datumwise, an independent open-source research project

Version 2.0
Published: 17 August 2026
DOI: 10.5281/zenodo.21980262
License: CC BY 4.0
Supersedes: A Primer on the Statistical Bridge, Version 1.1, DOI 10.5281/zenodo.21966876.
Companion framework: The Statistical Bridge, Version 3.0, DOI 10.5281/zenodo.21979821.

Statistics is often approached from two opposite directions.

From one side comes the thought:

If the mathematics is correct, isn’t the statistical conclusion correct?

From the other:

If the data are accurate and the calculation is correct, isn’t the reported conclusion simply a fact?

Both questions are tempting.

Both miss something essential.

Correct mathematics tells us what follows from formal premises. Correct data processing tells us what was recorded or calculated from records. Neither, by itself, tells us why those records support a claim about a population, a future outcome, a mechanism, a causal effect, or the world more generally.

Something has to connect the two.

That something is not a decorative step around statistical analysis.

Statistical analysis is the governed bridge between evidence and the claims we want evidence to support.

This primer introduces that idea.

It is intentionally incomplete.

Its purpose is not to reproduce the full architecture of The Statistical Bridge. It is to make a few distinctions difficult to forget, show why they matter in ordinary analysis, and leave visible the deeper questions that the full framework answers.

A familiar request

Suppose someone asks:

What is average revenue per active customer this quarter, with a standard error?

The request sounds routine.

You may have login records, transaction records, customer identifiers, dates, amounts, and status fields.

A natural analysis is:

1. decide which customers are active;
2. total revenue by customer;
3. compute the average;
4. compute a standard error;
5. report the result.

Every line of code could be correct.

And yet almost every important statistical question could still be unresolved.

What is one point in the analysis?

A customer? An account? A customer-quarter?

Which points exist if no transaction occurred?

Does no transaction mean revenue is zero, or that the transaction stream is incomplete?

Are the observed customers the entire target population, or a sample from something larger?

Why is a standard error needed?

Where did the probability behind that standard error come from?

And even if the number is valid for this quarter, what would allow us to say anything about next quarter?

These are not edge cases.

They are the bridge.

The bridge in one picture

At the broadest level, statistical work relates realized evidence to a target we want to reason about.

[ \text{realized evidence} \rightleftarrows \text{formal target} ]

Realized evidence might include transactions, measurements, survey responses, failures, experimental outcomes, or other recorded empirical material.

A target might be a finite-population mean, a parameter, a future outcome, a causal effect, a latent state, a model, or an explanation.

The arrow from target toward possible evidence asks:

If this account were relevant, how could evidence like this arise?

The arrow from realized evidence back toward the target asks:

Given what actually happened, what does the evidence support?

That already tells us something important:

Evidence does not explain itself, and mathematics does not apply itself.

But Version 3 of the Statistical Bridge goes one step further.

It asks us to keep four obligations distinct:

[ \boxed{ \text{Bridge Constitution} \rightarrow \text{Probability Source} \rightarrow \text{Inference Certificate} \rightarrow \text{Claim License} } ]

Do not read this as a mandatory workflow.

A real investigation may move backward, forward, and sideways.

The sequence is a list of questions that eventually have to be answered.

1. Bridge constitution: what exactly are we analyzing?

Before asking which statistical method to use, there is a more basic question:

What empirical object has actually been constituted?

A table does not answer this merely by having rows.

Consider transactions.

A transaction record is an event: the point exists because an occurrence was recorded.

Now consider all registered customer-quarters during a completed quarter.

Those points can exist whether or not any transaction happened. They form a spine: the analytical points are established independently and then await values.

The distinction is easy to remember:

Events generate points. Spines establish points and await values.

This matters because an event table naturally omits customers with no events.

If we calculate average revenue from transaction-bearing customers only, we may have silently changed the target from:

all eligible customer-quarters

to:

customer-quarters with recorded transactions.

The arithmetic may still be perfect.

The population is different.

Event and spine are not the whole Statistical Bridge. They are one recurring geometry that makes a common problem visible:

[ E^{(r)} \rightleftarrows S^{(r)}. ]

The superscript (r) reminds us that values arise under some regime or arrangement: observational, experimental, policy, future, or otherwise.

But geometry alone is not enough.

Two things can line up without one being evidence for the other.

The full paper asks what must make the crossing lawful.

2. Variation is not a probability source

Now return to the customer-quarter example.

Suppose the quarter is complete.

Every eligible customer-quarter is known.

Every qualifying transaction has been captured.

Revenue has been established for every point, including legitimate zeroes.

The revenues vary.

Do we need a standard error?

Not necessarily.

If the question is simply:

What was mean revenue over this complete finite population this quarter?

then the answer is determined by the constituted data:

[ \bar y

\frac{1}{N} \sum_{i=1}^{N} y_i. ]

The individual values can vary enormously.

That variation does not by itself create uncertainty about the finite mean we just computed.

This gives one of the most useful rules in the Statistical Bridge:

[ \boxed{ \text{variation} \neq \text{probability source}. } ]

Before choosing frequentist, Bayesian, bootstrap, or any other machinery, ask:

What is uncertain?

