Why fact-checking single claims isn't enough

A health claim goes viral. A fact-checker picks it up, reads the studies, and publishes a careful verdict: false, or mostly false, or missing context. The correction earns a fraction of the original's reach, the creator posts three new videos that week, and the cycle resets. Everyone involved did their job, and somehow the audience is barely better off.

The problem is not that fact-checking is done badly. It is that the unit of analysis is wrong. A claim is one sentence from one moment; the thing audiences actually rely on is a person, speaking across years, with positions that shift and incentives that don't appear in any single post.

What a single-claim verdict can't see

Treat one statement in isolation and you lose at least four dimensions that matter more than the verdict.

• Stance across time. Was this claim a one-off, or the fortieth time the creator has taken the same position on the same topic? Did they oppose it two years ago and flip? A verdict on today's post can't say.
• Certainty. There is a real difference between "early research is interesting" and "this will change your sleep, use my code." Binary true/false flattens exactly the dimension — how sure the speaker claimed to be — that determines how much the audience trusts them.
• Incentives. A fact check evaluates the proposition. It rarely asks whether the person asserting it has a commercial relationship with a brand in that category — the question the FTC's disclosure rules treat as central to reading any endorsement honestly.
• Selection. Fact-checkers triage by virality, so the record of any one creator is a scattering of their most shareable moments. The everyday claims — repeated weekly to a loyal audience, too small to check — are where most of the influence lives.

Medicine faced a structurally similar problem and named the fix decades ago: stop over-reading single studies and synthesize the whole body of evidence — the logic behind systematic reviews of the kind Cochrane publishes. Judging a creator by one viral claim is like judging a research question by one small trial. The honest answer is almost always the aggregate.

From verdicts to a ledger

This is the design premise of StanceGraph. Instead of adjudicating one post, we build the aggregate object: every checkable health claim a creator has made in public, each recorded with its stance, its certainty as phrased, an evidence grade for the underlying proposition, and any commercial tie on the same topic — with a receipt behind every record. Source, verbatim quote, timestamp, extraction model version.

A ledger like that answers questions a verdict cannot:

• Does this person's confidence track the evidence, or run ahead of it?
• Which topics do they only ever discuss while a sponsor is attached?
• When the research moved, did their stance move with it?
• Is the claim in my feed today consistent with what they said last year?

Notice that none of these questions requires declaring anyone a fraud. Most creators are not frauds. They are people with audiences, partial information and real financial incentives, and their reliability varies by topic in ways a ledger makes legible. Some report cards will read as reassuring — consistent stances, hedged where evidence is thin, ties disclosed. That result is just as valuable as the damning one, and a single-claim fact check can produce neither.

Corrections still matter

None of this replaces claim-level verification — when a specific viral assertion needs a careful rebuttal, fact-checkers do work we rely on and link to. A stance ledger is the complement: it is how you decide how much weight to give the next claim before anyone has checked it, which is the situation audiences are in almost all of the time.

There is also a practical reason to prefer the ledger: it scales where rebuttals cannot. A verdict has to be written; a claim record has to be extracted, receipted and reviewed — slower per item, but it accumulates. Every new video adds rows to an existing report card instead of starting a new argument from zero.

One sentence can be checked. A person has to be mapped. We think the second job has been the missing one, and it is the one StanceGraph is built to do — receipts included.