Identity before interpretation
The fixture contains twelve observation months, two features and two revisions of each feature-period pair. All were constructed in the current edition. Each row retains its observation month, feature name, consecutive revision number, value, fictional release time and fictional ingestion time. Release must occur after the observation month ends; ingestion cannot precede release.
Eligibility at an information cutoff
For each feature-period pair, choose the highest revision whose ingestion timestamp is at or before the selected cutoff. Equality is eligible. Release and ingestion use explicit UTC timestamps. A later revision is never pulled into an earlier reconstruction. Missing observations are not replaced with zero or carried from a different period.
An explicit null is still a revision
Select the record before deciding whether its value is usable. An eligible null marks a withdrawn input. It leaves the output Unavailable even if an older revision had a number. This avoids silently undoing a withdrawal while preserving the audit trail that explains it.
The threshold rule
Subtract the chosen growth threshold from the growth score and the inflation threshold from the inflation score. If either absolute difference is at most the boundary half-width, return Boundary. Otherwise label the two signs G+ or G-, and I+ or I-. Missing or null inputs take precedence and return Unavailable. Default thresholds are zero; the inclusive half-width is 0.1 points. Comparisons use a 1e-12 tolerance for floating-point equality.
Two different comparisons
The archive contrasts original cutoff eligibility with the latest fixture revisions under the same rule. The sensitivity tool instead holds each cutoff information set fixed and changes thresholds or the boundary band. It counts every period, including Boundary and Unavailable, in a twelve-record denominator. Neither disagreement measure is a forecast accuracy, confidence level or return.
Scope of verification
A separate Python Decimal implementation checks feature selection and labels against the browser engine across release boundaries and parameter combinations. Unit tests reject invalid timelines and verify that future-vintage edits leave earlier reconstructions unchanged. These are software-correctness checks on fictional inputs, not independent market validation or evidence of a profitable model.