rasch 1.12.1

• CRAN checks use shorter workflow fixtures and recorded vignette bootstrap calculations. The full tests and scripts to regenerate those calculations remain in the repository; estimation and replication defaults are unchanged. Export examples use a smaller demonstration dataset and screen-resolution plots to keep their running times below CRAN’s per-example limit.

• Automatic DIF resolution withholds its verdict when the final assessment could not estimate a single item-term test, instead of reporting that no DIF remains. Where only some tests were estimable the stop reason says how many were not, and the counts of remaining and non-uniform DIF are reported as unavailable rather than as zero. Saved resolutions from earlier versions are dropped when a project is reopened.

• Planned DIF contrasts withhold only the affected item when a split refit is refused, and say why, instead of abandoning the whole call. The refusal reason is reported in place of an unrelated response-category note.

• A linear trend across an ordered factor is no longer withheld because the middle level carries the rounding residue that contr.poly() leaves in place of an exact zero. Contrast weights below a relative tolerance are zeroed, so a level with no weight is no longer treated as a required cell. Saved planned contrasts from earlier versions are dropped on reopening.

• The notes recording which requested DIF tests were not estimable reach every surface that reports the analysis: the console, the HTML report, the Word and PDF report, and the saved summary, not only the app.

• Item and person tables in the app colour and embolden cells by their own value again. A conditional style was passed to the table as a function where an expression was expected, so every flag was decided from the row index: whole columns read as misfitting or significant, and residual and DIF magnitude flags never appeared.

• Switching an analysis to Comparative Judgement clears the Rasch results that do not carry over, so a CJ analysis can be saved after planned contrasts, automatic DIF resolution or a tailored analysis was run on the same data.

• Uploaded equating references, Comparative Judgement banks and item panel maps are checked when they are supplied rather than when the analysis is saved. An unusable upload is refused at the point of upload and leaves the rest of the analysis saveable.

• The app’s dependence and spread panels print the note explaining why a statistic was withheld, as the console does, rather than an unexplained NA.

• sim_recovery() accepts a fit made without an identifier column, verifying the response rows without treating row numbers as person identifiers, and accepts a fit whose estimator dropped an item with no variation. Both cases previously failed with a claim that the fit did not belong to the simulation. Person and group-unit recovery requires verified identifiers; legacy person truth can instead be paired by unique response patterns. When identifiers are absent, group allocation is explicitly unchecked.

• Test characteristic, test information and targeting displays enumerate the administration designs actually observed, including distinct missing-item patterns. A design that no person was administered is never formed. Large design collections no longer overflow the legend or recycle colours; person-item maps retain their person and threshold distributions.

• Fixed-subset dimensionality comparisons now require bootstrap calibration for a verdict, as residual-derived splits do. Unequal targeting can distort the binomial reference even when the subsets were chosen in advance. Observed proportions and intervals remain available descriptively.

• DIF resolution counts unavailable uniform and non-uniform hypotheses separately and distinguishes missing interaction cells from failed tests. Refused comparisons retain their explanation in the console and app.

• Opening a saved analysis clears the previous dimensionality result. Old dimensionality, DIF-resolution and frame-invariance results are omitted when their calculation has been superseded; source files are unchanged.

• test_information(items = ) labels each curve with the items that produced it, and no longer returns a curve twice.

• Extended frame score curves use the same design enumeration and labels as test_information() and the curve plots, so a fit describes its designs one way. A design label always names its item set, and a set or item name containing the separator is quoted. Saved extended frame analyses holding the earlier curves must be refitted before reopening; their source files remain unchanged.

• The EFRM unit-test family counts the hypotheses that are free. With two groups or two sets the second reported coordinate restates the first, and now leaves its adjusted probability and flag to the row it restates instead of being counted twice.

• frame_invariance(se_method = "bootstrap") refers its Wald statistics to a t distribution on the bootstrap degrees of freedom rather than to the normal. A 30-draw bootstrap rejected an invariant item about 6.6 per cent of the time at the nominal 5 per cent.

• The two-panel Comparative Judgement frame family counts the single panel-unit hypothesis once, so a panel-unit difference is reported with the evidence it has and the step-down multiplier for the scale rows is the number of distinct questions.

• btl_equate() records that its precision-weighted shift standard error treats the estimated weights as fixed, which understates the shift’s uncertainty in small panels.

• btl_dif() no longer reports the judge-clustered chi-square boilerplate as a finding about a particular object.

• Explanatory coefficient tests refer to a t distribution on the residual degrees of freedom for a fit without repeated persons. The normal reference rejected a true null about 8.6 per cent of the time at the nominal 5 per cent with the smallest supported person count.

• The item-fit bootstrap keeps the observed class-interval count in every replicate under the "resample", "normal" and "fixed" generators, so the null and the observed statistic use the same number of intervals. Probabilities no longer depend on which side of a 50-person boundary the sample happens to sit.

• dimensionality_test() reports the difference between the two subsets’ mean estimates as a description and no longer as a paired t-test. The test rejected on every unidimensional replicate whenever the subsets differed in difficulty, which is targeting rather than dimensionality. Saved person-subset tests are dropped when a project is reopened.

• Printing a dependence result saved by an earlier version reports its statistic again instead of failing, and a withheld item-trait probability prints as unavailable rather than as an empty value, in the console and in the saved summary alike.

• Notes containing parentheses or brackets appear as prose in the Word, PDF and HTML reports. They were escaped in a way the report renderer read as mathematics.

• The simulation-validation studies run from the tree as committed: the conditional DIF bootstrap confirmation pins the committed helper, the item-fit interval checker verifies against the current sources, and the README describes the recorded dimensionality guard results and how to reproduce them.

• DIF magnitudes leave the ETS category unavailable when a required test probability is unavailable or its standard error is zero, rather than treating the missing test as evidence for category A or B.

• EFRM conditional calibration and CJ frame panel-ratio fits check the exact likelihood curvature as well as the score. Stationary saddles can no longer supply converged fits, unit inference or bootstrap covariance. Previously saved frame analyses without a current likelihood-check record must be refitted before reopening; their source files remain unchanged. Resolving every common item in a set is also refused when it leaves the groups’ relative origins unidentified, even if another set links their units.

• CJ frame linking checks partly separated outcomes, including comparisons with zero loading along a separating direction. Such links have no finite estimate. An exhausted iteration limit can no longer certify convergence.

• The installed Shiny app resolves the internal helpers needed for item tables, saved CJ analyses and simulation recovery, including standalone launches.

• CJ dimensionality compares observed and fitted expected points within each object pair, retaining category thresholds and fitted position or history effects. Simulated histories use their own fitted expectations. Older saved dimensionality results require recalculation; fitted models are unchanged.

• EFRM bootstrap replicates are discarded when any supported set link fails numerically or does not converge, including redundant links. Failed draws cannot enter the covariance or count as usable replicates. Full-bootstrap refits with unidentified group units are also discarded.

• CSV, HTML, Word and PDF exports include supplied DIF results based on external person factors. Bulk exports require a new or empty folder, preventing files from separate analyses being mixed.

• CJ pair-surprise diagnostics do not assign a stronger or weaker direction to tied object locations. Relabelling objects no longer changes these flags.

• Explicit item selectors can distinguish item columns from separately supplied person factors with the same names in Rasch and EFRM fits.

• CJ model comparisons recognise unchanged within-judge sequences after relabelling their order values.

• EFRM item sets can link through intermediate sets without direct overlap between every pair. Insufficient overlap is distinguished from a numerical link failure; the retained linking graph must still connect all sets.

• Structural Rasch and EFRM refits preserve external factors whose names match item names, including frame-invariance bootstrap refits.

• Downloaded simulation scripts load rasch before recreating the data.

• EFRM DIF bootstraps preserve frame and person-factor names, including an ordinary factor named group.

• Calculations that restore random-number streams refuse the Box-Muller normal generator before changing the stream. R does not expose its cached normal value for restoration. The default Inversion generator is unaffected; ?rasch_rng describes the supported behaviour.

• BTL-EFRM documents its different treatment of response-boundary objects and distinguishes separated outcomes from disconnected comparison designs.

• Saved CJ DIF results without verified judge-role alignment are omitted on reopening. Reports refuse older automatic DIF magnitudes that may have used stale factor values; the analysis must be rerun.

• Bootstrap workers retain the coordinator’s random-number generator settings, including non-default generators, so a fixed seed gives the same draws in serial and parallel execution. Installations with source references retained also select the coordinator’s package library before loading the worker namespace.

• BTL simulation recovery checks the original comparisons before extreme objects were set aside, then assesses the calibrated objects only.

• Principal-component calibration uses the corrected kurtosis polynomial. The full four-component model matches unrestricted PCM through four thresholds; longer scales retain a restricted polynomial structure. Saved fits that used the earlier kurtosis polynomial require a refit.

• Automatic DIF magnitudes retain the factor values supplied to that analysis, including replacements for factors stored in the fit.

• BTL-EFRM panel-unit reconciliation retains covariance between object sets assessed by the same judges. Point-estimation weights are unchanged.

• Saved fits record their person-scoring algorithm. Older fits are checked before reopening; materially changed scores require a refit rather than silently restoring superseded estimates and diagnostics.

• CJ DIF in the app uses the judge role from the fitted analysis, not a subsequently changed column selector. Unavailable overall fit probabilities are labelled explicitly in summary tables.

• Incomplete repeated-measures DIF fits occasion and person factors jointly, with equal total weight per person and person-cluster CR3 covariance. Marginal occasion centering no longer introduces group effects when occasion coverage and person-factor composition differ. Complete-panel analyses retain their existing covariance references.

• WLE scoring compares competing maxima for separated item banks instead of accepting the first score-equation root. This applies to ordinary and extended-frame scoring, including externally weighted measures. Saved weighted tables from earlier solvers are authenticated and recomputed.

• BTL-EFRM refuses a linking unit at the zero boundary rather than replacing it with one. The same check applies inside bootstrap refits; genuinely indistinguishable within-set locations retain their placement convention.

• CJ model comparisons check each fit against its reference separately. Adding another fit cannot erase a known difference in comparison order.

• Judge-clustered CJ dimensionality is descriptive by default. independent_comparisons = TRUE requests the conditional sensitivity reference; it does not account for general within-judge dependence. Unsupported saved references and superseded mixed-panel DIF results are omitted on reopening, without removing the source data or fitted models.

• Explanatory diagnostics use stable added directions and retain failed candidates in the Holm family, allowing other departures to be reported.

• Loading another dataset in the app clears the active analysis and its derived results. Explicitly kept model snapshots remain available. HTML reports label an unavailable total-fit probability explicitly.

• CJ dimensionality withholds probabilities and reference bands wherever its design checks withhold inference, including incomplete pair coverage in BTL-EFRM. The observed residual decomposition remains available. Saved analyses omit superseded unsupported references after integrity checks, retaining the data, fitted models and analysis history.

• Explanatory Rasch and CJ fits accept distinct predictors whose generated coefficient labels coincide. Labels receive unique suffixes without changing the design, estimates or model comparisons. Fixed departures also avoid existing coefficient names.

• R code for a completed CJ DIF analysis retains the judge-factor maps used in that run, including metadata corrected after calibration. Reopened analyses retain those assignments for subsequent DIF runs; new metadata can still replace them.

