Translation and untranslatability in elections and polling

What arrives

Four streams run into any competent campaign shop, and none of them stop. Voter registration files update daily as counties process new registrations, deaths, moves and party re-registrations. Turnout signals arrive as early-vote and absentee-ballot counts, updated hourly in states that report them, and as vote-history flags appended to the file after each election. Survey flow arrives in waves: a tracking poll fielded over three days, a robo-poll overnight, an internet panel refreshed weekly, each with its own sample frame, weighting scheme and margin of error. Media coverage arrives as sentiment, as earned-media volume, as the framing a candidate's gaffe receives on the two nights it is news before it stops being news.

Each of these streams carries terms whose meaning depends on the stream, not on a dictionary. "Likely voter" is not a fact about a person; it is a probability model's output, re-estimated against a turnout screen that itself gets rebuilt when a primary produces an unexpected result. "Independent" denotes whoever the state's registration form allows to tick that box, and what that box permits varies by state and sometimes by year. "Swing district" denotes whatever the last two cycles' margins say it denotes, and a single redistricting cycle can dissolve the referent entirely, folding the district that used to swing into one that no longer exists under that number.

What is held

A campaign does not, and should not, treat any of these terms as fixed. What it holds, if it is doing the work properly, is a belief about what each term currently denotes, tagged with where that belief came from and when it was last checked. "Likely voter" as of this Tuesday means: registered, contacted twice, self-reported vote intention above 7 on a 10-point scale, weighted against a turnout model calibrated on the 2022 midterm and not yet recalibrated for this cycle's early-vote surge. That whole sentence is the referent. Drop any clause and the term still gets used in memos, but it has quietly detached from what it is meant to point at.

The honest version of this holding is a provenance record: this poll's crosstabs, fielded these dates, this sample size, this weighting target, superseded by that later poll on this date because turnout composition shifted. Most shops keep something like this informally, in the analyst's head or in a spreadsheet nobody else reads. The formal version — versioned belief, explicit supersession, an audit trail showing which numbers were live on which day — is rare, and its absence is where the failure below comes from.

What triggers revision

Revision is triggered by disagreement between streams, not by any single stream crossing a threshold. A tracking poll shows the race steady at plus-three. Early-vote returns from a county that leaned the candidate's way in the last two cycles come in 4,000 ballots short of the pace needed to hit the turnout model's assumption. Registration data shows a spike in new independent registrations in the suburban counties that the poll's likely-voter screen was built on data that predates. None of these alone forces a rewrite. Together they are the signal that the term "our voters" has drifted from what it denoted when the strategy was set.

The mechanism that should catch this is boring and specific: a standing comparison between what the turnout model predicted for early vote by this date and what actually arrived, refreshed daily, with an alert when the gap exceeds a set number of standard deviations. Absent that mechanism, the disagreement sits unnoticed in three separate spreadsheets maintained by three separate vendors, and nobody's job is to reconcile them until the count is final and it no longer matters.

What the operator sees

The campaign analyst sees a dashboard, or the equivalent stitched together from vendor exports, and the dashboard is where untranslatability becomes visible or stays hidden. If the dashboard shows only the current poll's topline, the analyst sees a number with no history: plus-three, this week, full stop. If it shows the provenance — this topline, built on a likely-voter screen last recalibrated six weeks ago, against a turnout assumption now diverging from early-vote reality by nine points in the counties that decide the race — the analyst sees the actual state of the belief, including its expiry.

The failure mode is the first dashboard mistaken for the second. A strategy gets fixed on a snapshot: message discipline built around "independents" as they polled in March, media buys allocated against a battleground map drawn from 2020 turnout patterns, a closing argument aimed at suburban women who, by the numbers taken in September, have already moved. The term "independent voter" or "the suburbs" survives in every subsequent memo. What it denoted in March is not what it denotes in October, and nothing in the memo format forces anyone to notice the substitution. The word is fluent. The referent has moved on without it.

What it costs

Coverage costs money: continuous tracking polling, real-time voter file refreshes, staff whose job is reconciliation rather than production, all cost more than a single benchmark poll and a strategy memo written once. Trust costs attention: a term flagged as uncertain — "likely voter, low confidence, screen under review" — is harder to act on than a clean number, and campaigns under deadline pressure prefer the clean number even when it is wrong. Elapsed time costs the most and is least visible: every day between the last recalibration and election day is a day the referent can move without the belief moving with it, and the cost compounds silently until the exit polls or the actual count reveal the gap all at once.

Two objections worth taking seriously

Continuous polling gives you more numbers, not more certainty. You can survey every day until election day and never rule out that your sample is systematically missing the voters who decide the race — the 2016 and 2020 problem of the respondent who will not answer the phone.

This is right, and it is the polling version of Quine's indeterminacy: no volume of observation forces a unique reading of "who will vote." Streaming data does not settle the question of what a term denotes; it narrows the range of readings consistent with the evidence and forces the campaign to notice when that range widens rather than assuming it has stayed narrow. The gain is not certainty. It is that an analyst working from continuous, provenance-tagged intake sees the range widen in real time — a growing gap between two independent turnout signals — where an analyst working from a single benchmark poll sees nothing at all until the result arrives.

Refresh the model. Field a new poll each week, rebuild the turnout screen after the primary, and drift is handled — this is a maintenance schedule, not evidence for anything more elaborate.

This works for facts whose identity is stable and whose value changed, such as a candidate's approval rating. It fails where identity itself changed underneath a stable label. "Independent" measured against a 2020 registration file and "independent" measured against a 2024 file, after a wave of party re-registration in reaction to a primary fight, are not the same population wearing the same name. A refreshed poll returns the new number with no flag that the underlying category shifted, and a strategy built on cross-cycle comparison — "independents moved eight points since last time" — silently compares two different things as though they were one. What is needed is not a fresher poll but a record that says the category itself was redefined on this date, for this reason, so that the eight-point comparison can be checked rather than trusted.

The lineage this points toward is not a better polling product. It is a discipline: hold every term that matters to a campaign's decisions — likely voter, battleground, base, swing — as a belief with a timestamp and a source, not a fact settled once and reused. The frozen memo is a Large Language Model's failure in miniature: fluent, confident, and quietly describing an electorate that has already moved.