The Finding That Changed After Publication

The story

After release 1.0.0 — fifteen rows change the finding

The invented Cedar Methods Group has already published a compact synthetic research package when its source steward discovers that fifteen Cohort D calibration profiles were classified as research profiles. All fifteen met the threshold in the first run, so the correction can change both the denominator and the conclusion.

The research team now has a bounded task: determine whether the correction follows the accepted analysis plan, identify every affected claim and surface, and decide whether a successor package may advance without rewriting release 1.0.0. There is no external deadline to beat; the correction and review states determine what may advance.

The released question was deliberately narrow: within this practice dataset, what proportion of eligible profiles improved by at least ten points between synthetic pre- and post-scores? The package contains a generated roster, an accepted descriptive plan, one execution, Table 2, a sentence in the abstract, a methods-seminar slide, review records, and an immutable manifest for CEDAR-REL-1.0.0.

Accepted plan PLAN-4 includes only records whose account_type is research, requires both scores, calculates change = post_score - pre_score, and marks a row when change is at least ten. It reports a pooled descriptive proportion and evaluates a predeclared two-thirds gate with exact integer arithmetic: three times the number that met the threshold must be at least two times the eligible count. The plan makes no causal, effectiveness, instructional, or real-world claim.

Roster r5 produces one hundred eligible synthetic profiles. Seventy meet the threshold. The displayed rate is 70.0 percent, and the exact comparison passes: 3 × 70 = 210, which is greater than 2 × 100 = 200. The table and abstract therefore carry one bounded sentence: in the synthetic dataset, 70 of 100 eligible profiles met the ten-point threshold and satisfied the descriptive gate.

The release is internally coherent. A reader can name its roster, plan, eligible set, run, table, claim, reviews, and package. The correction does not make that history disappear.

Roster r6 — a correction, not a convenient exclusion

Roster r6 corrects the fifteen profiles' account type and records that it supersedes r5 for current analysis.

The correction does not arrive as an analyst deleting inconvenient rows. The accepted source owner names the affected synthetic identities, basis, time, and revision relationship. Because PLAN-4 already excludes non-research profiles, using the corrected account type appears to implement the accepted plan rather than change it.

“Appears” matters. The source steward is authorized to correct the source but not to declare the method unchanged. The analyst can prepare both executions and expose their difference but cannot strengthen the interpretation or publish a successor. A methods reviewer must decide whether PLAN-4 truly governs the correction or whether the team has changed its method after seeing the result.

Before recalculating, the model identifies the full stale set: every dependent item that can no longer be treated as current. Roster r5 supports eligible set ELIGIBLE-SET-18, execution RUN-18, the Cohort D row, the total row, CLAIM-2 r2, the abstract sentence, the seminar slide, and the pointer that identifies release 1.0.0 as current. Once the source steward issues r6 as the authoritative current source revision, each dependent object becomes stale for current use. None is erased. The original package remains the evidence of what the group knew, calculated, reviewed, and released at that time.

The candidate rerun — the declared plan produces a different conclusion

The candidate corrected execution applies PLAN-4 to roster r6 and produces ELIGIBLE-SET-19. Eighty-five profiles remain eligible. Fifty-five meet the threshold. The exact fraction is 11 / 17, displayed as 64.7 percent. The change from the first release is approximately negative 5.3 percentage points. The calculation is deterministic; whether applying PLAN-4 counts as an unchanged method still belongs to methods review.

More importantly, the exact gate now fails: 3 × 55 = 165, while 2 × 85 = 170. Rounded display values do not decide the result. The candidate claim must say that the corrected synthetic dataset did not meet the predeclared gate and must explain why the denominator changed.

The new result does not prove that the earlier authors acted improperly, that an intervention failed, or that a scientific claim was replicated or refuted. This is a synthetic descriptive fixture. Its job is to make the transition answerable, not to borrow the authority of science for a software story.

An AI assistant can help identify affected dependencies, compare package manifests, propose failure tests, and draft corrected language. Every proposal retains its source and status. The assistant cannot accept the correction, decide that the method remained unchanged, settle a reviewer objection, determine authorship, or authorize release. Fluent prose does not collapse those responsibilities.

Release 1.0.1 — history remains visible

In the passing path, the methods reviewer records a substantive disposition: accept the source correction; PLAN-4 remains unchanged because the exclusion was predeclared; the corrected claim is required.

A domain reviewer then challenges the interpretation and limitations. The domain reviewer records that the final package resolves the challenge by retaining only the narrow synthetic descriptive claim and the existing limits: no causal, effectiveness, instructional, or real-world conclusion. The research lead then accepts the exact package within the scenario's recorded delegation.

That acceptance is still not publication. The repository or publisher performs a separate release act and issues immutable successor CEDAR-REL-1.0.1. Its manifest binds roster r6, the accepted plan, corrected execution, Table 2, revised abstract sentence, review dispositions, and its supersession relationship to 1.0.0.

Release 1.0.0 remains reconstructable from its exact retained package and visibly corrected or superseded. It cannot continue to masquerade as the current package through an unbound slide or export. Release 1.0.1 cannot silently replace the historical evidence. A future reader can reconstruct both states and understand precisely why they differ.

That is the promise of an auditable executable paper in this Scenario: not a magical document that decides what science means, but a connected research object in which one consequential sentence remains attached to its source, method, arithmetic, artifacts, review, and release.

Where the model stops

The model can represent sources, executions, claims, objections, dispositions, and releases. It cannot determine scientific truth, method validity, interpretation, authorship, peer-review status, or publication acceptance. Each institutional act requires its own accountable evidence.

Verification note

Grid FYI fixture revision 2026-08-27.2 freezes the synthetic roster logic, independently reviewed expected results, exact arithmetic, affected dependencies, roles, failure cases, predecessor immutability, release identities, and prohibited authority transitions. Eighteen deterministic tests passed, and a distinct verifier reconciled the retained execution with those independently reviewed expected results and the public evidence boundary. That advances this repository-owned publication fixture to R3. It does not validate Grid-runtime behavior, an institutional research workflow, or any scientific result.

Institutional use would require more: governed data and permissions, participant protections where applicable, identity and review systems, repository and publisher integration, rights handling, and explicitly selected and validated provenance, research-object, citation, contributor, or journal-interchange profiles. Naming a standard is not conformance.

The Scenario succeeds when a reader can change one declared source, predict the calculation, find every stale artifact, identify who may decide what, and reconstruct both releases. Software can preserve the chain. It cannot decide what the finding means.

What remains to prove

A controlled intended-environment evaluation still needs qualified reviewers, governed data and permissions, selected standards profiles, institutional source and repository integrations, failure cases for those integrations, confirmation that every derivative binds the same accepted revision, and retained acceptance evidence. No real research claim or institutional workflow is qualified by this verified publication fixture.