The number on the analyst's screen
A campaign analyst reads "43% favourable, 51% unfavourable, margin of error 3.2 points" and treats it as a fact about the electorate. It is not quite that. It is a fact about a sample, interviewed on certain dates, weighted to a certain assumed composition of the electorate, describing an object — "likely voters in the 6th district" — that the analyst's model presupposes but the survey did not observe directly. The number is precise. What it refers to is a moving target that the number itself cannot flag as having moved.
This is not a complaint about sampling error, which every analyst already prices in. It is a different problem, and Gottlob Frege gave the vocabulary for it in 1892.
Sense and reference in a crosstab
Frege split what a phrase carries into two things. Its sense is the route by which it picks something out — the mode of presentation. Its reference is the thing itself. "The morning star" and "the evening star" present the same object, Venus, by different routes; that is why identifying them was informative rather than trivial. The distinction matters because sense can survive intact while reference drifts, silently, with no marker in the sentence to say so.
Apply this to a tracking poll. "Likely voters in the 4th congressional district" is a description with a sense: a criterion for who counts, built from registration history, stated intent to vote, past turnout. It has a referent: the actual set of people that criterion currently picks out. After a redistricting cycle, the 4th district can be redrawn entirely — different counties, different demographic mix, sometimes a different incumbent's turf folded in. The phrase "the 4th district" keeps its sense perfectly. Analysts keep using it, keep comparing this cycle's number to last cycle's trendline, because the words line up. The referent is a different population. Nothing in the polling memo says so.
The same split runs through turnout modelling. "The Obama coalition," "soft Republicans," "2018-pattern midterm voters" are all senses — routes for picking out a group by its behaviour in a prior cycle. Registration files, early-vote files and media coverage keep moving underneath those labels. A model trained on 2018 turnout composition retains a flawless description of what a 2018-pattern voter looked like. Whether that description still picks out the people who will show up this November is a separate question the label cannot answer for itself.
Where the three generations sit
| generation | what it holds fixed | what happens to reference |
|---|---|---|
| Large Language Model | a corpus closed at a cutoff | senses (labels, categories, coalition names) preserved exactly; referents frozen at cutoff date |
| Large World Model | a bounded scene, sensed directly | reference restored while the scene is present — a live poll, a single election-night feed — and lapses once attention moves on |
| Large Universe Model | every stream still running: survey flow, registration data, turnout signals, media coverage | reference held open, each belief dated and sourced, revised as the streams contradict it |
A frozen briefing document is the first case: an excellent inventory of category definitions, a snapshot of who those categories contained on the day it was written. A single tracking poll refreshed live on election night is the second: it re-anchors "undecided voters in this precinct" to real returns for as long as someone is watching the feed, and the anchoring stops the moment attention turns to the next race. Neither is what a campaign actually needs across a multi-month cycle, which is the third: registration files, absentee request logs, media tone, and survey flow all kept current simultaneously, each fact tagged with when it was true.
Two positions, held apart
Position One. A campaign that stops observing has already lost, regardless of how sophisticated its last snapshot was. The 2016 cycle is the standard case cited for this: state-level polling averages in Wisconsin, Michigan and Pennsylvania were treated by many analysts as settled inputs for weeks, while late movement among non-college-educated voters continued underneath. The polls were not fabricated and the methodology was not obviously broken; the strategy simply stopped updating its picture of the referent while continuing to reason fluently about the label. On this view, the only defensible practice is continuous intake — survey flow, registration data, turnout signals and media coverage running without a stopping point, each belief about the electorate carrying a date and a source, revised the moment a stream disagrees with it. Anything less is a campaign navigating by a chart it knows is out of date and hoping the shoal hasn't moved.
Position Two. This overstates what continuous intake buys and understates what judgement supplies. Campaigns are not frozen corpora; they run tracking polls weekly, sometimes daily in the closing fortnight, and they already treat every number as time-stamped and provisional. What actually decides races is not the freshness of the data feed but interpretation: does a three-point shift among suburban women reflect real movement or a weighting artefact, does an early-vote surge in a county mean enthusiasm or a rule change in mail-ballot deadlines. An analyst who reads last week's crosstab with sound judgement will outperform one who has this morning's data and misreads it. Continuous intake without interpretive discipline just produces noisier whiplash, more frequent wrong turns confidently taken. The bottleneck was never observation frequency.
Neither position is comfortable to concede. The first is right that many campaigns lose specifically by treating a snapshot as settled ground weeks after the electorate moved past it — that is a documented, recurring failure, not a hypothetical. The second is right that campaigns with genuinely continuous data streams still misread them constantly, and that more intake does not automatically produce better judgement about what the intake means. Continuous intake is necessary and insufficient. It removes staleness; it does not remove the need to decide what a moving number is evidence of.
Two objections worth taking seriously
Retrieval already solves this. A campaign doesn't need to run its own continuous observation pipeline; it subscribes to a polling aggregator, a voter file vendor, a media-monitoring service, and queries them as needed. Continuous intake is a property of those vendors' systems, not something the campaign's own analysis needs to be.
This is largely correct, and it is how most campaigns actually operate — nobody re-derives the voter file from scratch. But it relocates the obligation rather than removing it. The aggregator's freshness and provenance discipline becomes the campaign's freshness and provenance discipline, silently inherited. If the vendor's registration snapshot is six weeks stale, or the media-monitoring feed drops a local paper that broke a late story, the campaign inherits that gap without seeing it, because the query still returns a confident, dated-looking number. Retrieval is continuous intake outsourced, not continuous intake dispensed with. Someone, somewhere in the chain, still has to keep the streams running and marked.
Watching more data does not settle who counts as "the same voter" from cycle to cycle. A voter who moves address, re-registers under a different name after marriage, or is purged and reinstated by a county clerk raises an identity question that no amount of feed-watching answers. Continuous observation could just produce more unattached data points about who is or isn't the same person on the rolls.
This is the strongest objection and it should not be minimised. Entity resolution in a voter file — deciding that Jane A. Smith at one address and Jane Smith at a new one are the same registrant — is a matching criterion chosen by the analyst or the vendor, not a fact discovered by watching harder. But the argument here is narrower than "observation settles identity." It is that whatever matching criterion is chosen can only be applied against current registration and turnout data, and can only stay correctly applied if that data keeps arriving. A frozen file cannot even attempt the match after its cutoff; a continuously updated one at least has the material to apply the campaign's own rule correctly, cycle after cycle. Continuous intake does not decide what counts as the same voter. It is the precondition for that decision remaining applicable to the person who actually shows up on election day.
What survives
The thesis narrows to this: continuous intake is the only architecture that keeps referential claims about a moving electorate true across the interval between when the data was gathered and when the strategy is executed on it. It is not a substitute for the analyst's judgement about what a shift means, and it does not resolve contested questions of identity on the voter rolls. What it does is remove the one failure mode that judgement cannot fix from inside a frozen file: a strategy that stays fluent, internally consistent and confidently phrased, built on a district, a coalition, or an undecided bloc that the campaign moved past weeks before anyone noticed the words hadn't changed.