dot’s exact spot.

Audience

Who is being asked: residents (the people who live in a place) or businesses (a simulated set of the place’s firms, each answered by a decision-maker). A real business name is never given an opinion.

Likely electorate

For a ballot-style question, residents weighted by how likely each is to vote. For a service, product or price question you can ask all residents instead.

Reading

One run of one question against one place’s twins: the direction it leans, the breakdowns by group, the interval, and a read-quality grade. A reading is a simulation, not a survey.

Framing

The shape of the question. Yes/no asks for a choice between two; options asks for a pick among up to 12; belief asks what the twin thinks is true rather than what it would do. A question with fewer than two options is read as yes/no.

Contrast

A difference between two things in the same reading — option B against option A, renters against homeowners, one neighbourhood against another. Contrasts are what a simulation is best suited to, and they are how Helm Express describes results: “B ahead of A among renters”, rather than a share of the vote.

Directional

Read for which way a result leans and how groups differ, not for its exact level. Every reading in Helm Express is directional in this sense.

Split (A/B)

Two versions of a message or a piece of news put to the same twins, so the gap between the two readings is not a gap between two different samples.

Segment

A group within the population that a reading is narrowed or cut to — for example renters, households with children, or businesses in food service. The smaller the segment, the wider its interval.

Breakdown (cut)

The same reading shown group by group: by age, household, work, housing, neighbourhood, or for businesses by industry and size. Cutting a reading does not ask the question again.

Cell

A group of twins who are alike on the things that matter for the question. One model call reads a whole cell, which is why the interval is priced on cells rather than on individual twins.

Weighting

Adjusting how much each twin counts so the population matches the make-up of the place. Twins carry unequal weights, so a panel of many twins is not the same as that many independent answers.

Post-stratification

The weighting step after the answers are in: each group’s answer is scaled back to that group’s real share of the place, so an over- or under-read group does not tilt the total.

Weighting to a customer file

Reweighting a place’s twins toward the make-up of a list you upload, such as your customers. The reading shows what that weighting costs in precision.

Effective sample

How many fully independent answers a reading is worth once unequal weights and shared cells are taken into account. It is usually much smaller than the number of twins, and it is what the interval is built on.

Also: effective sample size, n_eff.

Design effect

How much weighting and grouping shrink the effective sample compared with the same number of independent answers. A design effect of two means the reading carries about half the information its size suggests.

Interval

The range a reading’s estimate could plausibly fall in, given how much independent information the reading has. When two options or two groups sit inside each other’s intervals, the reading cannot separate them — and says so.

Read-quality grade

A plain label on every result for how much weight it can bear. Tight read: precise enough to compare options. Comparable: large differences are meaningful, small ones may be noise. Directional only: shows which way it leans, not by how much. Incomplete: part of the result didn’t come back — run it again.

As-of date

The date a reading is set at. Asking as of an earlier date lets you set a question next to the same question at another time.

Event (what-if)

A piece of news the twins are told has just happened, before they hear the question. Reading the same question with and without it shows what the news shifts.

Backtest

Asking a past question whose real outcome is already known, and scoring the reading against it. Backtests are not shown any news, so the answer is not handed to them. The model may still know a past outcome from its training, so read a backtest as a check, not proof.

Twin interview

A conversation with up to five twins from a reading about a question, each in its own words, shown with who it is modeled on. A twin interview explains a reading; it is not a result on its own.

Interview with real people

An interview or short study with members of the public or credential-checked experts, booked with Helm. Its answers are labelled as real people and kept separate from twin readings.

How these pieces fit together is in the notes: what a twin reads before it answers, why the interval is priced on effective sample, and why results are reported as contrasts.