Evidence
The data can be right. The mathematics can be right. The claim can still be unlicensed.
Evidence is not a kind of data. It is the governed passage by which operational material becomes evidence, evidence bears on a formal target, formal machinery derives a result, and that result returns as a bounded claim about the world.
The Statistical Bridge The foundational paper. This page compresses and connects it; the deposit governs.
The wound
Everything worked. The claim still does not follow.
A quarter has closed. Every eligible account-quarter point is known. Every qualifying transaction has been captured. Revenue has been established for every point, including the legitimate zeroes. The analyst is asked for average revenue per customer, with a standard error, and produces both.
For the enumerated finite population, the mean is the quantity itself, not an estimate of that population mean — determined by the constituted data:
τ = (1 / |P|) · Σ revenue(p)
The revenues vary a great deal. The numerical calculation over that variation is arithmetically correct. What has not been established is its standing as a standard error for the requested target: enumerating the population settles the sampling question that such a standard error would ordinarily answer. The interval is not wrong the way a miscalculation is wrong; it is unwarranted — a probability statement with no probability source behind it.
What the census removes, and what it does not
A census does not forbid probability. It removes one source of it: sampling uncertainty about the enumerated finite population.
A declared superpopulation model, a measurement-error model, a stochastic process, a forecast target — any of these can still license probabilistic machinery over exactly this table, provided it is declared as the source. The failure shown here is narrower, and worse: inferential machinery invoked without saying what uncertainty it represents.
If neither bad data nor bad mathematics explains the failure, what is missing?
The corpus answers the obvious follow-up — frequentist interval or Bayesian one? — before it is finished being asked:
Neither yet. What is uncertain?
Variation is not a probability source. Where probability is needed it has to enter from somewhere and be named — a design, an observation model, a process for what can vary between now and later, a prior over target-side unknowns. The number on the screen has not moved. What it is being asked to support has. What is missing is the governed passage between them, and the warrant that passage carries.
The structural turn
Two lawful histories can leave the same record.
A Bayesian analysis is usually displayed as a relation between a parameter and the data, p(θ | X). The declaration under which that relation holds is normally understood and not recorded. Suppress it, and something sharper than sloppiness follows: two analyses that differ in what they did can become indistinguishable in what they show.
Figure 1 · record-undecidability
Two lawful histories. One transcript. The difference is not in the record.
History hA · fixed declaration
- M₀
- cond(X₁)
- M₀
- cond(X₂)
- M₀
Every transition is conditionalization under one declared model. Bayes applies on each edge. Lawful.
History hB · declaration revised
- M₀
- cond(X₁)
- M₀
- const
- M₁
- cond(X₁ ∪ X₂)
- M₁
A stopping rule is discovered after X₁. The declaration is replaced, and the retained evidence re-enters under it. Equally lawful.
ρ — a projection into a standard transcript language, suppressing model version and transition type
Projected transcript — ρ(hA) = ρ(hB)
Beta(5, 7) ⟶ Beta(9, 13)
The defect is here. Whether the displayed step was conditionalization under one unchanged declaration — the predicate Vj — is not recoverable from this record. No decision function on the transcript alone can return the right answer for both histories.
Why the collision is not a contrived one
Under the initial declaration the design is fixed-n sampling; under the revision, sampling continued until the eighth success. The two realized likelihoods differ in their combinatorial factors and agree up to proportionality in θ, so the second displayed posterior is Beta(9, 13) either way. Sample spaces, stopping-rule declaration, predictive questions and replication semantics all differ. The displayed numbers do not.
This is what “undecidable” means here, and only this: validity is underdetermined by a projection that erases a load-bearing coordinate. It is a statement about record semantics. It makes no claim about Turing undecidability, halting, or computational complexity.
The repair
Conditioning is not constitution.
The erased coordinate can be put back. It takes two primitive transition types, and the difference between them is not nomenclature — they are different kinds of inferential move, and only one of them is the interior move of Bayes.
- cond(x) Conditioning
-
Evidence enters an unchanged declaration. The model version is the same on both sides of the edge, the new evidence is admissible under the active evidence contract, the carried state is sufficient for the update or the root evidence remains available, and the prior contributes exactly once. Bayes applies on this edge.