• Structural refits and bootstrap procedures refuse older anchored or component-constrained fits when their original refit settings are missing. They no longer release fitted restrictions silently. Simulation recovery also retains an anchored fit’s origin when those settings are unavailable.

• lr_test() checks restrictions recorded by the fitted estimator as well as its refit settings. Older anchored or principal-component fits no longer bypass the requirement for an unrestricted PCM comparison.

• App reports and result archives stop when a selected DIF analysis or custom dimensionality test is unavailable, rather than substituting a default analysis. Analysis files can still save settings without a computed result.

• sim_recovery() refuses comparisons after fitted categories have been removed or merged. Paired-comparison frame recovery aligns generating origins to the fitted reference set and refuses a changed reference whose generating unit is not one. The model help identifies the alphabetical reference-set rule explicitly.

• EFRM and externally weighted person scoring retain accuracy under changes of measurement unit. The score equations and SE formulas are unchanged. Saved weighted tables from earlier solvers are authenticated and recomputed when an analysis is loaded.

• Keyed analyses with repeated person IDs now export the available results and an explanation that distractor analysis is unavailable, instead of stopping the export or leaving an empty report section.

• BTL-EFRM unit tables in rendered reports show very small adjusted probabilities as < 0.001, not zero.

• Dichotomous CJ information remains symmetric when a pair is reversed, including in the far tails. Targeting plots refuse a grid whose reference information is numerically unavailable. The help now distinguishes the zero-gap information peak for dichotomous comparisons without a position effect from the more general ordered-comparison case.

• MFRM person factors supplied as a data frame retain each person’s observed value when other rows are missing. Results no longer depend on which response row comes first; conflicting observed values are still refused.

• Common-unit person scoring scales its root-search tolerance with the unit and calculates SEs without squaring the discrimination. Large changes of unit no longer produce inaccurate WLEs, MLEs or extrapolated-score SEs.

• Missing-data codes apply to raw responses, not scores generated by a key or earlier category renumbering. Keyed scoring and structural refits no longer delete valid scores that coincide with an original missing code. Affected analyses should be rerun from the raw data.

• combine_items() validates its requested model for every fit. Explanatory fits refuse an RSM request rather than silently returning a different model; their default refit still retains predictors and frees new superitems.

• Tailored analysis retains keyed scoring and DIF-split records in its returned fits. The tailored option data exclude censored responses, keeping subsequent distractor analysis and refits on the retained data.

• The RSM fit returned by lr_test() retains keyed scoring, DIF splits and superitem definitions for subsequent analyses.

• Alpha if deleted checks each reduced scale’s covariance matrix separately. Missing covariances involving a deleted item no longer suppress an otherwise estimable reduced-scale alpha; unsupported reductions remain unavailable.

• Saved analyses authenticate and recompute weighted person tables from the earlier solver, preserving the data, fits and requested weights. Current tables retain strict integrity and reproduction checks. Available-case CTT summaries with missing responses withhold SEM rather than combine pairwise alpha with a complete-case score variance.

• Relaxing explanatory restrictions preserves earlier DIF-split and superitem records for subsequent anchor checks and dependence diagnostics. Weighted person estimation rescales weights within each response pattern, preventing missing estimates when only very small-weight items were answered.

• Refits preserve automatic and explicitly requested class-interval rules. Item-fit bootstraps retain per-item allocation with missing responses; earlier Rasch bootstrap results must be recomputed before reuse. Category plots, plot exports and downstream refits accept fits whose tied locations produce one interval. Fixed tailored fits are labelled as person scoring against a fixed item calibration. The missing-response validation now includes 200 datasets and 99,900 usable bootstrap refits under the corrected rule.

• The final tailored scoring fit records its fixed thresholds explicitly. Downstream item changes and refit-based bootstraps refuse fully anchored scoring fits, including older saved fits, instead of releasing their anchors.

• Explanatory Rasch and CJ formulas now refuse unsupported offset() terms instead of silently omitting their fixed contributions.

• Model comparisons withhold generic EFRM likelihood differences, which need not use the same item pairs. The dedicated group-unit comparison is unchanged. CJ comparisons now require the same declared response scale, including endpoint categories absent from the observed responses.

• Explanatory refits align keyed responses with retained persons. Simulated scree and DIF refits no longer inherit observed answer options; blank response rows no longer make every keyed-response scree replicate fail. Predictor matching and item relaxation preserve exact item names, including leading or trailing spaces, and refuse ambiguous whitespace-normalised matches.

• Unused ordinal columns in predictor tables no longer prevent explanatory Rasch or CJ models from fitting. Ordinal contrasts are applied to the predictors included in the formula.

• Explanatory Rasch and CJ models use rescaled design columns for estimation and numerical checks. Changing predictor units no longer causes premature convergence or prevents Newton updates. Coefficients, covariance and model comparisons retain the supplied predictor units and equivalent restrictions. Stable column centring also prevents large constant offsets from making a varying predictor appear unidentifiable.

• Frame-invariance tests count distinct persons contributing informative item pairs within each set and frame. Adding all-zero or all-maximum responses no longer enables inference in an otherwise undersupported frame. Support uses the separate calibration’s retained items and category structure, including when a frame lacks an endpoint category.

• Equating retains the shift fixed by an exact common anchor when other common items or objects have unavailable SEs. CJ equated-location SEs now include shift uncertainty and its covariance with the reference locations; unavailable joint uncertainty is reported rather than ignored.

• Dimensionality magnitudes compare reliability on matched complete-response rows, retaining both calibrations. Missing subscale scores no longer change the PSI comparison sample. Tables report rows used and excluded; saved app results from the earlier calculation must be recomputed.

• Named-column arguments in reshaping, CJ, and Multiple Ratings require character strings. Numeric selectors can no longer pass a name check and then read a different column by position. Names such as "1" remain valid.

• Repeated-measures DIF reports unavailable within-person tests as NA when missing factorial cells confound their adjusted mean with the between-person design. These tests remain in the multiplicity family. Missing within-person combinations no longer abort the analysis, and item-specific missing levels no longer change the test’s requested design.

• Group-restricted person-item maps now show only the EFRM or MFRM response patterns observed in the selected subgroup. Information curves from other subgroups no longer accompany that group’s person distribution.

• EFRM and MFRM information and test-characteristic curves now retain each observed item pattern. Partly answered sets or facet conditions no longer contribute unobserved items, overstate information, or understate SEM.

• Crossed WrightMap panels keep distinct groups separate when their combined labels coincide. Supplied person-panel estimates must be real numbers. Large crossed-factor designs also retain exact cell identities when their potential combinations exceed integer precision in floating-point numbers. Grouping keys also treat column names such as collapse as data, so these facet names cannot merge separate MFRM administration designs.

• CJ design information now retains fitted position and history effects. Position-adjusted pair recommendations use the displayed presentation order. History-dependent fits no longer offer context-free recommendations or a single new-comparison information curve. Their response plots also retain these effects, comparing observed means with fitted expectations for each opponent and, when selected, judge group. Graded category curves include position effects and explicitly label the full linear-predictor axis for history-dependent fits.

• Structural refits now check the scoring structure after calibration, not just the categories present beforehand. Dropping or combining items cannot silently rescore retained items when a category loses conditional information. EFRM checks this within each frame; explanatory refits also preserve inherited item scoring. Dependence-magnitude refits apply the same check to every resolved copy before comparing their thresholds.

• Tailored bootstraps retain no-tailoring resamples as zero shifts and count non-convergence at each calibration stage separately from other failures. Saved results from the earlier bootstrap are omitted on restore. The app starts at 999 draws and displays warnings when too few usable draws make Holm-adjusted significance unattainable.

• Rasch simulation keeps numbered groups in numerical order, so DIF targets the intended last group when there are ten or more groups. Not-reached responses are documented as independently selected missing tails, not incorrectly labelled as non-ignorable departures from the response model.

• DIF resolution no longer reports zero remaining DIF when its final diagnostic assessment fails; it reports the failed assessment instead. Item splitting also checks the fitted category structure: a within-group category observed only in extreme patterns can no longer be merged and presented as a comparable split calibration.

• The conditional DIF bootstrap now compares the observed and replicated F-reference probabilities for both its marginal and familywise results. It previously used raw F values for the marginal result even when sparse refits changed the term’s degrees of freedom, making unlike statistics comparable. Results record the revised algorithm and saved analyses drop bootstrap DIF output from the superseded calculation.

• Tailored analyses now carry fitted-model and result signatures. Saved analyses omit older unauthenticated tailored tables, and current tailored item shifts are retained in app reports and result archives. Generated app code preserves whether anchors were selected automatically or supplied, so bootstrap reruns use the same anchor-selection procedure.

• Frame-invariance bootstrap replicates now retain the exact observed item comparison family. A category that was absent from an observed frame can no longer reappear in a resample and silently change the items used to centre that replicate’s frame origin. The result records the revised algorithm, and saved pre-correction frame-invariance results are omitted on restore.

• Multi-set EFRM fits with boot_reps = 0 now withhold common-unit item and threshold standard errors. Those errors require the omitted set-link uncertainty; reporting the conditional calibration component alone made them too small. Their set-unit inference flag now also records that the covariance is unavailable, rather than describing sample support alone. Single-set analytic errors are unchanged. Frame-unit plots also omit confidence intervals whenever the fitted unit families do not meet their inferential support conditions; a finite descriptive standard error is no longer drawn as an inferential interval.

• Calibration-sandwich support and row-based diagnostic support are now distinguished. A repeated person row that contributes a fitted residual but no conditional item pair still withholds row-independent fit probabilities and null generators, without incorrectly changing the calibration covariance. Residual DIF procedures use the same row-based distinction. Their paired contrasts now withhold inference when a required between-person stratum cannot supply a variance, rather than dropping it and changing the planned equal-stratum estimand.

• BTL sandwich inference now checks the empirical score-covariance rank for both independent comparisons and judge clusters, rather than relying on the number of sampling units. Repeated identical profiles can no longer produce zero standard errors and highly significant probabilities when the sandwich has no variation in the fitted directions. With identified judges, the row-based pair-fit chi-square remains descriptive and its probability is withheld; the same applies to BTL–EFRM, where judges are required. Object-location plots use the finite judge reference carried by the fitted covariance. BTL DIF follow-up adjustment retains every contrast opened by a significant term even when no resolved cell meets its support threshold. In a one-panel BTL–EFRM fit, the structurally fixed panel unit is no longer counted as a hypothesis in the Holm family for the estimated set units and origins.

• Explanatory Rasch coefficient tests now retain repeated-person cluster support from the conditional calibration. Supported repeated-person fits use a linearised delete-one-person covariance correction and a t reference with degrees of freedom equal to the number of person clusters contributing conditional information minus one; unsupported fits withhold coefficient inference, while supported fits without repeated identifiers retain the limiting normal reference. The multivariate Kent model comparison remains a first-order asymptotic test.

