The Evidence Walkthrough

A stakeholder asks:

“What was average revenue per active customer in Q2 — and can we count on it?”

The first afternoon taught the first half: avg_revenue_per_active_customer @ {Q2} is a governed number — declared population, declared law, lawful derivation. Suppose all of that is done. The number is $312.40, and it is entitled to be served.

“Can we count on it?” is a different kind of question, and it is secretly three questions wearing one sentence:

1. Is $312.40 the true Q2 average for our active customers?
2. Will next quarter look like this?
3. Is revenue this high because these customers are active?

Same number. Three claims. Three different verdicts are coming — and the work of this page is showing you why the difference was always there, hidden inside “count on it.”

A claim that wants service states its scope. The declaration begins:

claim A: mean(revenue) over active customers, Q2, this company    # about the enrolled population itself
claim B: expected revenue per active customer, Q3                 # about a period not yet observed
claim C: activity increases revenue                               # about what causes what

Nothing has been decided yet. Writing the claims apart prepares the evidence problem — and it has already exposed something upstream of evidence: an intent ambiguity. The stakeholder’s one sentence turned out to contain a fact question, a forecast question, and a causal question, and no amount of computation settles which one was meant. Before evidence adjudication begins, the system returns its first outcome:

The stakeholder picks claim A, then B. The declaration continues with the evidence side:

evidence: revenue @ {customer, Q2}, all customers active in Q2
frame:    the declared active-customer roster, Q2                  # who could be observed
coverage: complete — every point in the frame is observed          # a census, not a sample

Read the coverage line twice, because it decides everything in the next step. The Q2 active-customer population is fully enumerated: every customer who could be in the denominator is in the data. There is no sampling, no missingness, no gap between the frame and the observations.

Why write these lines at all, rather than trust the number to speak for itself? Because the number cannot speak for itself. Two lawful histories can end at the same displayed number — a complete enumeration and a well-designed sample; a fixed analysis plan and one revised along the way — and nothing in the output distinguishes them. Neither history is at fault; each can be perfectly governed on its own terms. The failure would belong to a record that erases the distinction between them. That is why the account is established before service and travels with the claim: the numeric output cannot reconstruct its own provenance; the distinguishing history has to survive in the governed record.

Statistical habit says: compute a confidence interval, attach a p-value, report $312.40 ± something. Habit is not law. The law asks first: what would the probability be about?

A census does not forbid probability. It removes one particular source of it: sampling uncertainty about the enumerated population. Claim A’s target is exactly that population — the Q2 roster, fully observed — so the removed source is the only one this claim puts in play. Other uncertainty could still enter, lawfully: a measurement-error model, a stochastic account of revenue itself, a forecast target. But each requires its own declared source, and claim A declares none. $312.40 is the finite-population Q2 average, and under this declaration there is no gap for inferential probability to be about.

Two verdicts, for two different requests.

• The descriptive claim:

  no inferential passage is needed,

  and the system Serves the arithmetic fact.

• The request to attach an interval anyway:

  an inferential passage is requested — and at the evidence boundary, the machinery names no uncertainty source, so no passage is established.

  The system therefore Refuses the request.

This pair is the walkthrough’s flagship, and it is the distinction conventional analytical workflows can leave implicit: the system says yes to the number and no to the inference. Two layers are visible in that pair, and keeping them apart is the discipline. The descriptive claim needs no inferential passage. The interval request has no established one. The serving verdict — Refuse — is the governed system acting on that second finding. Evidence adjudicates the passage; the system serves the outcome. And the failure the refusal names is not “a standard error on a census,” as though the taboo were syntactic. It is inferential machinery invoked without saying what uncertainty it represents. An interval on claim A would not be caution; it would be a claim about a gap this declaration does not contain — decoration wearing the costume of rigor. A governed system distinguishes the two requests, and distinguishing them is not a limitation. It is the system knowing what its numbers are.

(The canonical deposited instance of this case — a census total and its unlicensed interval — is worked on the Evidence page and in the papers. Different number, same law.)

Claim B — next quarter. Now there is a gap: Q3 is unobserved. Probability must enter, and it enters only through declared structure — a probability source saying what the probability is over, a premise with its standing, a certificate established on them, and a license bounding what the certificate supports. This is the guard’s other half in motion: the uncertainty that claim A could not invoke without a source, claim B now declares one for. The declaration grows, and the objects stay distinct:

claim B target:       expected revenue per active customer, Q3

claim B source:       the predictive distribution of Q3 active-customer
                      revenue, under the declared continuation account

claim B premise:      under the stated stability conditions, the mean
                      of that Q3 distribution equals the observed Q2
                      finite-population mean                        # evidential basis: argued

claim B certificate:  conditional on that source and premise,
                      expected Q3 revenue per active customer
                      is established as $312.40

claim B license:      Q3 point expectation only, conditional on
                      the declared source and premise

Claim C — because. The current account establishes Q2 revenue for the active-customer population — and nothing more. It declares no exposure contrast, no comparison, no identification account that could establish the causal effect of activity on revenue; within this record, activity is who the population is, not a variable that varies across it. No computation on this record can manufacture the missing passage.