Maybe the quarter is only a sample from a larger population.

Maybe the transactions are measured with error.

Maybe the completed quarter is being used to predict future quarters.

Maybe we care about a latent process rather than the finite realized population.

Those are all possible reasons for probability.

But the probability has to come from somewhere.

The visible spread of the numbers is not enough.

Return to the revenue example. Suppose the completed-quarter mean is exactly:

[ $137. ]

For the finite completed population, that number may need no inferential probability at all.

But now change the question:

What does $137 tell us about next quarter?

The number has not changed. The target has.

A future-quarter claim needs a bridge to a future-time target and some probability source for what can vary between now and then.

The full Statistical Bridge develops this question much more sharply:

Where does probability enter the analysis?

That question turns out to separate several things statistical practice often bundles together.

3. A probability source is not an inference certificate

Suppose probability really is needed.

Then we still need to ask what role it plays.

A randomized sample gives one kind of probability source.

A stochastic measurement process gives another.

A model for future outcomes gives another.

A Bayesian prior places probability on target-side unknowns.

These are not interchangeable.

Version 3 makes a useful distinction between two broad locations.

An evidence-side probability source governs how possible evidence could arise.

A target-side probability source places probability over something on the target side.

For a simple Bayesian model, the distinction can be seen as:

[ p(y\mid\theta) ]

on the evidence side, and

[ \pi(\theta) ]

on the target side.

After the evidence (y) is observed, we might derive:

[ \pi(\theta\mid y). ]

That posterior is not the original probability source.

It is an inferential result produced under the declared construction.

The Statistical Bridge calls the formal statement carrying inferential authority an inference certificate.

This gives another memorable distinction:

[ \boxed{ \text{probability source} \neq \text{inference certificate}. } ]

A frequentist confidence procedure may carry one kind of certificate.

A Bayesian posterior statement carries another.

Likelihood, fiducial, predictive, betting, and other approaches can carry still others.

The Primer does not need to classify them.

The important point is simpler:

A number, interval, or posterior does not manufacture the probability structure that gives it meaning.

If someone reports a “95% interval,” there is still a question:

What exactly is the 95% statement, and relative to what probability source?

Now the $137 example acquires another layer.

Suppose we build a future-revenue model and obtain a predictive interval for next quarter.

The observed $137 is still part of the evidence. But the predictive interval is a different object: it is an inference certificate produced under the declared future-revenue probability source.

So:

[ $137 ]

does not carry one fixed statistical meaning merely because the number is unchanged.

The full paper answers this with an open-typed certificate architecture.

That is one place where this Primer deliberately stops.

4. A valid certificate is not yet a world claim

Suppose now that the probability source is appropriate and the inference certificate is valid.

Are we done?

No.

Imagine that the analysis supports this statement:

Mean revenue among eligible account-quarter points in Q2 2026 was $137.

That may be a perfectly valid completed-quarter claim.

But now someone says:

So customers will average about $137 next quarter.

Something changed.

The calculation did not.

The claim did.

The new statement travels to a different time.

It needs additional assumptions about stability and transport.

Now suppose someone says:

Engagement causes about $137 of quarterly revenue.

The claim moved even farther.

We now need a causal target, an intervention regime, identification assumptions, and a lawful relation between the evidence and that causal claim.

A correct statistic cannot supply these merely by being correct.

The Statistical Bridge calls this final boundary claim license.

The same $137 can therefore participate in three very different statements:

Descriptive: Mean revenue over all eligible account-quarter points in the completed quarter was $137.

Predictive: Under a declared future-revenue model, next-quarter mean revenue is expected to lie within some stated range.

Causal: Customer engagement causes approximately $137 of quarterly revenue.

The first may be fully supported by complete constituted data. The second needs a future-time probability source and a predictive certificate. The third needs an intervention target, causal identification, and an appropriate regime passage.

[ \boxed{ $137 ;\not\Rightarrow; \text{one fixed statistical meaning}. } ]

A useful review asks:

• Which population?
• Which time?
• Under which regime?
• What permits transport elsewhere?
• How sensitive is the conclusion to uncertain assumptions?

The principle is:

[ \boxed{ \text{valid inference certificate} \not\Rightarrow \text{unlimited world-facing claim}. } ]

Reporting says what result was produced.

Claim licensing asks what that result is actually allowed to mean.

The full paper develops this as the fourth bridge obligation.

And this is one of the places where the deeper framework becomes necessary, because a single phrase like “generalizable” hides several different kinds of travel.

5. Putting an assumption inside a model does not make it better supported

Modern statistical models can be extraordinarily expressive.

They can place measurement error, missingness, latent structure, priors, causal relations, and observation processes inside one joint probability model.

That is computationally powerful.

But one danger comes with the convenience.

Things that appear together mathematically can look as though they have the same empirical standing.

They do not.

A treatment assignment may have been physically randomized.

A measurement model may be supported by validation data.

A prior may be elicited.

A causal graph may be argued from domain knowledge.

A transport assumption may simply be assumed.

All can participate in one model.

Writing them in the same probability expression does not give them the same evidence.