• Pairwise conditional Rasch calibrations now withhold item-parameter covariance and standard errors when the independent or effective person count, parameter count, or empirical score rank cannot support them. The same conditions apply to person clusters when IDs repeat. A single shared ID can therefore no longer produce false zero standard errors that equating would treat as exact; genuinely fixed anchor locations remain exact. Unit and effective-unit counts use informative conditional item-pair contributions, so appending empty or deterministic response rows cannot manufacture inferential support. EFRM conditional-stage covariance now applies the same person-support and projected-rank checks; unsupported fits retain descriptive point estimates only and cannot enter hybrid or full-bootstrap unit inference. EFRM also refuses repeated supplied person identifiers: its set-link likelihood and person bootstrap require one response row per person. Missing identifiers remain separate unknown persons. Common-item equating from supported clustered calibrations now uses contrast-specific Welch–Satterthwaite degrees of freedom rather than a normal reference. Resolved DIF magnitudes likewise use the independent person-cluster count minus one for their probabilities, confidence intervals and ETS interval-null decisions. Response-dependence magnitudes and the spread-parameter screen use the same finite-cluster reference. Ordinary and paired-comparison equating plots use the corresponding contrast-specific critical values for their interval displays. MFRM fits now retain the structural calibration’s Godambe ingredients, so compare_fits() reports CL-AIC and CL-BIC for comparable MFRM structures rather than withholding them with the two-stage EFRM criteria.

• Fixed-origin equating now refuses contradictory exact anchors instead of averaging them into a spurious shift and apparent drift in the remaining items or objects. A location-anchored polytomous item also retains its exact zero item-location standard error when sparse categories make its free threshold-spread standard errors unavailable.

• Missing-response codes are now matched before numeric conversion in ordinary Rasch and many-facet analyses. Text labels and numerically equivalent forms, such as "09" and 9, therefore behave consistently. Superitem refits also discard a stored missing code when it would become a valid summed score.

• Paired-comparison order diagnostics now assess variation between judges, not repeated presentations within a judge. Position-only effects are not treated as sequential dependence, and simulated comparisons use independent judge orders so planted dependence remains assessable. Static first-position effects are labelled separately from within-judge history effects in printed, downloaded, and app summaries.

• Fixed-origin equating can use one common item or object; estimating a shift still requires at least two. A normal-parametric fit bootstrap now uses a point mass when corrected person variance is zero and refuses an unavailable variance instead of silently changing to an empirical resampling method.

• Legacy paired-comparison dependence results without adjusted probabilities now report that Holm probabilities are unavailable in direct printing, plots, and the app. The same rule applies to residual-dimensionality printing. Raw probabilities are not substituted for adjusted inference.

• Weighted person scoring now refuses set maps that become duplicated after item-name matching, rather than silently replacing one assignment. Saved application results are authenticated against the active calibration and retained in HTML, Word and PDF reports and the results archive as externally weighted secondary person measures. The archive also carries the resolved response-cell weights needed to interpret and reproduce them, and a successful switch to Comparative Judgement clears an earlier Rasch weighted table with the rest of the calibration-specific results. A table returned by weighted_person_estimates() can be supplied explicitly through person_weights when saving outputs or writing a report; it is reproduced against the fitted model before use.

• simulate_mfrm() now treats item_sd as the realised sample standard deviation of its item difficulties, consistent with the argument name and with the other simulator scale controls. A value of zero still gives equal item difficulties.

• Simulation controls now refuse departures that the generated design cannot distinguish: DIF must leave an invariant item, EFRM drift must leave one in each affected set, a second attribute cannot duplicate the first exactly, and random or halo contamination cannot replace every identifying group, panel, attribute camp or sampling unit. A response style must leave an unstyled person because a category weighting applied to everyone is absorbed by the fitted thresholds. Halo needs differing item difficulties, while guessing, heterogeneous slopes and non-uniform DIF need person variation. Common item discrimination is recorded as a change of logit unit. Recovery output labels any deliberate departure outside the fitted model as a descriptive comparison rather than an unbiased recovery target. Partial-credit threshold structures are therefore identified as unsupported when the same simulated responses are fitted with a rating-scale model, and MFRM recovery checks every fitted facet allocation carried by the simulation. The application also requires slope departures to leave one reference item. A nonzero local-dependence request must contain an item pair, and scale controls described as realised standard deviations are refused when their spread cannot be represented numerically.

• Fit summaries and the app now distinguish ordinary pairwise conditional calibration, response-cell calibration, and semiparametric EFRM set linking.

• Generic model comparison now withholds information criteria and likelihood differences for BTL–EFRM fits. Their combined objective is evaluated at a two-stage estimate rather than maximised jointly; the fit’s equal_unit component retains the corresponding labelled descriptive comparison.

• Explanatory diagnostics and fixed relaxations now add only the part of a polytomous threshold-departure block not already represented by the active predictor design.

• BTL–EFRM judge-bootstrap degrees of freedom now follow each set’s supported path to the reference set. A strong terminal link can no longer mask a weak link earlier in the chain, including when the fitted locations are used for common-object equating. Set pairs below min_link are now omitted from the linking fit, as documented, rather than entering the likelihood after being excluded only from the connectivity check. Unit plots use these row-specific references for their intervals and omit an interval when inference is unavailable.

• EFRM set-unit support now follows the strongest bottleneck path through the linking graph. A strong terminal edge can no longer lend support across a weak upstream link, and a weak redundant edge no longer suppresses a set that has a stronger route.

• EFRM score curves are now keyed to the exact observed-item pattern within each group. A partial booklet form no longer receives the expected weighted score and information of every item in a set merely because it contains one item from that set.

• EFRM now refuses repeated non-missing person identifiers because its set-link likelihood and person bootstrap require one response row per person. Missing or blank identifiers remain distinct unknown persons.

• simulate_efrm() now refuses a zero person spread when it generates more than one item set, because relative set units then have no shared person variation from which to identify their scale. It also refuses non-zero item drift with one person group, where the shift is an ordinary item-location change rather than a violation of invariance across frames.

• Item-fit and dimensionality bootstraps now reject a sparse polytomous replicate if it loses a category and would therefore be fitted on a shorter item scale. Such draws are included in the existing failure accounting.

• A full EFRM person bootstrap now supplies every covariance exposed in unit_cov, including the joint within-frame threshold/group-unit block and the set-unit/group-unit cross-block. These matrices therefore agree with the bootstrap standard errors and frame-unit tests returned by the same fit.

• Single-value display arguments no longer flatten one-cell matrices: plot_facets() requires a plain facet name, compare_fits() requires a plain reference name, and report titles require plain character scalars. The app also withholds legacy BTL dependence decisions when an adjusted probability is absent, rather than highlighting the raw probability.

• simulate_rasch() now guarantees that an item requested through disordered has an adjacent threshold reversal, including PCM simulations with randomly varying threshold spans. Generating threshold lists are also named by item. simulate_rasch() and simulate_efrm() now refuse missing = 1, which would remove every response and leave no model to fit.

• The two equating functions now refuse a fitted model from the other response family directly, rather than passing it to data-frame coercion as though it were an item or object bank.

• split_items() now requires a plain grouping vector. A one-column matrix can no longer be flattened silently into a person-factor allocation.

• Tailored-analysis anchor items now require unique, non-missing item names in a plain vector. Matrix-shaped or repeated anchors are refused before the recalibrations begin.

• A data-driven dimensionality bootstrap now withholds its verdict when failed refits leave a probability floor above the requested alpha level. Printed, app and exported results distinguish that case from an uncalibrated residual split and retain the bootstrap resolution and replicate accounting.

• Traditional SEM is now withheld when coefficient alpha is negative. The tailored-bootstrap resolution warning also covers a Holm probability floor exactly equal to .05 and gives the first usable replicate count.

• Item-curve overlays now retain each fitted item’s class intervals under missing data, honour observed = FALSE with group displays, and reject repeated selectors. Plot palettes cycle for larger category and group sets, and residual components beyond the tenth can be selected directly.

• rasch_mfrm() now refuses non-data-frame input at entry with a direct message instead of failing later through a long-format subscript or an internal assertion. rack_data() and stack_data() now check source column names before coercion, so duplicate columns cannot be silently renamed and reshaped as different items. The built-in person-item and Wright targeting plots now also refuse a failed calibration or EFRM link instead of drawing its last numerical iterate as an estimate. Direct BTL scale, category, characteristic, dependence and frame-unit plots apply the same convergence rule. BTL transitivity and residual-dimension plots now give a direct class error when passed the wrong result object. Complete threshold, item-fit, person-fit, person diagnostic, fit-residual, facet and frame-unit plots follow the same rule. Complete exports and reports reject a failed calibration before creating output, rather than stopping later with a partial result directory. Rest-measure distractor analyses and plots also refuse a failed calibration. plot_recovery() now gives a direct class error for an unrelated object.

• A paired-comparison solution rejected by the late quasi-separation check now withdraws its complete covariance-based inference state. Composite information criteria, object and graded-threshold standard errors, threshold components and the pairwise fit probability are no longer reported from a calibration that the estimator has marked as non-convergent. Composite information criteria for ordinary Rasch fits likewise require a finite positive-semidefinite Godambe covariance and an identified positive- definite sensitivity matrix; an absolute trace can no longer turn invalid ingredients into an apparently usable penalty. A non-converged BTL-EFRM fit now likewise retains its final estimates only for diagnosis and withholds standard errors and inferential probabilities. The same rule now applies to ordinary response-data and paired-comparison optimisers that stop before convergence: locations and residual patterns remain for diagnosis, while standard errors, separation reliability and probabilities are unavailable. The exported low-level pcml() and pcml_pc() entry points now warn on the same condition and withhold their covariance-based uncertainty. Explanatory fits and their model-comparison helper now also require the unrestricted reference calibration to have converged; a simpler restricted model can no longer be assessed against a failed reference fit.

• Rasch simulation now refuses duplicate dependence pairs, repeated item selectors and non-scalar response-style types. Zero-sized DIF and MFRM interaction requests, and zero-strength dependence, are no longer reported as planted departures. Zero-size EFRM item drift is treated in the same way. Response-style/careless and halo/erratic-rater proportions are refused when they cannot be realised without one mechanism erasing the other. Completely-at-random missing cells are now sampled only from responses not already removed by planted speededness, so both requested counts remain observable; incompatible proportions are refused.

• Simulation recovery now verifies that the fitted responses and allocations belong to the supplied simulation before comparing parameters. Row and item order remain immaterial. EFRM sets, person groups and paired-comparison panels are matched by membership, so renamed frames retain their recovery summaries; a different partition is refused.

• Fixed departures generated by the app for explanatory models are now orthogonal to the nominated predictor design. The requested magnitude can no longer be partly, or in a saturated design wholly, absorbed by the explanatory coefficients.

• compare_fits() now treats row order, item-column order, arbitrary person or judge labels, and count compression as changes of presentation rather than changes of data. It still distinguishes different person or judge allocations, and a half-scored tie remains one comparison rather than two independent opposing judgements.

• Simulation recovery now retains the identified origin for externally anchored Rasch and paired-comparison fits. distractor_rescore() refuses to propose scores when no respondent selected the full-credit option. Export functions also reject an invalid fitted object before creating a partial output directory.

• Person scoring, test information and model-curve plots now follow an externally anchored calibration origin instead of searching or drawing on a fixed zero-centred interval. Model curves are withheld when calibration or EFRM linking did not converge.

• When an ordinary Rasch analysis contains repeated person identifiers, its pairwise item-calibration sandwich now sums score contributions within person before estimating uncertainty. Point estimates still use every response row; duplicating an occasion no longer makes item standard errors spuriously smaller. This covariance is also used by explanatory refits and resolved DIF magnitudes. CL-BIC counts contributing persons rather than response occasions. Model comparisons now include the independent-person allocation in their data check and withhold composite information criteria when it differs from the reference fit. Unique-identifier analyses are unchanged. Item-trait and class-interval ANOVA probabilities are now withheld when identifiers repeat because their reference distributions count response rows rather than independent persons. The ordinary item-fit bootstrap is refused for the same reason; descriptive fit statistics and repeated-measures DIF remain available.

