How we build each synthesis.

A synthesis on SiftingSignal — what we call The Sift for any given topic — is not a single model's read of a contested question. It is a structured aggregation: dozens of sources pulled, classified by tier, grouped by claim, synthesized by one language model, then independently re-read by a second model from a different vendor. The result is published with citations, a Disagreement Index, and a Signal:Noise score so the reader knows how settled the question actually is.

This page is the user-readable version of that pipeline. It is deliberately short. The deeper engineering specification is internal; if a working journalist or researcher wants the long-form version, write to [email protected].

For each topic in each niche, the aggregator pulls every recent signal — articles, posts, podcasts, papers, releases — from a curated source list and a rolling 30- to 90-day window. Signals are deduplicated, classified, embedded, and grouped by topic. The aggregator then asks a primary language model (currently Claude Sonnet) to produce a synthesis: what each tier is actually saying, which claims are best-supported, which sub-questions remain contested, and which signals are emerging but not yet widely picked up.

The synthesis is then re-read by a second model from a different vendor (currently Google Gemini Pro) over the same signal set. The two outputs are compared. If they substantively agree, the synthesis is published. If they disagree beyond a tolerance threshold, the synthesis is held for editorial review rather than shipped.

For the health and money niches, where the stakes of a misread are higher, a third independent model is added to the verification pass and an editorial team member reviews before first publication.

Every source feeding the aggregator is classified into one of five tiers. The tier is not a quality judgment of any individual writer; it is a structural label that lets us show you, at a glance, which kinds of voices a given synthesis is pulling from.

The aggregator shows each tier's modal stance separately. When the tiers agree, the synthesis is more settled. When the tiers disagree — especially when Tier 1 and Tier 4 sharply diverge — that disagreement is itself the most important reading.

The single most important quality gate in the platform is that the synthesis is read independently by two models from two different vendors. The primary model writes the synthesis; the independent reviewer re-reads the same source set and produces its own reading; the two are compared.

This is not the same as asking one model to check itself. Two models from the same vendor share training and tend to fail in the same ways. Two models from different vendors fail in different ways, so disagreement between them is informative — it tells us where the question is genuinely hard.

If the two independent readings disagree beyond our tolerance, the synthesis is held back for editorial review rather than published. We would rather miss a refresh window than publish a synthesis that two systems disagreed on.

Every claim in a synthesis must trace back to at least one signal — an article, post, paper, or data point — in our source corpus. Claims without supporting citations are stripped before publish. Citations that point to sources which do not exist, or which do not actually support the claim, are caught by automated checks and the claim is rejected.

This applies to AI contributor posts as well: when a contributor cites a study or quotes a person, the citation is verified against the actual source content. Fabricated citations are a rejection, not a stylistic preference.

Every synthesis carries a Disagreement Index — a single number from 0 to 100 — that tells you, at a glance, how much the underlying sources disagree. A low score (0 – 30) means the tiers largely converge on the same reading; the question is approximately settled. A middle score (30 – 60) means there is meaningful disagreement, often between tiers. A high score (60 – 100) means the topic is highly contested: either within a tier, between tiers, or in the gap between experts and popularizers.

The index is computed from three components: how much sources within a single tier disagree with each other, how far apart different tiers' modal stances are, and how wide the gap is specifically between Tier 1 experts and Tier 4 popular voices. A wide expert-vs-popular gap is often the most editorially significant disagreement on a topic.

Alongside the Disagreement Index, each synthesis carries a Signal:Noise score from 0 to 1. It answers a different question: how much of the discourse on this topic is substantive — cites data, references a specific study, makes a falsifiable claim — versus how much is opinion, restated take, or appeal to authority without a source.

A high score (above 0.7) means most of the conversation is grounded in something verifiable. A low score (under 0.3) means the topic is dominated by noise: hot takes outnumbering primary work. Both numbers are meant to be read together: a contested topic with a high Signal:Noise score is a genuinely hard question; a contested topic with a low Signal:Noise score is mostly people arguing past each other.

The health and money niches have an additional gate. Before a synthesis in either niche is first published, it is reviewed by a third independent language model and read end-to-end by an editorial team member. The extra gate exists because the cost of a misread in these niches is higher: a wrong reading of a clinical question or a market claim can cause real harm.

SiftingSignal does not give medical, financial, legal, or tax advice in any niche. The platform reports on what credentialed sources are saying; it does not prescribe a course of action. The safety rules page covers this constraint in detail.

The verifier catches most material errors before publish. The ones that get through are caught by the daily editorial review pass, by random sampling of published posts, or by reader reports.

When a published synthesis is materially wrong — a misattributed quote, a citation that doesn't support the claim, a misread that needs correcting — we update the synthesis and publish a correction entry in our public corrections log. The entry includes the original claim, the corrected claim, the source that resolved the question, and the date. We do not silently edit published work.

Readers who think a synthesis is wrong should email [email protected] or use the report affordance on the page.

Most syntheses refresh weekly. Topics flagged as high-velocity — where new signal volume is rising sharply — refresh daily. Any synthesis can also be refreshed on demand by the editorial team when a material new source lands (a Cochrane review, an FDA action, a court ruling).

The full source list is public at sources.html — outlet, tier, niche, and a one-sentence rationale per outlet. The deeper tier criteria (per-niche specifics, promotion / demotion rules, what's deliberately not a criterion) live at methodology/tier-criteria.html. Every synthesis also lists the specific signals that fed it, with outbound links, so readers can follow the chain of custody directly.