Independence

The intuition that breaks first is that more sources means more confidence. Five providers agreeing is stronger evidence than one only if the five are looking at different things. If they all read one feed, five agreements are one observation reported five times, and the confidence built on it is confidence in a copy. The failure is quiet, because a consensus label looks the same whether it was earned or manufactured.

What went wrong before it was measured

Against five providers trading under five names and reading one shared upstream, the engine produced 575 confidently wrong consensus deliveries. Afterwards, zero. The data did not improve and no provider changed anything. The label stopped claiming corroboration that never existed.

Two social providers tested at 100% exact agreement across eleven handles, identical to the byte. Real independent measurements of a social account do not agree to the byte; they disagree in the last digit constantly. Byte-identical agreement is the signature of one pipeline, not two.

How it is measured

Sources are compared on the residual correlation of their disagreements over time rather than on what they say about their own pipelines. Two exchange feeds tested at r=0.154are treated as independent because the measurement shows them moving apart under conditions where a shared upstream would keep them locked together. A pair that tracks too closely is collapsed into a single observation regardless of who operates it.

The estimate is maintained continuously rather than fixed at onboarding, because independence is a property that decays. A provider that switches to a cheaper upstream has not told anyone, and will not. The measurement notices.

Independence of the anchor, not only of peers

The same question applies one level up. An anchor is only worth the independence it has from the answer it is checking. Chainlink publishes the basis of each Solana feed, and the trace now carries it: a feed computed from on-chain pool state agreeing with a price also derived from on-chain pools may be reading the same liquidity twice.

The judgement is per call rather than per pair, because it depends on which source answered. The same token can be checked by an independent reference on one call and by a near-relative on the next, and the trace says which happened. Where the two share a basis it says so plainly: agreement here is weaker than an independent check.

What it does not do is call that a correlation. Nobody has measured the residual correlation between those two, the way it was measured for the exchange feeds. Reporting a shared basis is a fact. Inferring a correlation from it would be the same sin as assuming independence, pointed the other way.

Declared upstreams help, but never decide

Providers are asked to name their upstreams when they apply, and to say whether they originate, aggregate or relay. Aggregators are asked how they aggregate, because two aggregators can look independent while reading the same three exchanges with different weights. That declaration is genuinely useful: it tells us where to look.

It is never the evidence. A provider can be wrong about their own supply chain without intending to mislead anyone, and the correlation measurement does not care what anyone intended.

The case that cannot be solved

Two sources that lie together, in a category with no ground truth, are informationally indistinguishable from the truth. No statistical judge resolves that. Metera does not pretend otherwise: answers in that position are labelled unconfirmed every time. Across every adversarial scenario the mislabelled rate was 0%, which is a claim about honesty rather than about accuracy, and honesty is the one that can actually be guaranteed.