• The threshold diagnostic now determines exactly which partial-credit categories are modal over a non-zero interval. It no longer depends on a finite plotting grid that could miss a very narrow category region. Thresholds that are equal by construction now leave only the extreme categories modal; their crossing points differ by rounding, and that gap was previously reported as an interval.

• The semiparametric EFRM set link now judges grid truncation from the persons with a finite location in at least one set. Persons at the same extreme tail in both sets remain in the likelihood but no longer count towards the two per cent edge-mass rule, since the masses explaining them sit on the end of any finite grid. The rule refused about one person resample in twelve on an ordinary design, so a bootstrap frame_invariance() rarely reached its usable-replicate minimum and its usable replicates under-represented the resamples richest in extreme scores. The rule is now documented.

• Paired-comparison EFRM simulation now requires plain vectors for panel units, set units and set origins instead of silently flattening matrices.

• Anchor-table columns must now contain one plain value per row. Matrix-valued data-frame columns can no longer be flattened or recycled into a different set of item or threshold anchors. Paired-comparison object anchors apply the same plain-vector rule.

• Externally imposed person-estimation weights and item-set maps now reject matrices instead of flattening them into a different weighting scheme. Set-map shape is checked before label canonicalisation. BTL-EFRM panel-unit reconciliation likewise refuses asymmetric within-set covariance matrices rather than silently symmetrising their precision weights.

• Numeric columns in external calibration banks must be vectors. Matrix-valued data-frame columns can no longer be flattened and misaligned with their item or object names.

• User-defined DIF contrast weights must be plain numeric vectors; shaped weights can no longer be flattened into a different contrast.

• When dichotomous paired-comparison ties are divided equally between the two outcomes without a judge column, the two half rows now remain one sampling unit in the sandwich covariance. Count-compressed ties retain their original replication count in that calculation.

• Item selectors in ordinary and EFRM fits, EFRM item-set maps, and paired-comparison margin columns now reject matrix-valued inputs rather than flattening them into a different fitted design or response scale. Paired-comparison EFRM panel and object-set maps apply the same rule. Multiple Ratings facet, item, interaction and factor selectors are likewise required to be plain vectors. By-value person identifiers in ordinary and EFRM fits must also be plain vectors. By-value Rasch, EFRM and Comparative Judgement person or judge factors and EFRM frame groups now apply the same rule.

• EFRM and paired-comparison EFRM omnibus Wald tests now refuse materially asymmetric covariances. A singular covariance is used only when the effect lies wholly in its estimable subspace; an effect in a discarded direction is no longer silently ignored. The Kent composite-likelihood calibration now applies the same covariance validity gate to its Godambe covariance and inverse sensitivity.

• Extrapolated extreme-score person locations are now reported only when the adjacent finite-score estimates support a stable geometric continuation. Otherwise score_table() refuses the extrapolation and person_extrapolated() retains the fitted Warm estimate.

• Uncertainty-weighted adaptive paired-comparison recommendations now require an aligned, finite, symmetric positive-semidefinite object covariance (or finite object standard errors when no covariance is stored). Information- only recommendations remain available with weight_se = FALSE.

• Covariance gates and public scalar controls now reject malformed shaped, classed or complex inputs consistently. Manual dimensionality subsets and DIF design selectors also refuse duplicate, overlapping or ambiguous specifications before fitting.

• Fit-derived person, targeting, information, scalogram and Wright-map outputs now verify EFRM link convergence from the model’s per-edge status as well as its overall convergence flag. This also protects restored or modified fits whose two records do not agree.

• Tailored-analysis bootstrap runs can take an explicit seed without changing the caller’s random-number state. The app now validates its replicate count and seed and includes that seed in the displayed R code.

• Judge-level unexpected-object and unexpected-pair flags now use Holm-adjusted approximate normal probabilities over the eligible family. Their earlier per-row z cut-off did not control familywise error.

• Paired-comparison DIF now retains anchored, failed or otherwise unavailable planned follow-up contrasts in the Holm family size. A failed resolution can no longer make the surviving adjusted probabilities smaller.

• Response-dependence magnitudes, resolved DIF contrasts, both equating procedures and conditional frame-invariance comparisons now refuse or withhold Wald inference when their fitted covariance is unavailable, asymmetric or materially indefinite. Point estimates remain available for description where the comparison itself remains defined. Multiple Ratings DIF also withholds its covariance-weighted pooled point estimate in this case, because the invalid covariance would otherwise determine the pooling weights themselves.

• EFRM corrected-moment set linking again runs its requested bootstrap. Its non-iterative edges record no optimiser status; those structural missing values were incorrectly treated as non-convergence, leaving unit standard errors unavailable.

• Crossed EFRM group-unit decompositions now withhold their GLS table when either the unit covariance or the derived coefficient covariance is asymmetric or materially indefinite.

• Residual PCA plots and biplots now use a finite neutral plotting range when their displayed loading vectors are identically zero.

• plot_facets() now refuses non-character facet selectors rather than allowing a factor code to be interpreted as a list position. Item selectors likewise refuse matrices and other shaped objects rather than treating a one-cell object as a scalar item index. Multiple-choice item subsets also reject duplicate canonical names. Batch item and person plot exports now validate their complete selection before opening an output file; fractional row numbers can no longer be truncated to a different person.

• The application DIF summary now distinguishes unavailable uniform and non-uniform tests from non-significant adjusted results.

• Available-case coefficient alpha now requires the pairwise covariance to be positive semidefinite at a scale-relative numerical tolerance. The former absolute tolerance could accept materially indefinite binary-item covariance matrices.

• Hybrid EFRM set linking no longer falls back silently to a person-only bootstrap when the joint stage-one covariance cannot be used. Requested unit inference is refused because that fallback omits calibration uncertainty; a descriptive fit remains available with boot_reps = 0.

• Equating now validates bank covariance symmetry, positive semidefiniteness and agreement with stated marginal standard errors at the covariance’s own numerical scale. Precision-weighted links and pooled Multiple Ratings DIF magnitudes no longer use an absolute variance floor.

• Whitespace-only DIF levels are treated as missing factor metadata rather than as a person or judge group. EFRM now fits the same trimmed frame-group labels that it validates, so visually identical labels cannot define different frame units. Whitespace-only data, item and factor names are refused.

• Custom DIF contrast weights are normalised by their direction rather than their absolute magnitude; rescaling every weight cannot make a valid contrast disappear.

• Structural item changes now reject missing and blank item names, and direct DIF splits treat blank group labels as missing. Externally weighted person estimates identify extreme response patterns from their responses rather than an absolute weighted-score tolerance, and no longer merge distinct weighted totals for computational caching.

• Numeric NaN values supplied as BTL judge-group labels, including grouped ICC displays, are now treated as missing rather than as a literal "NaN" group. Missing and whitespace-only BTL anchor names are also refused explicitly.

• Externally weighted person estimates now normalise relative weights without underflow when all positive weights are close to the smallest finite double. Exact fitted item names take precedence when item weights or item-set maps are matched, so literal leading or trailing spaces in a data-column name are not discarded; padded selector input remains available when unambiguous. Native EFRM person estimates likewise identify extreme responses from the response pattern rather than an absolute weighted-score tolerance, and no longer merge distinct weighted scores for computational caching.

• MFRM marginal fit summaries return an unavailable value when every contributing response-cell residual is unavailable, rather than emitting a warning and storing NaN. MFRM simulation recovery now identifies the planted rater facet from the strongest level-label match, using its name only to resolve a tie, instead of assuming that it is the first fitted facet. sim_recovery() also refuses a non-convergent fit or a model family that does not match the generating simulation, rather than comparing coincident parameter labels. Replicated simulations also retain a valid seed at the upper integer boundary instead of overflowing during seed incrementation.

• The paired-comparison fit bootstrap now gives a deliberate refusal when a compressed comparison count exceeds the integer range accepted by R’s multinomial generator.

• In the app, planned Multiple Ratings DIF contrasts now select underlying items rather than internal item-by-facet response cells. Saved-analysis notices now also name the superseded derived result actually omitted, including frame-invariance results, rather than always referring to a fit bootstrap.

• Parallel bootstrap requests from an installed package now start the requested socket workers. An incorrect namespace-name check had made EFRM, BTL–EFRM, DIF, dimensionality and fit bootstraps run serially while still recording the requested worker count. The application now refuses fractional or unavailable worker selections instead of truncating them silently.

• Joint bootstrap references now require every transformed statistic or fitted parameter in a replicate to be finite. Infinite log mean-squares, frame units, or replicated eigenvalues are treated as failed draws rather than entering maximum-statistic or covariance calculations. EFRM NPML compression also retains the full precision of weighted sufficient statistics instead of pooling values that agree only to 12 significant digits.

• Kent-adjusted model comparisons and unit-family Wald tests now withhold inference when their estimated covariance is singular in a tested direction or materially indefinite. BTL–EFRM panel-unit reconciliation no longer adds a fixed variance floor to its precision weights. Test-information curves are withheld for an unconverged calibration, and zero-degree-of-freedom model summaries no longer display an infinite or undefined chi-square ratio. Targeting summaries and WrightMap displays now likewise refuse a non-convergent calibration. Item-trait detail retains an interval with unavailable or zero model variance but does not manufacture an infinite residual from it. Fully fixed item locations are excluded from item separation reliability, rather than being counted as estimates with zero standard error.

• PSI bands are now described as person separation quality rather than as the power of fit tests. The old power_of_fit component remains as an alias for saved-object compatibility. EFRM documentation now states the identification conventions for ordinary and paired-comparison frame models exactly.

• Person-subset dimensionality tests and simulated residual references now refuse repeated person identifiers. Their current null generators treat response rows as independent and cannot preserve within-person dependence. The fixed-split interval is described as Clopper–Pearson rather than as an exact test of dimensionality.

• frame_invariance() with bootstrap uncertainty now requires the standard 90% usable-refit rate and reports usable, non-converged and other-failure counts. A contrast with zero bootstrap uncertainty is retained descriptively, but its Wald probability is withheld. Every observed frame in a compared item set must also yield a usable separate calibration; unavailable frames are no longer silently omitted. An item dropped or rescored by one separate calibration is recorded as an unavailable comparison and retained in the multiplicity family. The procedure now gives a clear refusal when no frame pair retains two items with a comparable category structure.

• Schema-2 application projects made with the earlier text signature now reach the intended maximum-statistic-result migration. Multi-design test information plots show standard-error-of-measurement curves as well as information curves.

• Wald follow-up tables now withhold probabilities when a contrast has zero estimated uncertainty. Positive finite standard errors remain testable, so rescaling a predictor or contrast cannot change inference. This applies consistently to DIF, frame units, explanatory coefficients, superitem spread and both ordinary and Comparative Judgement equating.

• Missing numeric identifiers, including NaN, are no longer converted into literal factor levels. Unknown person identifiers remain separate in DIF and tailored-bootstrap calculations, and malformed NaN anchor indices are refused.

• A Comparative Judgement anchor whose object has no usable comparison after data preparation is now refused instead of being silently removed from the fitted anchor set.

• Paired-comparison equating now refuses malformed bank df_location and polytomous score-scale attributes rather than silently replacing the former with an asymptotic reference or coercing the latter from another class.

