How Soval works

How we research, classify, and stand behind a verdict — and why the method matters more than the label.

We do not sell verdicts. We sell a method that produces verdicts you can defend.

The problem with TRUE / FALSE

Most automated fact-checkers collapse everything into true, false, or unverified — with a confidence percentage attached. That shape fails in three ways. It misrepresents disagreement: when credible experts genuinely disagree, a "48% confidence" reads as "leaning false," which is a statement the evidence doesn't make. It misrepresents absence: a number can't distinguish "we couldn't find enough evidence" from "the evidence is against this." And it isn't auditable: you can challenge a stated classification with named sources; you can't meaningfully challenge a 72%.

The claims people actually submit — viral political, health, and financial claims — concentrate exactly where that single-axis shape breaks down. So we rebuilt the verdict around how evidence actually behaves.

Five states, not a score

Every verified claim receives one of five evidence states:

The states aren't arranged on a true-to-false line — they describe distinct evidence situations, separated by how much quality evidence exists, which way it points, whether credible sources disagree, and whether the story is still developing. The full definitions live on the Evidence states page.

A confidence score is computed and stored internally, but never displayed. Exposing a percentage invites readers to over-weight numerical precision the underlying evidence doesn't justify.

How a claim becomes a verdict

A claim moves through five stages from submission to a stored, auditable evidence object:

1. Source gathering — a curated retrieval, not a search. We don't simply search the open web. Retrieval applies a maintained blocklist that excludes social platforms, user-generated content, and known low-quality aggregators. If the first pass returns no quality sources, a second pass with a reformulated query runs automatically.
2. Four axes before a state. Before any state is assigned, the evidence situation is characterized on four named axes: evidence weight (abundant, moderate, or sparse — judged by quality, not count), evidence direction (supports, contradicts, or genuinely mixed), dissent level (none, fringe only, or substantial), and whether the topic is actively developing.
3. State assignment. The state follows from the rules those axes imply, anchored by a hand-labeled calibration set of claims spanning every boundary case — maintained and sharpened as real submissions test the boundaries.
4. The named-subject safeguard. Claims about identifiable people and businesses carry a different category of consequence. A Debunked classification against a named subject must clear an additional threshold: abundant contradicting evidence, unambiguous direction, and a high confidence bar. If any condition fails, the verdict is held at Insufficient evidence — and the downgrade is logged, so every conservative call is visible in the audit trail rather than buried.
5. The evidence object. The output isn't a label. It's a structured record: the state, the four axes behind it, summaries of each side where sources disagree, and every source tagged with its own stance (supports / contradicts / neutral) and quality tier (official body, peer-reviewed, reputable media, aggregator).

Refusing to label is also a verdict. We track every time we made it, and why.

Why you see the sources

Two design choices matter here. Dissent is a first-class field — when credible sources disagree, both sides are summarized and shown, not footnoted. And because every source carries its own stance and quality tier, the rule that one authoritative source outweighs many weak ones is visible rather than merely asserted. Looking at a claim, you can see immediately whether the contradicting source is an official body while the supporting sources are aggregators.

What we deliberately refuse to do

A methodology is defined as much by what it refuses as by what it produces. We do not display confidence percentages — they invite false precision. We do not auto-balance verdicts — if the evidence is one-sided, the verdict is one-sided, and we don't manufacture dissent to appear neutral. We do not let source count override source quality. We do not return Debunked verdicts against named individuals or businesses without crossing the higher evidence threshold above. And we do not score people — no reputation profiles of named individuals, no credibility rankings.

The verdict is the headline. The method is the product.

The limits we name

No methodology is finished, and we'd rather name our limits than have you discover them. A verdict is a snapshot of the evidence at the moment of checking — every claim shows its check date, and developing claims are flagged for re-verification. Our retrieval and quality tiers are currently best-calibrated to English-language sources. Image-based claims rest on provenance tools whose state of the art is uneven, so we disclose the basis of every image verdict and decline strong verdicts where provenance is unsettled.

We update this page when the limits change, when new ones surface, or when the method sharpens.

Corrections

A verdict is only as good as our willingness to revisit it. If you believe a claim's evidence state is wrong — based on flawed sources, outdated evidence, or an error in our analysis — flag it directly from the claim page, or email us at info [at] sovalsocial [dot] com.

When an error is confirmed, we update the evidence object in place and log the change. We don't quietly revise a verdict and move on — the correction is visible on the claim page, consistent with the same audit-trail principle behind every refusal we track.

Corrections are evaluated against the same four axes and named-subject safeguards as the original verdict. A correction is not exempt from the thresholds that produced the verdict in the first place.