This is the intuition behind a Version 3 principle:

How a premise is represented does not, by itself, change the warrant for that premise.

Or more compactly:

[ \boxed{ \text{representation} \not\Rightarrow \text{warrant}. } ]

This matters far beyond Bayesian modeling.

Reformatting an assumption does not verify it.

Transforming data does not strengthen the evidence for an upstream premise.

Encoding a causal assumption probabilistically does not identify the causal effect.

The full paper calls the broader idea representation-invariance of warrant and shows why evidence status needs to remain visible even when the mathematics becomes unified.

6. Some bridges can be run

There is also a positive lesson.

A forward account does not always have to remain prose.

If the analysis claims:

Under this model, design, or mechanism, evidence like this could arise,

sometimes we can make that account executable.

We can simulate possible evidence.

That lets us ask questions before trusting the final result.

Does the prior imply absurd observations?

Can the intended procedure recover a known quantity from synthetic data?

Does the fitted model reproduce important features of what was observed?

Do conclusions survive plausible changes in uncertain assumptions?

These checks do not prove that the model is true.

They do something more modest and very valuable:

They make it harder for us to misunderstand what our own assumptions imply.

So another useful distinction is:

[ \boxed{ \text{internal reviewability} \neq \text{truth about the world}. } ]

A simulator may faithfully implement a bad scientific account.

But an executable account gives us more opportunities to detect internal failures, coding mismatches, weak identification, and unintended implications.

The full Statistical Bridge treats executability as a strengthening of the forward account, not as a substitute for empirical warrant.

Three errors to remember

The framework can now be compressed into three recurring mistakes.

Error 1: Evidence carries its own interpretation

It does not.

A recorded pattern may be real while the proposed population, mechanism, explanation, or causal interpretation is wrong. In the revenue example, this is how “customers with transactions” can silently replace “all eligible customers.”

Error 2: Mathematics carries its own empirical applicability

It does not.

A theorem can be correct while the operational objects supplied to it fail to satisfy the bridge that would make the theorem relevant. This is how a perfectly computed standard error can be reported for a complete census that needed no sampling uncertainty.

Error 3: Once the bridge is named, the work is finished

It is not.

Knowing that evidence and target must be connected still leaves the hard questions. A declared model can still hide an unsupported observation process, a weak transport assumption, or a claim that travels beyond its evidence.

• What empirical points exist?
• How were they constituted?
• How could evidence arise?
• Where does probability enter?
• What does the inferential result actually certify?
• Which premises are well supported and which remain assumptions?
• What world-facing claim is licensed?

An arrow is not yet a bridge.

The deeper questions now visible

At this point the Primer has done its job if several questions feel newly unavoidable.

For example:

If event evidence and spine targets are different analytical objects, what exactly licenses the crossing between them?

The full paper distinguishes ordinary data transformation from cross-universe construction and shows why shared coordinates are not enough.

If the same realized data can participate in different probability structures, can the same evidence support different kinds of inference certificate?

Yes.

The full paper gives the architecture for saying exactly how.

If probability can enter on either the evidence side or the target side, where does the frequentist-Bayesian disagreement really live?

Not everywhere.

The full paper types the bridge so school differences can be located rather than allowed to define the whole problem.

If a model contains premises with different evidential standing, how do we keep one strong component from lending false authority to a weak one?

The full paper separates mathematical role from evidence status.

If a certificate is valid, how do we decide whether a claim about another population, future time, or intervention regime is licensed?

That is the claim-license problem.

Version 3 makes it explicit rather than leaving it inside the vague word “interpretation.”

These are not complications added by the Statistical Bridge.

They are complications that were already present in the analysis.

The framework makes them visible.

A compact way to look at an analysis

When confronted with a statistical result, try asking five questions:

1. What exactly is the evidence?
2. What exactly is the target?
3. Where, if anywhere, did probability enter?
4. What does the formal result actually certify?
5. What claim is that certificate allowed to support?

If any answer is unclear, the analysis may still be correct.

But some bridge work remains hidden.

And hidden bridge work is where surprisingly many silent analytical failures live.

Where to go next

This Primer intentionally stops where the architecture becomes more interesting.

The full framework is:

Huayin Wang, The Statistical Bridge: From Governed Evidence to Inference Certificates and Licensed Claims, Version 3.0, 2026. DOI 10.5281/zenodo.21979821.

Read it if you want to know:

• why bridge geography and bridge obligations are not the same thing;
• how event and spine fit inside the broader architecture without becoming universal primitives;
• how evidence-side and target-side probability sources differ;
• how one realized evidence object can support different certificates under different declared sources;
• why causal identification, estimation, and transport are separate obligations;
• why premise warrant should survive re-representation unchanged;
• how claim license differs from evidential warrant and informativeness;
• how the five recurrent bridge failures can be diagnosed in practice;
• and what a statistical system would need to store if the reasoning behind a result were meant to be genuinely reviewable.

The deepest question is no longer merely:

What statistical method should I use?

It is:

What makes this evidence capable of supporting this target, through this inferential statement, for this claim about the world?

That is the Statistical Bridge.

And answering that question is where the full paper begins.