# Counterevidence, Boundary Cases, and Unsupported Claims

The strongest conclusion is conditional: revisions can change decision-relevant conclusions, not that public data are generally unstable. This file preserves evidence that narrows or challenges broader readings.

## Most scoped conclusions did not change

- BLS employment and total-wage adjacent-growth direction was stable in all 11 comparable transitions for each measure. Their median level revisions were only 0.154% and 0.083%.
- Census growth direction was stable in 199 of 204 state-year comparisons; the 1% threshold was stable in 196; top-10 membership was stable in 198.
- BEA growth direction was stable in 19 of 20 quarters. The observed one-quarter sign flip demonstrates possibility, not a typical rate for all GDP history.

This counters the popular claim that “revisions make official numbers useless.” A later vintage often refines a level without changing the conclusion.

## Threshold results depend on the rule itself

The BEA rule is strictly `growth > 2.0%`. Two threshold changes begin exactly at 2.0% and move above it. A rule using `>= 2.0%`, a rounding tolerance, or a broader “near 2%” category would produce a different count. The four-of-20 result therefore describes the preregistered rule, not an intrinsic property of GDP.

Similarly, “top 10” and “above 1%” are decision conventions. The ledger keeps denominators and rules visible so they cannot be mistaken for natural breakpoints.

## Level magnitude and narrative fragility are not the same

The largest BLS establishment revisions also have meaningful level magnitude, but the general mechanism is denominator-dependent: a small revision can dominate a tiny adjacent change, and a larger revision can leave direction unchanged. Revision-to-change ratios use a denominator floor to avoid infinite or explosive claims near zero. They are sensitivity diagnostics, not stable physical constants.

## The Census frame does not isolate revision causes

Census explains that annual vintages incorporate newer administrative data and methodology changes, with migration particularly assumption-sensitive. This run observes combined vintage differences. It cannot attribute New York's or DC's revision to one component, nor infer that the same state will be revised in the same direction next year. Geography aligns for the sampled state/DC rows, but the source schemas widened from 75 to 97 columns.

## The population scenario fixture is intentionally simple

Constant annual growth plus/minus 0.5 percentage point is transparent, not demographically complete. It omits age structure, births, deaths, migration components, policy, and shocks. The V2024-versus-V2025 starting-level difference in the base 2030 result is about 114 thousand, while the V2025 high–low assumption spread is about 21.4 million. Vintage uncertainty is therefore small relative to the selected scenario spread in this fixture. The layers still should not be merged, but this run does not show that they are equally material.

## EIA case spread is not forecast calibration

The AEO cases demonstrate assumption sensitivity; they do not establish predictive coverage. Near 2025 the three selected values differ by only 0.002 quads, and divergence grows over time. The 2050 extremes are neither confidence limits nor proof that actual energy use will lie between them. A retrospective error study would be needed to evaluate forecast performance.

The bulk archive also contradicts its own declared start field: metadata says 2024, data begin in 2025. This reduces confidence in metadata-driven period generation but does not invalidate the 2025–2050 values that are actually present.

## The missingness result is a controlled fixture

Zero-fill falls below the documented lower bound in S019, proving a failure mode. It does not estimate how often this happens in BLS, Census, BEA, or EIA data. The empirical canonical panels selected here contain no missing numeric canonical values because unresolved or absent periods were excluded by the frozen inclusion rules and recorded as quality/coverage constraints.

## BLS coverage is deliberately narrow

The official linked revision CSV returned HTTP 403 to this runtime. No user-agent spoofing or bypass was attempted. The labor result uses 36 source-located national table rows (three measures × 12 quarters), not the full state-level or historical file. The publisher's statement that large revisions are rare and the observed establishment outliers can coexist: “rare” is not a bound, and this sample cannot estimate the long-run frequency.

## Hypotheses not fully tested

- H03, extra fragility at turning points, remains inconclusive because a turning-point definition was not frozen precisely enough to support a non-post-hoc comparison.
- H07, false conflicts from seasonally adjusted versus unadjusted joins, was not tested because the frozen matched frames do not expose both variants at the same target grain. The compatibility rule is retained as a documented guardrail, not reported as an observed failure.
- H06, transform amplification, is supported for selected BLS change denominators but not established as universal across all transforms and datasets.

## Popular claims not supported

This study does **not** support:

- “the newest number is always the right number”;
- “two official values can be averaged to resolve a conflict”;
- “a baseline scenario is the most likely forecast”;
- “scenario min/max is a confidence interval”;
- “missing values are safely equivalent to zero”;
- “every first release is unreliable”;
- causal attribution of a revision from its direction or magnitude;
- prevalence estimates for revision risk across all public datasets.

These boundaries are reflected in the claim ledger, methodology, article packets, and product tickets.