• Saved schema-2 application analyses carrying frame-invariance results from the earlier comparison rules now reopen after that derived result is authenticated and omitted; the application asks the analyst to recompute it.

• A tailored-analysis bootstrap that cannot retain enough refits now signals the standard structured refusal, including requested, usable, non-converged and other-failure counts.

• frame_invariance() now retains excluded comparisons in its Holm family. Bootstrap discrimination probabilities are withheld when either fitted slope is on its imposed boundary; the ratio remains descriptive.

• dimensionality_test() documents that its default split, chosen from the residuals, makes the ordinary fixed-split binomial rule anti-conservative. The earlier documentation claimed the procedure held an exact null. A new B argument calibrates the split by a parametric bootstrap under the fitted model (score-conditional replicates, each refitted and split afresh from its own residuals), giving p_boot, the mean replicate proportion prop_null, and a verdict at the chosen alpha level. An automatic split without bootstrap calibration is now descriptive and has no binary verdict. The application can run the bootstrap and carries the exact item split and result into saved analyses, tables and reports. Refactored residual_pca() and the fit-bootstrap refit so the replicates travel the observed computation.

• A category observed only in extreme response patterns (every answered item at its floor, or every one at its ceiling) carries no pairwise conditional information, and its partial-credit threshold diverged during estimation until the projected information matrix was reported singular. rasch() now merges such a category with its neighbour during data preparation, with a note, and drops an item with no conditional information at all. An item with a free threshold in that position is refused when it is anchored; an item whose every threshold is anchored keeps its coding, since nothing is estimated for it. pcml() names the item and category instead of reporting the singular matrix.

• Tailored analysis step 3 now uses average item anchoring (RUMM2030’s convention): the retained items are constrained to the mean location of the step-2 calibration rather than each being fixed individually, so the step-3 standard errors reflect the uncertainty of the re-estimated locations. rasch() and pcml() accept anchors with average = TRUE, and the application offers the option. Anchor documentation now distinguishes a change of origin from the extra relative constraints imposed by several fixed parameters, and states that standard errors are conditional on the supplied anchor values. drop_items() refuses to remove an externally anchored item because that would silently change the fitted scale.

• report_html(), report_document() and save_outputs() now say why a scree plot is absent: a residual-PCA refusal or a reference that could not be completed is warned about and, in the reports, noted in place of the plot, where the precomputed-scree change had dropped it without a word. The document template wraps a long refusal so it stays inside the figure. spread_test() documents that Andrich’s tabulated bounds are exact for two- and three-item subtests and conservative for larger ones. distractor_analysis() no longer warns when nobody chose the keyed option.

• Bootstrap maximum-statistic probabilities now standardise each simulated row against the other rows. This gives the simulated and observed statistics the same external standardisation and removes the anti-conservative shrinkage caused by using a null row in its own mean and standard deviation. Constant statistics remain in the declared family. Saved projects omit fit-bootstrap results made with the earlier adjustment and ask the analyst to recompute them.

• Residual scree references and tailored-analysis bootstraps now require the same usable-refit proportion as the fitted-model bootstrap: at least 90% for 30 or more requested refits. Their results report requested, usable, non-converged and other-failure counts. Scree and paired-comparison dimensionality functions accept a seed and restore the caller’s random- number state.

• The application keeps the computed dimensionality and EFRM invariance analyses with the saved project and passes them to reports and complete exports. The selected EFRM uncertainty method is therefore not replaced by a conditional rerun. Complete exports and editable reports now include the dimensionality tables and figures for Rasch and Comparative Judgement analyses.

• Simulated response styles now preserve declared local dependence by rebuilding dependent responses from the styled source responses, and their log-probability tilt remains stable at large finite strengths. btl() now refuses count-compressed rows with judgment order: the row does not retain the sequence needed to estimate exposure and carry-over effects. The dimensionality diagnostic retains the same guard for older saved fits.

• Scree parallel analysis now draws from the exact Rasch distribution conditional on each person’s observed score and missingness pattern. Its shaded reference band runs from the simulated mean to a finite-simulation familywise 5% upper critical curve. Each simulated maximum is standardised against the other draws, so its reference matches the externally standardised observed value. The returned table gives marginal and adjusted upper-tail probabilities for every displayed component. Saved result and application-project signatures use platform-independent binary serialisation; analysis files carrying the earlier text signatures remain valid across Windows and Unix line endings.

• Crossed Extended Frames groups retain their original column names in fitted objects and factorial unit tables. Frame-defining groups are therefore excluded correctly from a subsequent DIF analysis even when their names contain spaces, reserved words or punctuation.

• Explanatory models in the application now preserve exact predictor names and distinguish literal colons from interactions. Ordered predictors keep unique adjacent-contrast names. Comparative Judgement DIF display settings are kept outside the signed result, and saved primary DIF analyses are checked against the active fit whether or not a bootstrap was requested.

• Fit-bootstrap and primary DIF results now carry their own integrity fingerprints. Before an analysis is restored, reported or exported, the package checks the fitted-model identity, replicate accounting, declared family, adjusted probabilities and the agreement between detailed and summary tables. Edited or incomplete results are refused.

• DIF bootstrap results are now checked before they are restored, reported or exported. The replicate matrices, accounting totals, declared test family, primary DIF tables and derived probabilities must agree; an incomplete or inconsistent result is refused before any output files are written.

• dif_bootstrap() adds an optional sensitivity analysis for DIF. It conditions on person scores for Rasch, explanatory and Multiple Ratings models, on item-set subtotals for Extended Frames, and on the fitted outcome model for Comparative Judgement. Each replicate refits the model and repeats the declared DIF analysis. Adjusted probabilities use a single-step minimum-p reference over the complete family; dif_anova() or btl_dif() remains primary. They describe the fitted global invariant null, not strong familywise control when another member has DIF. The application runs the analysis in the background and retains it with its primary DIF specification in saved projects, reports and exports. Validation now separates power for the affected member from familywise error among invariant members under partial alternatives in every supported model family.

• Fitted-model provenance now includes the complete fit, including person identities, comparative-judgement thresholds and sequential-dependence estimates. Application project files use a sealed schema that detects a changed dataset, fit, history or saved result before restoration, and structurally validates every stored fit and history entry. Schema-1 files are structurally checked and upgraded when opened; saved bootstrap or DIF results that predate the current fingerprints are omitted with a warning, while the data, fit and history are retained. Saving the analysis again records the new integrity information.

• Fit bootstraps now withhold inference when fewer than 90% of 30 or more requested refits are usable. This prevents sparse polytomous analyses from reporting a null distribution selectively thinned by failed category structures. Conditional generation is evaluated on the log scale, so finite extreme thresholds no longer overflow. Printed bootstrap objects give a short summary rather than dumping the stored replicate matrices.

• DIF results carry the identity of their fitted model and cannot be exported or reported with a different calibration. Saved application projects now validate every fitted-model history entry, not only the active one. A saved DIF bootstrap must also match its primary DIF specification and restored controls. Reports include the corresponding DIF magnitude estimates.

• Adjusted probabilities for DIF magnitudes and contrasts, paired-comparison DIF magnitudes, explanatory coefficients and diagnostics, MFRM interaction follow-ups, and tailored analysis now retain the complete declared family when one probability is unavailable.

• The paired-comparison application example separates sequential dependence from residual dimensionality. Its main dataset has planted exposure, null carry-over and null judge factors; a second dataset supplies the dimensionality example.

• Paired-comparison fitted-design bootstraps with exposure and carry-over have been calibrated for dichotomous and four-category responses at 30 judges. Total and adjusted pair, object and judge error remained near 5%; all 238,800 refits were usable.

• Fit inference now retains its predeclared family when an item or frame-unit probability is unavailable. BTL-EFRM omnibus tests are withheld rather than reduced when a requested covariance coordinate is missing. Fixed-location item bootstraps keep each location with its own response row, and whole-test fit-residual nulls use one fixed item set. Saved application bootstraps are checked against the active restored fit. Parallel bootstraps loaded from a development source tree run serially instead of finding another installed package on socket workers.

• fit_bootstrap() now calibrates person fit under its default score-conditional null. It reports marginal probabilities and a maximum-statistic adjustment across persons for each fit statistic. Paired-comparison fits are also supported: outcomes are generated on the fitted comparison design, the model is refitted, and bootstrap probabilities are returned for the total pairwise chi-square and pair, object and judge fit. Ordered thresholds and sequential history effects are reproduced. The application runs either fit bootstrap in a cancellable background process and carries the result into its tables, saved analyses and reports.

• Fit bootstraps now refuse a non-converged observed fit, reject non-converged refits, and report non-convergence separately from other failed replicates. A maximum-statistic adjustment is withheld for the complete predeclared family when any member lacks a usable joint null. Bootstrap results carry their model identity and can only be exported with the fit that produced them. The application shows unavailable inference neutrally. Documentation now states that adjustment is within each statistic under the fitted global null; it is not strong familywise control for otherwise fitting members when another member misfits.

• Bootstrap accounting now follows the quantities actually used. EFRM and BTL-EFRM size their rank requirement from the free directions in the largest covariance block, rather than the concatenated row retained from each replicate. Item-fit bootstrap probabilities are withheld when too few refits or too few values of a particular statistic remain; the item table records the usable count for every statistic, and Holm adjustment retains the full family when one result is unavailable. Seeds must be non-negative whole numbers. The asymptotic item-fit probabilities remain available as descriptive diagnostics but are labelled as approximate in summaries, reports and the application. A calibrated bootstrap probability is the operative result after fit_bootstrap() is run.

• A person-item map now refuses an EFRM item-by-frame cell and person-group selection with no common response cell. It no longer replaces the requested item silently with all items from the selected group.

• Release review, second pass. fit_bootstrap() warns when the requested replicates cannot reach Holm-adjusted significance at .05 for the item count — the floor is 2L/(B + 1) two-sided, so a run whose adjusted probabilities could never flag anything says so before it spends the time. A restricted person-item map carries the information of its own selection: test_information() gains an items argument and plot_pimap() passes its item subset through, where the curve previously described the whole instrument over a subset’s distributions. A person group is addressed unambiguously: plot_pimap() accepts the qualified "factor: level" form and refuses a bare level that two factors share, naming the qualified candidates, where the first factor was silently taken; the application’s selector now offers the qualified form throughout, and its generated code quotes names by deparse() so a level carrying a quote still yields runnable code. A saved analysis carries its bootstrap: the project stores and restores fit_bootstrap() results, and reports and complete exports render the analysis as run — the application’s configured DIF model and any bootstrap null — rather than recomputing defaults (save_outputs(), report_html() and report_document() gain dif and bootstrap arguments), and every bootstrap probability a report renders is the Holm-adjusted one. Bootstrap covariance acceptance is dimension-aware: successes must also exceed the free directions in the largest covariance block used by the fit, since the stored row contains several separate constrained blocks. The alpha–phi cross-covariance requires 30 joint draws and a majority, and is withheld with a warning below that. On an extended-frame map restricted to one group, the information curves are that group’s designs alone, and test_information() refuses fractional indices and accepts an extended-frame fit’s underlying item names.