- const(ΔM) Constitution
-
A new declaration is established. The edge records what changed, why, which evidence or criticism motivated it, which prior objects are retained and in what role, which root evidence is re-admitted under the new construction, and any transport map used in place of recomputation. Bayes is silent about this edge — and resumes inside the newly declared model once the new episode exists.
Crossing a constitution edge is not a fault. A workflow may cross it freely whenever criticism warrants revision; the type exists so that the crossing is visible in the lineage, not so that it can be discouraged. What the type forbids is a revision that presents itself as a continuation of the same inferential episode.
With the coordinate restored, the predicate becomes decidable by inspection: the displayed transition was conditionalization exactly when the segment behind it consists only of conditioning edges under one unchanged model version. Any constitution edge in that segment settles the question the other way. The remaining premises of the inference keep their own obligations — that is what the rest of this page is about.
Bayes’ theorem needs no repair. Its workflow record needs types.
What a governed passage must expose
Five questions, and the fourth is the one that goes missing.
A typed record is not the whole repair. The transition type says what kind of move was made; it does not say whether the passage as a whole was entitled to its conclusion. For that, a mature analysis has to expose five different things — and answering four of them well is the most common way to be confidently wrong.
-
Bridge constitution
A dependency in the obligation chain.
What empirical objects are being connected, and why is there a statistical connection between them at all?
The evidence object, the target, the anchor geometry, the regime under which values arise, a forward account of how the evidence could have arisen, and the relation that makes the evidence bear on the quantity of interest. An inference method cannot repair an undefined bridge endpoint.
-
Probability source
A dependency in the obligation chain.
Where did probability enter this analysis, and on which side of the bridge?
Evidence-side sources govern how possible evidence could arise — a randomized design, an assignment mechanism, a measurement-error model, a missingness model. Target-side sources place probability over target-side unknowns; a prior is the canonical case. Observing that values vary does not create either one.
-
Inference certificate
A dependency in the obligation chain.
What formal statement carries inferential authority under the declared sources?
A coverage guarantee, a posterior statement, a likelihood ratio, a predictive interval, an e-value. The basis is an open type: frequentist and Bayesian forms are the historically dominant instances, not the definition. A number, interval or posterior does not manufacture the probability structure that gives it meaning.
-
Evidential standing of material premises
Cross-cutting. It qualifies every other exposure, and occupies no position in the chain.
What is the evidential standing of material premises?
A randomized assignment may be instituted and auditable. A measurement model may be corroborated by validation data. A prior may be elicited. A transport assumption may simply be assumed. All of them can appear inside one joint model. Mathematical co-location does not equalize warrant.
-
Claim license
A dependency in the obligation chain.
What may the certificate legitimately mean in the world?
The bound on population, time, regime, transport and sensitivity. A valid certificate has a home — an evidence object, a target, a design, a regime, a population and time domain, a set of premises. Moving beyond that home requires another argument.
Why the corpus says four in one place and five in another
Four of these are the obligation chain: bridge constitution → probability source → inference certificate → claim license. The arrows indicate dependence, not chronology — one declaration can discharge several, and a later diagnostic can reopen an earlier one. What a later claim may not do is borrow authority from an undeclared earlier relation.
Evidential standing is the fifth exposure and not a fifth arrow, because it attaches to components rather than to stages: it is the standing of each premise the passage leans on, and it applies to all four at once. That is precisely why it is the one that disappears when the five are compressed into the chain — and why the next section exists.
Warrant conservation
Writing a premise down more formally does not make it better supported.
Modern models are expressive enough to hold almost everything at once: measurement error, missingness, latent structure, priors, causal relations, observation processes. That is a genuine strength. It also creates a specific illusion — components 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 corroborated by validation data. A prior may be elicited. A transport assumption may simply be assumed. All four can sit inside one joint distribution, and writing them in the same expression does not give them the same evidence.
The corpus names this representation-invariance of warrant: how a premise is written down does not by itself change the evidence for it. Re-expression, transformation and probabilistic encoding are all evidence-neutral with respect to the premise they re-describe. A status strengthens only when an explicit evidence-producing event or rule supplies new warrant — a validation study, a design verification, a calibration check, an external measurement.
This is not an argument against modelling, and it does not say that transformation degrades anything. Evidence-neutral transformation may preserve warrant or weaken it. What it cannot do is manufacture warrant that was not there.
