Modern message intelligence.
Torus uses emergent data and large-language-model processing to deliver an entirely new messaging capability and campaign management toolkit — built to replace guesswork with a defensible, district-grounded map of belief.
Six stages, one loop — cluster, interview, frame, assign, deploy, listen.
A topline and a quarterly memo aren't the cycle a campaign actually runs on. Each stage of the Torus feeds the next: what voters say reshapes what we produce, and what lands reshapes what we say next.
Cluster
Archetype formation. Voters are cut from deep, auditable records on stated dimensions — so anyone can inspect exactly how a voter was grouped.
Interview
The district in its own words. AI-driven, open-ended interviews invite voters to raise their own issues and priorities, not march through assumed themes.
Frame
LLM decomposition and analysis. Every statement is broken into its components — the problem, who is responsible, what they want done, and how strongly they feel it.
Assign
Starting position on the message strategy matrix: how often a voter turns out, by how closely they align with the candidate.
Deploy
Media in their language, on their issues — the stacked priorities of an archetype decide what gets said first, and to whom.
Listen
Measure, correct, reassign. Response moves voters on the matrix, and the next round of messaging starts from where they actually are.
Three deliverables, one continuous brief.
Torus doesn’t discard polling — it supplies what a topline can’t: the district’s own structure, in the district’s own words, in a workspace your strategy team can act on today.
The Archetype Map
Who is in the district and how they were grouped — archetypes cut on stated dimensions, each with the stacked priorities that rank what to say first.
Bridge Narratives
The convictions that move disparate archetypes at once — messages built on insight, capable of forming broad coalition, in the language voters actually use.
The Message Strategy Matrix
Every voter on one square — turnout by candidate alignment — so the play is explicit: persuade, wake up, turn out, hold, or leave alone.
Let the data organize itself.
Emergent data is organized by its own properties rather than categories imposed by the viewer. That distinction matters because bias enters research at the moment of design: the buckets a pollster picks and the themes they write into a questionnaire decide most of what can be found. We let the structure and the themes come out of the data instead, and the findings are both authentic and accurately applied.
Archetypes, the map of who
Traditional sampling stratifies on a handful of buckets chosen in advance. Archetype clustering forms on scores of verifiable attributes and finds the groupings that already exist — each one internally coherent, so a few hundred interviews represent the district far more faithfully than quotas ever could.
A sampling frame drawn from real structureStacked identities, the map of belief
Open-ended interviews, not a checklist of assumed themes. Nobody hands the district a set of issues in advance; the themes that matter come out of what voters chose to talk about.
Discovered, not assumed · 750–2,500 interviews per district“They cut the bus that got my mother to her appointments. The county board did that, and nobody asked us. Put the morning run back and I’ll listen to whoever did it.”
Interview excerpt · senior homeowners
Archetypes are statistically derived, not presumed.
Every voter in the district is matched to a deep record of who they verifiably are — household, housing, work, and place — and BIRCH-Ward clustering groups them from those facts alone. Nobody decides in advance how many groups a district has: the clustering finds the number that actually fits, and every voter’s grouping can be traced back to the facts that produced it.
Every voter occupies exactly one square.
How often they vote, by how closely they align with the candidate. Position sets the play — and tells the campaign what mix of persuasion and mobilization is worth the spend. No voter is counted twice; none is moved along a generic ladder of touches.
votes
sometimes
generals
every
election
Built on transcripts, not toplines.
The core limit of polling is methodological and intrinsic: it imposes demographic buckets and assumed framing before a single response is collected, then reads the results as if they were discovered rather than presupposed. The bias lives in the architecture, not the execution.
Archetypes cut before a question is asked
Groupings come from a deep, non-public dataset on stated dimensions — auditable, and never a set of demos the campaign hands us in advance.
Open-ended interviews, then decomposition
750–2,500 conversations per district. Each statement is broken into problem, blame, remedy, and intensity — then stacked across an archetype into ranked priorities.
Every voter assigned exactly one square
Turnout by candidate alignment. No voter is counted twice, and no one is moved along a generic ladder of touches.
Transparent provenance
Every recommendation links to the voter quotes and archetype evidence it rests on. Your team can audit the chain back to the human voice.
A campaign that listens first
wins more than once.
Tell us about your race. We'll send a sample capability brief and walk you through the method.