matchednull 0.2.2

• matched_null_test() now stops with an informative error when cluster_fn returns an invalid value (missing, non-finite, non-numeric, or not of length one), on the real data or on any twin. The error names the call (“the real data” or “null twin r of R”). Previously an invalid twin was silently dropped from the numerator of p_exceed while the denominator kept the requested R, and the printed summary still reported all R twins, so a run in which most twins had failed could print a p-value and an outcome as if nothing had gone wrong. Errors raised inside cluster_fn propagate unchanged, as before. Numerical results for valid inputs are unchanged: seeded runs reproduce exactly.
• The printed exceedance line is relabelled. It read “p (real >= nulls)”, but the numerator counts twins whose statistic reaches the real value, so the label named the wrong direction. It now reads “MC exceedance (null >= real)”, and the other printed labels are re-padded to keep the block aligned. The value is unchanged: (1 + sum(null >= real)) / (R + 1), as before.
• The printed outcome line is renamed. The three outcomes now read “no detected excess (within the reference interval)”, “departure from the reference model (above the interval; kind not settled by this test)”, and “below the reference range (a diagnostic, not a verdict)”. They replace “null-like (within the twins’ interval)”, “exceeds the null (beyond margins
  • covariance)“, and”below the null interval”. Returned values are unchanged.
• Documentation corrected. The twins preserve the margins exactly (each column is a permutation of the observed values) but reproduce the correlations only approximately: the observed correlation matrix is used as the latent correlation of the copula, and non-normal margins change the output Pearson correlations, so the earlier “to within sampling error” wording was wrong. p_exceed is now described as a plug-in Monte Carlo exceedance quantity under a reference fitted to the same data, not an exact finite-sample p-value, and copula = "t" as a sensitivity variant with heavier-tailed dependence that does not by itself establish or rule out clusters.
• citation("matchednull") now gives the title of the manuscript under review, A Matched-Null Test of Whether the Clusters a Pipeline Reports Are Real, With an Application to Personality Types. It previously gave the title of the preregistration, Types Without Taxa.
• The calibration summary in the README now reports runs in which the search range was wide enough that the tested count never reached its upper limit: on skewed clusterless data the test flags 4 of 100 datasets at n = 500 and at n = 2,000 (nominal 2.5%), while BIC and ICL choose more than one cluster in all of them. The earlier summary, which said the test flagged none, came from a capped range, under which the twins’ 97.5th percentile sat at the cap and could not be exceeded.

matchednull 0.2.1

• matched_null_test() gains parallel. With parallel = TRUE the twin loop runs through ‘future.apply’, so the backend is whatever future::plan() the caller has set. Seeded results are reproducible and independent of the number of workers. The default parallel = FALSE path is unchanged, so existing seeded analyses reproduce exactly.
• The maintainer address now uses a permanent personal inbox.

matchednull 0.2.0

• copula_null() and matched_null_test() gain copula = "t" and df: t-copula twins preserve the same margins and correlations but add tail dependence. A verdict of “exceeds the Gaussian null” can now be stress-tested against heavier-tailed dependence (df = 8, then 3) before it is read as evidence of types.
• matched_null_test() results record the null family used ($copula, $df), and the printed summary names it.

matchednull 0.1.0

• First release: copula_null() builds Gaussian-copula matched-null twins; matched_null_test() tests any clustering pipeline against them.