Evidence independence: count mechanisms, not documents
Evidence independence depends on how observations were produced, not on how many documents repeat them. Separate the number of files, the number of underlying measurements and the number of genuinely different tests of your explanation. Ten reports can share a survey, an assumption or a selection rule, making them less informative than their apparent variety suggests.
Go beyond tracing quotations
A citation chain tells you who repeated whose statement. Independence analysis asks another question: could these observations be wrong for the same reason? Two publications can contain original writing and still rely on the same customer panel. Conversely, two measurements from one organization may use different collection methods, although their institutional incentives may remain related.
The W3C PROV overview describes provenance through entities, activities and agents. Use that vocabulary to map inputs and production steps; it does not provide a statistical independence certificate. The evidence-cluster worksheet here is an editorial tool for exposing shared mechanisms, rather than a formal confidence score endorsed by W3C.
A hypothetical demand investigation
Assume you are checking whether demand for an invented industrial component is broadening. Report A surveys distributors. Report B estimates revenue from that same distributor survey. Report C interviews analysts who cite the survey. Report D examines a separate sample of end-customer purchase orders. Four files do not constitute four independent observations of demand.
| Record | Measurement behind it | Shared vulnerability | Research use |
|---|---|---|---|
| A | Distributor panel | Panel excludes small buyers | Direct but limited observation |
| B | A’s panel plus pricing assumption | Same exclusions and model assumptions | Derived interpretation |
| C | Commentary based on A | Same panel plus quotation selection | Context, not extra confirmation |
| D | End-customer order sample | Different sample and order definitions | Another test, subject to reconciliation |
Group A, B and C into one measurement cluster. D deserves its own description, but it does not automatically validate A: orders may be cancellable, while distributors report shipments. Before combining them, align period, product scope and the part of the demand process each observes. Different sources can disagree without either being fabricated.
Record three kinds of dependence
First, identify shared data. Look for the survey provider, company disclosure, raw dataset or transaction feed behind each record. Different presentation and branding do not create new measurements. Mark an origin as unknown if you cannot establish it; an untraceable dataset should not be promoted to independent evidence merely because its provenance is obscure.
Second, identify shared assumptions. Two forecasts built from separate datasets may both assume the same price increase, exchange rate or market size estimate. Agreement in their final number can reflect that common input. Request the assumption table or record its absence. Do not treat a model’s distinctive wording as proof that it reasoned from different information.
Third, identify shared selection. Studies may independently collect data but choose only successful customers, large public issuers or recently active wallets. Those shared filters can leave the same blind spot. Write the omitted population beside the cluster rather than burying it in a general disclaimer at the end of the memo.
Use clusters without inventing weights
Summarize what each cluster tests and which uncertainty remains. A practical memo might say one distributor measurement and one end-customer measurement support different stages of the mechanism, with cancellation risk unresolved. That is more inspectable than stating four sources confirm demand or assigning an unexplained confidence percentage.
Do not multiply probabilities or award equal points to clusters unless you have a defensible model for their dependence and error. The worksheet is not that model. When a shared source is corrected, revisit the entire dependent cluster; do not count three corrections as three new adverse events. Preserve the distinct contribution of genuinely new measurement methods while keeping their own limitations visible.
FAQ
Does a separate publisher mean independent evidence?
No. Inspect the underlying observations, assumptions and population. Separate editorial teams can still inherit the same measurement error.
Should all dependent reports be discarded?
No. They can offer useful interpretation or reveal disagreement. Retain their role without counting each as another independent observation of the same fact.