• fit_bootstrap() generates its replicates conditionally by default: each person’s responses are drawn from the Rasch conditional distribution given their observed raw score over their own observed items, with sufficiency cancelling the person parameter. Release review caught two defects in the ability-resampling default this replaces. Resampled estimates carry their estimation error, and the standardised statistics feel the inflated spread as the sample grows — infit z rejected 14.4% of correctly fitting items at 4,000 persons, against 2.1% conditionally, with the chi-square at 5.2% and the fit residual at 3.7%. And abilities drawn independently of each person’s missingness sever any tie between who answers and what is missing: a linked-booklet design’s observed 1.29-logit group difference fell to 0.12 in resampled replicates and is reproduced at 1.28 conditionally, exactly, because the scores that carry it are held; the recorded booklet scenario holds the item-level rate at 1.9% and Holm familywise error at 3% for the chi-square and 6% for the fit residual. The ability-sampling schemes remain by name. Review also tightened the arithmetic around the resolution floor — a two-sided bootstrap probability cannot fall below 2/(B + 1), so Holm across L items needs B of at least 40L to flag at .05, which the documentation, the application’s guidance and its adjusted fit-residual display now state and follow — and two operational faults: the shared bootstrap harness now degrades to a serial run with a warning when its sockets cannot be opened (every draw is made before dispatch, so the result is identical), where it previously errored inside documented examples; and bootstrap acceptance again honours the documented contract of at least 30 successful replicates and a majority of those requested, where a bare majority — 16 of 30 — had been allowed to price a covariance.

• fit_bootstrap() gives the item fit statistics a null distribution they can be referred to. Every one of them is computed at estimated person locations and referred to a distribution derived as though those locations were known, and each is miscalibrated by an amount that moves with the sample. The item-trait chi-square forms its class intervals by selecting on the estimates, so the expected interval means carry a bias that does not shrink while the intervals grow: at eight items a correctly fitting item is rejected 8.9% of the time at 250 persons and 99.9% at 4,000, and the whole-test total rejects every correctly fitting instrument from 1,000 persons on. The fit residual fails the other way, its null SD running from 0.74 at 250 persons to 1.00 at 4,000 with a mean drifting to -0.36, so the conventional 2.5 cut flags under 1% of correctly fitting items and means something different at each sample size. Infit z beyond 1.96 flags 11.5% at 250 persons and 68.4% at 4,000. The bootstrap generates from the fitted model and sends every replicate down the same road as the observed data, so the same bias enters the null and cancels: under the score-conditional default, item-level error across all five sample sizes is 0.039-0.051 for the chi-square, 0.039-0.054 for the fit residual, 0.033-0.050 for infit z and 0.037-0.055 for outfit z, and the whole-test total holds at 0.020-0.080. Power is retained in full — an item generated at discrimination 2.5 is detected 1.000 of the time at 2,000 persons while the clean items beside it come back from 0.723 to 0.037. One set of replicates serves every statistic, since a replicate must refit the whole model in any case, and the whole-test readings include the mean and SD of the item fit residuals: the convention reads that SD against 1, and it is nowhere near 1 until the sample runs to a few thousand.

• adjust_N is removed from rasch(), rasch_mfrm(), rasch_efrm() and rasch_explanatory(), and from the application’s run settings. Rescaling every item’s chi-square to a reference sample size applies one constant across items, so it can move how many items are flagged but not which: at 2,000 persons its power on a planted 2.5-slope item equalled its error rate on the fitting items beside it at every reference value tried, and at a reference of 500 it silenced that item in all 120 replicates. The justification offered for it — that a fitting item’s chi-square is unchanged by sample size while a misfitting item’s grows — does not hold: under a model true by construction the mean chi-square runs 4.0, 7.9, 11.5, 18.2 and 30.7 for 250 to 4,000 persons on fixed degrees of freedom. chisq_detail() loses adjust_factor, chisq_unadjusted and the chisq_adjusted interval column with it; its intervals now sum to the item’s chi-square directly. Saved application projects carrying the setting still open, the value being ignored.

• The application’s Summary panel renders after a comparative judgement fit. Its headline tiles and test-of-fit table sat inside the panel shown only for the other models, where their own show-for-comparative-judgement condition contradicted the ancestor’s and the panel stayed empty; the override history moves out of that panel with them, matching the object estimates it already reported.

• plot_btl_dependence() includes a fitted first-position advantage in its partial-residual baseline. The baseline summed only the exposure and carry-over terms, and the expected-score map is nonlinear, so on a fit made with position = TRUE every observed and fitted departure in the plot was displaced; the estimates were never affected. Found by review; the regression test pins the corrected baseline’s partial residuals at zero under the fitted model.

• The application’s Items page can bootstrap the fit statistics on request: one button runs fit_bootstrap() on the active fit and swaps the chi-square and fit-residual probabilities in the table, its CSV, the headline tile and the note for the calibrated ones, labelled as such and cleared by the next estimation. The default 999 replicates clears the familywise resolution floor — the smallest Holm-adjusted probability is the item count over B + 1, so flagging at .05 needs at least 20 replicates per item, which the control’s help states.

• The application’s Wright map controls sit above the plot. They rendered in the card footer, where choosing the WrightMap renderer revealed its panel and display controls two screen-heights down; a control that changes what is drawn belongs where the change is made. The data preview also shows 25 rows rather than 10 — the sidebar it sits beside is longer than ten rows ever were.

• The application’s panels start closed, except where the accordion is the page: Local dependence, Trait dimensionality and DIF open their first panel, since everything those pages show lives inside one. Elsewhere twelve accordions opened a panel on arrival — ten naming one and two relying on the default, which opens the first — so every page presented a choice already made. More also loses the only route to reopening a saved analysis: Open saved analysis now sits beside Upload data on the welcome screen, since reopening one is a way to start.

• The workflow and comparative judgement vignettes demonstrate the application over the analysis they set out in code, rather than a first panel or two: the workflow vignette walks Data, Summary, Items, the class-interval detail, Persons, Targeting, local dependence, DIF, the R code disclosure and export, and the comparative judgement vignette gains the section it never had, covering what changes when the roles are two objects and an observed preference. The data structures vignette shows an explanatory design at both levels with the same predictors, so the difference between an item-level and a threshold-level file is visible as the shape it is, with the number of rows a threshold design needs derived from the data and the three errors that refuse one named.

• plot_pimap() can restrict either side of the map: group takes one level of a fitted person factor, and items takes item names or one item-set name of an extended-frame fit, whose virtual item-by-group cells are matched through their underlying items. The selection is named in the legend, so a restricted map cannot be read as the whole instrument. The application offers both as person-group and item-set controls beside the person-item map.

• The application colours its headline tiles by the side of the model they describe: persons blue and items amber, the colours the Wright and person-item maps already use. The data page follows, with rows blue and columns amber. Tiles that report a check keep their own reading, now green when it passes and red when it wants attention: item and person misfit, disordered thresholds, separation, coefficient alpha, the item-trait probability and the explanatory comparison. Amber is left to mean the item side alone, so the two readings cannot be confused.

• The application’s comparative judgement example is dichotomous. It also carried an ordered preference and a margin, which invited a polytomous fit on its ten judges; a polytomous fit with the judging-order covariates estimates as many parameters as there are judge clusters, so the example withheld its own cluster-robust inference. The ordered response is demonstrated on designs that can carry it.

• The application’s post-estimation notification lists its notes. They were joined on a newline, which the notification renders as ordinary whitespace, so a fit reporting more than one note read as a single run-on sentence – withheld cluster-robust inference and a withheld carry-over probability ran together into one unreadable line.

• A judge factor named by judge may name judges the fit set aside – one who only ever tied, or whose rows were dropped – since a map built from the source data necessarily does. Those entries are ignored and reported in the notes rather than refused, in btl_dif() and plot_btl_icc() alike; every fitted judge must still carry an entry, which is what catches a mistyped judge name.

• Bootstrap covariance estimation requires at least 30 usable replicates and a majority of those requested. The absolute floor prevents a covariance from being reported from 16 draws when boot_reps = 30; the rank rule may raise the minimum for a larger covariance block.

• pcml_pc() now labels unnamed response matrices consistently, and direct pcml() calls reject an empty anchor table. Available-case item-rest correlations exclude respondents with no observed rest score.

• Frame-invariance bootstraps keep singleton person-group strata in their original group. Paired-comparison order effects are refused when the comparison design confounds them with the object locations.

• Automatic DIF follow-ups now use the adjustment method and significance level supplied to dif_anova(). dif_posthoc() also checks item selectors before fitting the contrast family.

• A saved app analysis now restores its data roles, estimation controls and embedded supporting data, so it can be re-estimated after it is reopened. Simulation recovery covers paired-comparison Extended Frames, including object locations, panel and set units, and set origins.

• Shiny background fits are tied to the data and analysis that launched them. A completed EFRM or paired-comparison frame fit is discarded if that context has changed, and opening a saved analysis cancels work still in progress. A new Comparative Judgement fit clears earlier requested DIF and frame results, while reopening a saved fit retains the results stored with it.

• The Shiny application can simulate ordinary and explanatory Rasch, Comparative Judgement, Multiple Ratings, Extended Frames and paired- comparison Extended Frames data. Model parameters can be varied and model-specific departures planted. simulate_efrm() now supports item drift, careless response and missingness, while simulate_btl_efrm() can plant erratic judges. A positive planted proportion affects at least one observation whenever the requested departures can coexist; incompatible mixtures are diagnosed. Rasch simulation truth records speeded persons and the cells selected for missingness.

• weighted_person_estimates() provides supplementary person measures from externally imposed item or item-set weights. It leaves the fitted calibration and the ordinary measures used for fit, reliability, targeting and DIF unchanged. Its Warm correction and sandwich standard error now both use the variability of the weighted score; the earlier correction treated weights as replicated observations. A new first-listed vignette gives the data structure required by every model family.

• Simulation truth now retains person identifiers, so recovery matches person estimates by ID after rows are reordered. A planted second trait has its requested realised correlation. Paired-comparison simulators balance comparisons across the declared judges and refuse designs with too few comparisons. MFRM interaction probabilities require adequate support in every item-by-facet cell rather than only at the pooled facet level.

• EFRM convergence now covers the fitted nonparametric masses as well as the set transformation. Linking uncertainty requires at least 30 usable resamples and more than half of those requested; the requested, usable and failed counts are retained in the fit. The full-bootstrap attempt remains recorded when too few refits succeed and hybrid standard errors are returned. BTL-EFRM applies the same usable-resample rule and accounting. Fixed-iteration EFRM linking no longer recalculates an unused likelihood at every intermediate mass update.

• Automatic DIF resolution now acts on the adjusted omnibus result. Pairwise contrasts describe the location of a multifactor effect but do not impose a second significance test. Thin or incompatible factor cells are still refused. A prior manual split now counts as one resolved source item rather than several anchors, and its provenance survives later item dropping or subtest formation. The reported residual count is the number of source items, not the number of item-by-factor terms.

• BTL-DIF retains the validated Welch reference for two-cell contrasts and uses the least-supported effective-judge cell as a conservative reference for contrasts spanning more cells. In 500 balanced four-cell null fits, the new reference rejected 3.6%, against 6.2% for the superseded pooled-count rule. BTL-EFRM no longer counts a set unit fixed after an identification failure as an estimated parameter.

• The application uses Holm-adjusted probabilities for its DIF, frame invariance, threshold-spread, explanatory and supplementary item-fit decisions. BTL judge-group factors that vary within judge are refused rather than reduced to the first comparison row.

