how the SNR score works
Every item on this site carries a signal-to-noise score from 1 to 5, and every scored registry figure carries one too. The score is not an opinion about whether a claim is important; it is a reading of how well the sources support it. Publishing an early signal at SNR 1 is the system working as intended. Publishing a weak claim dressed as a certainty is the failure it exists to prevent. Here is the whole mechanism.
the scale
| score | what it means |
|---|---|
| SNR 1 | Low confidence. A single source, a rumour, or an out-of-pattern claim with little behind it. |
| SNR 2 | A little more: more than one source, one usually-reliable source, or an early signal that later reporting matched (a 1 upgraded retroactively; see below). |
| SNR 3 | A few reputable sources: trade press, established media, or an industry-leader account. |
| SNR 4 | Widely reported: many sources, an established aggregator, or a long-standing claim nothing has contradicted. |
| SNR 5 | Quasi-certainty. The actor itself or an official record: a release on the company's own site, an official filing, or direct observational data. |
The math is code, not judgment calls: a deterministic scoring engine computes every score from recorded inputs, and the agent that drafts an item can attest facts about its sources but can never set the number.
where a score starts
The best source attached to an item sets the base tier. First-party statements, official records, and directly computed data start at 5. Press-wire copy and established aggregators start at 4. Trade and mainstream press start at 3, as does a person on our curated signals list before any floor applies. An informal but identifiable source starts at 1. Sources that cannot be named at all do not publish, whatever else they would score.
The test for first-party is strict: could the linked page be wrong about the fact without the actor or an official record being wrong? If yes, it is not first-party. And a source that has repeatedly burned us can lose its class entirely (see grading, below).
what counts as one source
Before corroboration is counted, sources are collapsed into corroboration units, because syndication is the cheapest way to fake breadth. URL variants of one article are one unit. Multiple pages on one registrable domain are one unit. And two articles whose headlines are near-identical are one unit even across domains: headlines are fingerprinted (a 64-bit SimHash over normalized title words) and anything within a Hamming distance of 3 is treated as the same wire story rewritten. The item keeps every link for the reader; the units drive the math; every collapse is logged in the sweep entry on the log.
how corroboration moves it
Independent corroboration raises a score, and each rule fires at most once per claim: a second distinct unit, a fourth, and pickup by a mainstream outlet beyond the lead reporter. Corroboration is tested, not assumed: for any claim resting on second-hand reporting, the machine actively searches the open web for other coverage. A search that ran and found nothing costs one point, because "nothing else reports this" is itself a claim about the world and it should hurt to be wrong about it. A direct source proves its own statement and pays no such penalty; and a search that never ran (the sweep's crawl budget ran out first) costs nothing, but is recorded as not attempted rather than dressed up as a result.
There is a hard ceiling. No amount of second-hand corroboration reaches 5. Wide reporting IS the definition of 4; 5 is reserved for a direct source, the actor speaking for itself or an official record.
scores keep moving after publication
Two rules lift a score with time, both bounded, both automatic, both visible:
Reinforcement. When a matching event lands 8 to 30 days after an item published at SNR 1 or 2, the item is bumped by one and the new source is attached. An early lone signal that later reporting matches was early, not wrong, and the score says so retroactively. Once per item, only from 1 or 2, only inside the window.
Persistence. An item still below 4 that survives 14 days with nothing contradicting it earns one point, once, and can never pass 4 this way. Time is weak evidence; it counts a little and caps early.
Every movement, up or down, is appended to the item's stored calculation and listed in that sweep's entry on the log. Trace history is append-only: earlier steps are never rewritten to flatter the present score.
claims that have to earn it
An out-of-pattern or extraordinary claim starts at 1 whatever its source count, and climbs only through corroboration and survival. Any claim big enough to reshape the market whose best source is below first-party is treated as extraordinary automatically, by code, and additionally queued for human review even while it publishes. Extraordinary claims are the honest-calibration stress test: they run on the same rules, just from the bottom.
when sources disagree
A mismatch of metrics (one source counts launched satellites, another counts operational ones) is annotated, never punished: both numbers can be true. A genuine conflict on the same metric lets the better-sourced side lead and costs the loser a point. Two equally sourced claims that conflict are both marked disputed, kept visible side by side, and queued for a human ruling; the site never quietly picks a winner it cannot justify.
the signals-list floor
The signals page lists people we have individually verified and chosen to trust. When one of them states an on-topic fact on a verified channel, the claim is floored at 4 as an observer, and at 5 when the person speaks for the actor concerned about itself. The floor covers factual statements only: jokes, opinions, and off-topic posts get nothing. The list is curated by a human and the software that ingests the web cannot edit it.
fakes and spoofs
Fake press releases are a documented attack on trackers like this one, so the two highest source classes are gated by domain: a page only counts as first-party or official record when its domain matches the actor's registry-recorded website or an official register. A press-wire copy of an announcement caps at 4 until the actor's own domain confirms it. A superlative in a first-party release ("largest constellation") is attributed as a statement, never scored or repeated as fact.
sources are graded too
Every source domain carries a rolling reliability record, rendered on the log: strikes for claims that lost a same-metric contradiction, credits for claims that started at 1 or 2 and were later confirmed. Repeated strikes inside a 90-day window demote a source's class in future scoring; demotion decays, and confirmed claims win the class back. Sources that repeatedly produce independently confirmed early claims are suggested for the signals list, but a human makes that call.
what the score is not
It is not importance. Importance is the separate impact label (seismic, major, notable, noise), and the two axes are independent: a seismic rumour is seismic AND low-SNR at the same time. It is not an endorsement of opinions: commentary items score the attribution (this person said this, here), never the take itself. And it is not a promise of truth; it is a promise that the confidence shown matches the evidence held.
stored, shown, and checked
Every score is saved with its full calculation at the moment it was set: the base source tier, every adjustment since, and the two components (source class and corroboration) that fused into the integer. You can open that calculation under any score mark on the site.
Whether the scores are honest is itself measured. Each claim's score at publication is recorded permanently, even after later bumps change what the item displays, and compared against how the claim resolves: confirmed independently, debunked, or expired quiet. The running tally per score level is public on the sweep log. If our 2s turn out right as often as our 4s, the scale is broken and the record will show it.