The evidence

What we measured, and what we still don't know.

The geometry is settled: we can show that the walk is short. Whether a postcard can talk somebody into taking it is a completely different question, nobody has a credible number for it, and we are not going to pretend otherwise.

Settled · the geometry

Ten precincts, one county, every address.

We took the public North Carolina voter file and voter history for ten precincts in Halifax County, geocoded the households, and sorted every resident into two groups.

Who could be a captain

Residents who voted the general election in 2020, 2022 and 2024 -- people for whom voting is simply a habit. There are 6,264 of them, about 38% of everyone registered in those precincts.

An earlier, stricter definition (every federal general since 2018, plus municipals) produced only 254 people across these ten precincts, and it failed in a specific way worth stating: the precincts with the greatest need had almost no one who qualified. One precinct had 777 neighbors to reach and zero eligible captains. A definition that cannot staff the places that need staffing is the wrong definition.

Who needs reaching

Registered residents who vote inconsistently (7,023) or who have no general election on record at all (2,803) -- 9,826 people.

Then the only question that matters for the model: how far apart are the two groups? 95% of those 9,826 live within ten doors of someone in the first group. If every one of those habitual voters walked only her ten nearest households, 9,306 of the 9,826 would be reached by someone who already lives on their street.

An average captain's ten doors hold about 7.6 neighbors who vote inconsistently and 2.6 who have never voted in a general, alongside roughly eight neighbors who already vote and could help her knock. That is the size of the job: not a turf, not a shift, a short walk with about ten conversations at the end of it.

Unsettled · the funnel

Nobody knows how many people say yes.

Every projection of this model runs through one number we have never measured: the share of mailed neighbors who opt in and actually walk their doors. Our working assumption is 1-3%. That is a guess. Measuring it is the whole point of the pilot.

Why it can't be borrowed.

Published canvassing effects assume a volunteer already exists. This model's cost, coverage and reach all hinge on recruitment happening by mail, cold, with no organizer in the county. There is no comparable number to lift.

Why it dominates everything.

At a 2% opt-in, the recruitment step dilutes any downstream turnout effect roughly fiftyfold. Which means a single small trial cannot honestly measure both recruitment and persuasion, and a design that claims to is overreaching.

The pilot · designed, not yet run

A randomized trial that answers the recruitment question cleanly.

So we split it. Run the recruitment question now as a properly powered mail experiment, and pre-register the turnout question as secondary -- as the input that sizes a second wave rather than a result we overclaim from this one.

Randomize neighborhoods

About 180 contiguous clusters of roughly 35 habitual voters and their neighbors, split one-to-one. A two-to-three door buffer is stripped at every cluster boundary. Because a captain only ever touches her ten nearest doors, keeping whole neighborhoods in one arm is what keeps treated and untreated neighbors genuinely separate.

Mail the ask

Roughly 3,100 recruitment postcards to the habitual voters in treatment clusters. Control clusters get nothing. Nonpartisan throughout: no party, no candidate, no issue.

Measure two things

Primary outcomes are the opt-in rate per postcard mailed, and the share of those who go on to open their block sheet. Neighbor turnout and the captains' own turnout are pre-registered as secondary and exploratory.

State the power honestly

At a true 2% opt-in, 3,100 pieces pin the recruitment rate to about six tenths of a percentage point either side. The turnout effect is underpowered by design, and a first pass understated how badly: correcting for the fact that we randomize whole neighborhoods rather than individuals, detecting the roughly one point we would actually expect takes on the order of 52,000 pieces and $18-29K -- something like seventeen counties this size. Better to know that before spending on it than after.

Commit to the decision rule in advance

At or above roughly 2% opt-in, the funnel clears and a multi-county activation trial is worth running. Below about 1%, the postcard alone does not recruit, and the honest response is to redesign the ask rather than spend more measuring turnout.

Cost: about 3,100 postcards at nonprofit print and postage runs $1,100-1,700. That is the mail only -- print, address certification, the nonprofit permit, producing the packs and the work of running it are not in that number. The voter-file match already exists. The opt-in capture page does not: the postcard QR currently resolves to this brochure site, and until that page is built the recruitment rate cannot actually be recorded. Building it is the first item of work, ahead of any of the arithmetic above.

What we don't claim

Honest about the limits.

  • The ten-door geometry is measured in one county. It is very likely to hold in similar places and we have not shown that it does.
  • The opt-in rate of 1-3% is an assumption, not a finding, and every downstream projection inherits that uncertainty.
  • An unpaid captain with no supervisor is an open question over an eight-week campaign. The pilot measures whether she starts, not whether she finishes.
  • Some addresses will not geocode and stay blank rather than being guessed at.
  • The voter file is current registration only; full historical address moves are not reconstructable.
  • Birth month and day are not in the public voter file, so date-based outputs use registration date as a public-data proxy and say so.
  • No paid APIs are required, and no voter data is uploaded to any third-party map or AI service. Address geocoding runs against local address-point files, with cached responses from the public Census geocoder as fallback.
How it's built

Local-first. Idempotent. Verifiable.

Local-first.

Reads statewide voter, history, polling-place, address-point and precinct-boundary files. Writes only to local output directories. Generated packets carry names and addresses and never enter version control.

Idempotent builds.

Every packet regenerates from scratch on each build, so there is no such thing as a stale one sitting in an output folder waiting to be mailed.

Verifiable.

Test suites, address-match-rate audits and run logs are first-class outputs. The sample packet on this site is regenerated by the same builders as a live one.

The pilot is designed and the county is picked.

What it needs is a few thousand dollars of mail, a capture page that does not exist yet, and a partner who wants the recruitment number to exist. If that is you, say so.

Fund or run a pilot →