• Subtests must be formed before DIF splitting. A group-specific split copy cannot be combined because it does not provide a common item across groups.

• BTL-EFRM now refits every object set after its panel units are reconciled. The reported object locations, expected probabilities, composite likelihood and equal-unit comparison therefore come from the same fitted parameters. Previously, stable sets retained their independently optimised locations while their probabilities were evaluated at the reconciled units.

• Multifactor DIF magnitudes now match the adjusted omnibus estimand. Ordinary and paired-comparison main effects average complete factor cells equally over nuisance factors; interactions use differences between differences. DIF contrasts are withheld when the compared groups do not share an observed score structure. ETS classifications use the adjusted probabilities for both significance and departure beyond category A.

• Structural refits preserve score categories. Subtests, DIF splits and EFRM frame resolutions are refused when a required score category is absent, rather than allowing ordinary data preparation to renumber the scores. Resampling refits treat category loss as an unsuccessful replicate. Supplied app anchors and scoring keys now fail closed when malformed, inapplicable or unmatched.

• Crossed EFRM factorial tests and BTL order-effect tests now retain raw probabilities but use Holm-adjusted probabilities for decisions. In fresh null simulations, crossed-EFRM familywise rejection was 5.55% over 2,000 fits (5.93% pooled over 3,000); the three marginal rates were 5.25–5.75%. BTL familywise rejection was 5.9% over 1,000 fits with position, exposure and carry-over fitted together.

• simulate_btl() now constructs a second object attribute with the requested realised correlation, rather than obtaining that correlation only in expectation. The largest error over 1,200 short and long object sets was 3.9e-16; dimensionality power was 87% in the re-run of the strong design. Structured simulator options are checked before their components are used.

• DIF splitting omits factor levels for which an item has fewer than two observed categories, and refuses a split unless at least two levels remain. EFRM refits and wide MFRM conversion use collision-free internal names; named BTL-EFRM panel maps take precedence over data-column names. Numeric NaN responses are refused rather than treated as ordinary missing data.

• EFRM and BTL-EFRM retain raw unit-test probabilities but use Holm-adjusted probabilities for decisions. Omnibus tests are adjusted across the reported unit families. BTL-EFRM follow-up contrasts form one family across panel units, set units and set origins, rather than three separately adjusted tables. The application uses the same adjusted probabilities. In null simulations, EFRM omnibus-family rejection was 3.5% and follow-up-family rejection was 1.6% among 489 analysed fits. After the reconciled-panel BTL-EFRM refit, the corresponding rates were 3.9% and 3.0% over 1,000 null fits in the six-judge-per-panel caution design. A 500-fit supported-design top-up gave 4.6% raw set-unit rejection and 0.934 interval coverage.

• Character role arguments are resolved from matching column names rather than their length. Named judge maps are matched by judge even when their length happens to equal the number of comparisons. Missing DIF identifiers remain as separate analysis units, conflicting external factor columns are refused, and scores outside R’s integer range no longer become missing on coercion. Paired-comparison equating now permits its documented descriptive two-object link while continuing to require three objects for drift tests.

• Common-item and common-object equating now label an unweighted descriptive shift when fewer than two common items or objects have usable variances. The functions no longer describe that fallback as precision-weighted or say that observations used in it were excluded from the shift.

• Every location axis is labelled at whole logits when the span allows, through one shared tick rule; the previous defaults could leave an axis extreme between labels. The person-item map begins and ends its proportion axis on labelled ticks, closes the axis corner, and drops its unused bottom margin.

• The Wright map and the person-item map mark the person and threshold means with dashed lines in their distributions’ colours.

• plot_person_fit() and plot_item_map() can display the standardised infit or outfit in place of the fit residual, under the same +/-2.5 band, and annotate the flagged count with its percentage. The person table reports infit_z alongside outfit_z.

• Input and selection boundaries are hardened package-wide, closing two further review rounds. Every estimator and the DIF family validate their arguments through shared checks: iteration caps, tolerances, class-interval counts, reference sample sizes, significance levels, adjustment methods, practical thresholds, linking minima, component and replication counts, and plot bins all reject fractional, non-finite, or out-of-range values instead of silently truncating. Selection can no longer alter the analysis silently: EFRM refuses misspelled or missized id, factor, and item inputs exactly as ordinary Rasch does, matrix input honours the items argument in both, multiple-choice keys refuse fractional option scores and duplicate item entries, replication counts must be whole and finite, explanatory threshold labels and predictor rows are validated, duplicate column names are refused in direct estimation, and unknown names in anchor, distractor, and object-set requests are errors rather than silent drops. Pooled MFRM items flow through the DIF follow-ups under the stricter item resolver.

• Nine defect families from an adversarial review are corrected. Wide-format many-facet data now scores factor columns by their labels, where reordered factor levels previously shifted item locations. Item, threshold, and anchor indices are validated: a fractional or unknown index errors instead of silently truncating to a different item, and the numeric fitting controls reject fractional interval counts and non-finite references. Duplicate named mappings – comparison anchors, item-set maps, judge-panel maps, and named judge factors – are refused instead of silently taking one of the conflicting values. A boundary comparison object keeps its extrapolated location when a dependence effect was dropped after its removal, reports count-weighted comparisons, stays inside the plotted range, and is refused by with its calibrated companions unaffected. Explanatory departure probabilities are withheld for items whose thresholds the calibration marks as weak, and the judge diagnostics report count-weighted comparison totals.

• A row that is dropped can no longer define the analysis it is dropped from. The paired-comparison response scale is derived after zero-count and unusable rows are removed, where a single zero-count row could take an otherwise identical model from three categories to six; margin levels and the presence of ties follow the kept rows for the same reason. An ordered factor’s declared levels remain the stated scale whatever the weights are, so an empty declared extreme is still refused as the identifiability question it is. A tie carries no margin, so its margin value no longer opens win and loss categories nothing was judged in, which had left both extremes empty and stopped the fit; and a response column that is entirely missing reports no usable comparisons rather than failing on an impossible vector length.

• A missing person identifier is unknown, not shared. The DIF family no longer reads missing identifiers as repeats – which declared a repeated-measures design and changed every test, and in dif_size() withheld every standard error and Wald test in the table – and the tailored bootstrap resamples each unidentified row as its own person rather than clustering them into one. A genuinely repeated design still withholds its Wald inference.

• Item and object banks are read through their labels. A factor column of locations, standard errors or maximum scores was passed to as.numeric(), which returns level codes, so a bank of -2.5, 0.25 and 4.0 equated as 1, 2 and 3. Numeric text remains accepted; invalid text and non-numeric classes are refused. An attached joint covariance now completes individual missing standard errors rather than doing so only when the bank omitted the whole column. Bank, predictor, key and anchor tables also require unique column names.

• A judging sequence must order the comparisons it describes. Values repeating within a judge left the order of those comparisons to the row order of the data, so the same data read in a different order carried different exposure and carry-over covariates; repeated and non-finite sequence values are now refused. A retained non-tie margin must be an ordered factor or a finite positive magnitude; zero denotes a tie. Margins on ties and excluded rows do not define the response scale. Logical, complex and Date columns are refused. A replication count must be real, since coercing a complex one discards its imaginary part in silence.

• A paired-comparison call must state one outcome. Supplying both winner and response fitted the response and ignored the winner, so changing every winner left the fit identical; the combination is now refused, in btl_explanatory() too. Every column-role argument names exactly one existing column before it is dereferenced, in the comparison, frame-adjusted comparison, explanatory comparison and many-facet entry points alike, and in the many-facet case through the wide entry as well as the long one, which needs at least one item column. Many-facet person, item, score, facet and person-factor roles must also be distinct.

• Paired-comparison equating uses the documented precision-weighted origin shift when two common objects have usable standard errors. Three remain necessary for object-level drift inference, but that inferential threshold no longer changes the descriptive link estimator.

• Named item-set, judge-panel, judge-factor and Wright-map assignments must cover their fitted units exactly. Missing entries no longer discard units, and extra entries no longer pass as unnoticed spelling errors. Empty or blank panels are refused. Repeated-measure person and time roles must be distinct.

• Simulation counts, parameters and seeds are read only as plain numeric values. Factors and classed or complex vectors can no longer be interpreted through their storage codes, and replicate seeds cannot overflow the integer range. A zero-effect BTL dependence specification is treated as no planted dependence.

• The low-level pcml() and pcml_pc() estimators now refuse negative, non-consecutive, constant and all-missing item-score columns. Their pair tables require at least two observed categories numbered from zero; negative values were previously omitted by tabulate() rather than diagnosed. The main rasch() entry point also refuses non-finite scores rather than treating them as ordinary non-numeric missing entries.

• MFRM and EFRM exports and app displays draw their observed residual scree without requesting the parallel reference that is unavailable for their virtual response-cell designs.

• A multiple-choice scoring table refuses a missing or blank item name in every key form – the option/score table, the item/key table and the named vector – as well as a missing or blank option, and a key refuses a blank value; both had scored their item zero throughout and then dropped it as constant under a misleading message. A saved analysis file’s schema is read as stored rather than coerced, so “1”, TRUE, 1.5 and a factor’s level code are no longer accepted as schema 1. An analysis file that declares a model type that is not one character – several, missing, or a factor whose integer code would select a branch by level order – is reported as unsupported instead of failing inside a length-one condition.

• External person factors supplied as a data frame are checked for constancy within person exactly as named columns are, instead of the first row’s value being kept. A frame group given by value no longer removes a person factor whose name matches one of the group labels, and a named panel map must give one stated panel for every judge.

• Set names are trimmed and validated in both frame families: a whitespace-only set name, or a blank set in the item-to-set map, is refused rather than fitted.

• An empty frame definition is refused rather than dropped: an item set or object set naming no identifier would have been fitted away, answering a different question from the one asked. A paired-comparison DIF call needs at least one judge factor.

• Administration patterns are built with explicit person-by-set dimensions. A single respondent – one person in a group, or a one-person fit – simplified to a vector that was then read transposed, so the design could name sets or facet cells the person never saw.

• A person factor may not take a name the fitted person table generates for itself: a factor called class_interval silently replaced the intervals every fit statistic is computed over, and one called theta stopped the fit. The generated names are reserved centrally. A frame design refuses repeated group columns, and the generated name for items no set lists is refused when a nominated set already uses it.

• The standalone HTML report names the estimator the fit used, as the fit summary table does, instead of always reporting pairwise conditional estimation.

• Names are carried as values rather than parsed out of labels: an item-set name containing the label separator keeps its items in the score curves; a predictor level no item or object carries is dropped before the design is built, where it added an all-zero column and made an identified model look rank-deficient; and a resolved comparison copy whose generated name already belongs to another object is refused with its magnitude withheld, rather than the two silently merging. A report table longer than its display limit carries the omission note as a caption, so it still renders as a table.

• EFRM score curves are keyed by the administration as well as the group: people in one group who sat different item sets have different maximum scores and different expected totals, and previously shared one curve. The table gains design and n_persons columns.

• A response style redraws from the probabilities the response was drawn under, so planted local dependence survives it, and a style of zero strength or zero prevalence is no longer recorded as an active effect. In a chain of dependence pairs the middle item’s expectation now includes its own carry-over, where the residual it passed on otherwise carried the first pair’s shift as a systematic mean into the second.

• An object set aside at a response boundary keeps its predictor row, so an explanatory comparison model no longer fails when one object always wins or always loses; a predictor row for an object the comparisons never mention is still an error.