A related distinction survives the same way. A certificate can be entirely warranted and still say very little: a wide interval, a diffuse posterior, a conclusion that only survives under a weak reading. Warrant is not informativeness, and neither one is evidence that the other is present.
Refusal at the inferential boundary
Refusal lives at the boundary.
Inside a fully specified probability model, a conditional question usually has a mathematical answer wherever the relevant conditional distribution is defined. A governed statistical system faces an earlier question:
Is the requested inferential passage established strongly enough to emit the requested claim?
The answer can be refusal.
Grounds a governed system may refuse on
- A required model declaration is missing.
- Evidence provenance is unresolved.
- The requested claim outruns the license.
- A model crossing lacks a transport contract.
Refusal here is an entitlement result, not a crash and not a failure of nerve. It corresponds to underivability in the wider certificate system: the machine may well have produced a state, and the governance layer declines to promote that state into the requested claim because the precondition is not derivable. The absence of a certificate is not repaired by the presence of a value.
Two refusals, one shape, different objects
- Analytical Governance
- asks whether a requested analytical result may be served.
- Evidence · the Statistical Bridge
- asks whether the inferential passage from governed material to a requested claim is established.
Analytical governance can refuse the answer. Evidence can refuse the inference.
The rhyme is a site compression, and it is only safe with the types beside it. These are the same constitutional shape — a governed system that cannot withhold permission is not governed — applied to two different objects under two different jurisdictions. Collapsing them would make the inferential boundary look like a serving policy, which is exactly the confusion the distinction exists to prevent.
What Evidence is
Evidence is a standing acquired through a governed crossing, not a sovereign jurisdiction.
Nothing on this page is a claim that evidence is a kind of object, a layer of the stack, or a third territory beside data and intelligence. Material becomes evidence by bearing on a declared target under a declared account of how it could have arisen — which is a status conferred by a passage, and revocable when the passage changes. That is why the same complete table was evidence for one claim and not for another in the opening case without a single byte of it changing. Evidence is a pillar of this site’s argument. It is not a province of the world.
Where this sits
- Theory of Data supplies the material this page starts from
- ToD governs analytical identity, existence, eligibility, observation, support and lawful derivation — what one point is, why it exists, and which transformations preserve its meaning. It reaches further down than statistics does. It does not establish inferential warrant, and it is not a general theory of inference.
- The Statistical Bridge is the law this page renders
- The Bridge governs the passage from governed material toward inferential claims. It does not replace the Theory of Data beneath it, and it does not legislate a philosophy of probability: it types where probabilistic commitments enter and what they are then entitled to support.
- Certifiable State Under Information Loss formalizes what a state may still certify
- Two states can hold identical numbers and support different claims, because what a state is entitled to assert depends on the contracts that produced it and the evidence behind them. It supplies claim transport, the conservation law that governs warrant, and refusal as underivability. It is not another statistics paper; it is the state-level account the Bridge leans on.
- Analytical Governance composes this crossing into a larger process
- Analytical governance decides whether a requested analytical result may be served, across a multi-world process. It may compose an inferential crossing into that process. It does not adjudicate the inference itself, and its refusal is not this page’s refusal.
- Intelligence begins after this page ends
- A licensed claim does not yet determine what a person or an agent may conclude from it, communicate, delegate, or act upon. That is a different jurisdiction with a different governed object, and it is not built here — named so that its absence is a boundary rather than a gap.
Read the corpus
This page compresses. The deposits govern, and they are open access. A serious reader should start with the foundation and take the rest in whatever order the argument demands.
-
Foundation
The Statistical Bridgethe governed passage, argued in full: constitution, probability source, certificate, evidence status, claim license.
-
Shorter entry
A Primer on the Statistical Bridgethe same distinctions on a running example, without the architecture.
-
Typed history
The Two Jobs of the Conditioning Barrecord-undecidability, and the two edge types that repair it.
-
Warrant and state
Certifiable State Under Information Losswhat a state is entitled to certify, and why transformation cannot manufacture warrant.
-
Probability source
Where Does Probability Live?where probability enters, and why that question precedes the choice of school.
Take the Evidence Walkthrough → Analytical Governance — what may be served The research corpus