• plot_pcc() and plot_kidmap() refuse a person identifier that appears in more than one row, naming the rows, rather than silently drawing the first. Repeated identifiers are the ordinary case in stacked and racked longitudinal data. dif_posthoc() validates the repeated-measures identifier where it is described rather than failing later on length.

• frame_invariance() computes the covariance of the centred location differences as C{V1 + V2}C’, not C{V1 + V2}C. The centring matrix is not symmetric when the compared items differ in maximum score, so the standard errors, statistics, p values, rmse and ratio were wrong in that case: a polytomous item’s standard error was inflated (22% in a five-category example) and every dichotomous item’s deflated, flagging short items and hiding long ones. Equal maximum scores were unaffected.

• Derived fits keep the controls they were built from. lr_test()‘s rating-scale refit carries the reference sample size and the person factors, so both parameterisations’ item-trait statistics are on one scale and the refit can be used for follow-up analyses; the fourth step of tailored_analysis() carries the reference sample size its three siblings use; and a subtest total is no longer read as missing when it happens to equal a missing-data code, which had deleted complete responses and renumbered the categories that remained.

• chisq_detail() reports the rescaled per-interval components beside the standardised ones, so the detail and the item table reconcile under adjust_N, and gives the adjustment factor and the unadjusted total.

• Item and object drift are refused for explanatory calibrations, where a location is a function of its predictors: a drifted item is smeared over every item sharing its design cell and the standard errors belong to the design coefficients. The dimensionality magnitude is refused for the same reason, since its subtest refit frees every superitem. An explanatory comparison design now centres on the objects actually calibrated, so its locations sit on the model’s own origin.

• An inestimable DIF analysis is refused rather than returned malformed, plot_frames() draws without intervals when the units carry no standard error instead of failing on an empty range, the frame-invariance summary counts item comparisons and items separately, and the observed points of the item displays follow the fit’s own per-item class intervals, so a graphical fit check shows the intervals of the test it illustrates.

• Simulation plants what it records, in three further cases: speededness needs a not-reached tail, a dependence source may not be regenerated as a later target, and a bias planted on a rater who answers at random is refused. Frame group units follow their group labels rather than the sorted level order, which attached the wrong unit to each group from ten groups upward. sim_apply() requires one atomic scalar per replicate.

• Reports and exports check their arguments before they write: the output path, title, plot dimensions and resolution are validated first, so a bad size can no longer leave a populated folder that reads as a complete export. A plot archive is written fresh rather than appended to, and the report closes only the devices it opened. Report tables state how many rows were omitted and print small probabilities in the package’s own vocabulary rather than as an impossible zero.

• compare_fits() treats presentation as data and the position covariate as a model term: paired-comparison fits differing only in which object was presented first, or in the judging sequence, are no longer reported as the same data, while a plain fit and a position-effect fit of the same comparisons still are.

• Four procedures no longer relax an explanatory restriction in silence. lr_test() refuses an explanatory fit, whose rating re-parameterisation would drop the design and leave the two models not nested; tailored_analysis() refuses one, because the tailored recalibration would differ from the original by the design as well as the tailoring; and btl_dif() refuses an explanatory comparison fit, whose resolved copies would either be forced equal by the design or estimated without it. The parallel scree reference now analyses each simulated draw under the model that was fitted, where an explanatory calibration was previously compared against an unrestricted refit.

• Simulation plants what it records. An item named as the second element of several dependence pairs now carries every one of them, where a later pair regenerated the item and erased the earlier dependence; a dichotomous item cannot be disordered, so the request is refused with a warning instead of recorded as planted; and a set or group unit ratio must be 1 when there is only one set or group, since a ratio between frames cannot be planted in a single frame.

• A failed plot export no longer closes the caller’s graphics device: it closes only a device it opened itself. Item names that sanitise to the same file stem now keep separate files, where the later plot silently overwrote the earlier and the export still looked complete.

• plot_pcc() draws the person characteristic curve from the fitted model’s own expectations. For a polytomous, many-facet, or explanatory fit it previously drew a dichotomous logistic curve, which ignores the thresholds and the frame’s units; the curve is now the expected proportion of the maximum score across the fitted item locations, and a dichotomous fit with a common discrimination keeps its exact logistic form.

• Selection can no longer alter a multiplicity family in silence. Duplicate items, objects, contrast names, and contrast cells are refused, because a repeated hypothesis quietly changes the Holm adjustment; a grouping given to the DIF family must name fitted factors or supply one value per person, rather than being recycled into a grouping the fit never contained; dimensionality subsets must be free of duplicates and disjoint; and a person-factor frame must carry unique, non-empty names whichever input branch assembles it.

• Sustained adversarial review closed the remaining input and display boundaries. Role selection is unambiguous in rasch() and rasch_efrm() alike: a vector as long as the person count is read by value even when its labels collide with column names, and a vector whose values match a data column exactly is refused, with items named as the resolution, where an item whose responses happened to agree with a role vector was previously dropped from the fit without notice. A repeated-measures identifier must carry one entry per fitted row: a short vector was recycled, understating the sampling units and changing every test in the DIF table. A person with a missing or blank frame group is refused rather than expanded into rows missing from every set. Reshaping refuses missing or blank person identifiers and occasions and requires one existing column name for the person and occasion; paired comparisons, frame-adjusted comparisons, and many-facet data refuse blank objects, judges, panels, persons, items, and facet levels, since a whitespace label is not a level but was calibrated as one; and every stated flag must be TRUE or FALSE. Display controls are checked before anything is drawn: an evaluation grid needs at least two finite locations, class intervals and bins must be whole numbers, limits must be two finite ascending values, label sizes and colour caps must be positive and finite, single-person, single-item, and single-facet displays take exactly one name, and limits admitting no thresholds report an empty range instead of drawing an empty panel. Batch plot exports and the HTML report name the plots they could not draw, where a file was previously written with the failures omitted; a batch in which nothing could be drawn is now an error rather than a returned path to an archive that was never created, and the archive is confirmed on disk before the path is returned. Export device dimensions are validated before any device is opened.

• The observed points of a paired-comparison ICC no longer abort the display when every comparator falls below the informativeness threshold; the model curve and the omission note draw on their own.

• A package-wide sweep for reporting-table misalignments found and corrected three defects. A numeric anchor index now resolves against the data as supplied, where previously a dropped constant item shifted the anchor to the wrong item. The class-interval detail refuses an item whose only responders are extreme, instead of failing obscurely and aborting exports. dependence_magnitude() withholds its standard error and probability when the resolved thresholds are weakly identified, reporting the magnitude descriptively, as the thresholds themselves already were.

• An undefeated or winless comparison object is now reported in the object table at an extrapolated location, its score moved half a point inside the boundary against the calibrated scale, with extreme = TRUE and its standard error withheld – the reporting practice already used for extreme person measures. The row takes no part in estimation, inference, or equating. Validated against sirt::btm: identical likelihood to machine precision on clean replicates, with the boundary policies agreeing in direction.

• The heaviest examples are smaller, and CRAN runs fewer scenario test blocks, keeping the check well inside the incoming pretest budget.

rasch 1.12.0

Models and inference

• wright_map() sends fitted person and item estimates to WrightMap. It supports several person distributions and the item-panel layout introduced in WrightMap 1.5, including the person-group and item-set structure of EFRM fits.

• rasch_explanatory() fits the linear logistic test model and linear partial credit model from continuous, categorical or ordinal item- or threshold-level predictors. Formulae may include selected interactions. explanatory_test() compares the restrictions with a free calibration using the Kent adjustment; explanatory_diagnostics() and relax_explanatory() support fixed item and threshold departures. Refitted departures propagate to item and person estimates and are retained through item deletion, DIF splitting, superitem construction and response-dependence resolution. Keyed option responses remain available after item deletion, splitting and fixed-departure refits.

• btl_explanatory() applies a fixed explanatory design to object locations in dichotomous or ordered comparative judgements. It supports the same model comparison, Holm-adjusted diagnostics and fixed departures while retaining the nominated ordered-response threshold structure.

• A worked case study on the documentation site uses the verbal aggression data to develop and check an explanatory partial credit model.

• explanatory_test() now places the Kent-calibrated probability in both p and p_kent. The unscaled composite-likelihood probability is named p_naive so it cannot be mistaken for the inferential result. The table also reports calibration R-squared, with an adjusted counterpart whose null expectation is near zero, against the free threshold or object calibration.

• Pairwise conditional calibrations now use the remaining Newton move as a second convergence check. This prevents numerical false refusals at large sample sizes without changing the estimates.

• Holm adjustment for item-fit statistics now excludes items whose tests are unavailable. Their probabilities remain NA.

• Sparse-unit safeguards now use the sampling units that inform each test. MFRM interaction tests use the least-supported item-by-level cell; EFRM unit tests require adequate persons on every group or set link; BTL-EFRM judge bootstraps require adequate effective judges in every panel or link; and frame-invariance tests exclude weak frame calibrations.

• BTL-EFRM judge bootstraps now distinguish refit errors from non-convergence and report the underlying worker error when parallel refits fail.

• rasch_mfrm() supports several facets and an optional item-by-facet interaction. Omnibus and cell follow-up tests use the fitted joint covariance.

• rasch_efrm() supports crossed person-group factors and reports their GLS factorial decomposition. Set-unit linking uses a finite-grid semiparametric likelihood, with a separate nuisance distribution for each observed person group. Hybrid standard errors retain the joint uncertainty of the within-frame calibration and set link; full person-bootstrap inference remains available. The convergence flag covers both estimation stages, and non-converged links are excluded from bootstrap covariance calculations. EFRM data require one response row per person.

• The repeated semiparametric linking calculations in the EFRM bootstrap now use a compiled numerical kernel. Bootstrap replicates can also be distributed over a reproducible, cross-platform worker cluster. The Shiny application runs EFRM fits in a background process, defaults to four workers where the system permits, records the bootstrap seed, reports progress and permits the fit to be cancelled without retaining a partial result.

• BTL-EFRM judge bootstraps likewise default to four workers where available. A fixed seed gives the same result for any worker count. The application runs these fits in the background and supports progress reporting and cancellation.

• frame_invariance() compares item locations and discrimination across separately calibrated frames. The conditional method tests locations and reports discrimination descriptively. The whole-person bootstrap within person group preserves item-set response patterns and provides inference for both, with one combined Holm family.

• MFRM and EFRM summaries report item estimates separately from the item-by-facet or item-by-frame response cells used in estimation. Coefficient alpha is not reported for the expanded response-cell matrix. EFRM DIF tests pool residual evidence by item and exclude the person factors that define the frames.

• btl(), btl_dif() and btl_efrm() add ordered paired comparisons, judge-clustered inference, judge-factor DIF, linked object sets and judge panels. Paired-comparison diagnostics now include equating, transitivity, residual dimensions, design information and adaptive pair selection.

• Carry-over probabilities are withheld below 30 judges. btl_equate() uses Welch–Satterthwaite degrees of freedom when fitted calibrations have a finite number of judge clusters. Conditional BTL-EFRM unit probabilities are withheld; the application defaults to the judge bootstrap. BTL-EFRM judge bootstraps use finite-judge references, whereas its independent-outcome parametric bootstrap uses normal and chi-square references.

Differential item functioning

• Confirmatory multiplicity defaults are now consistently Holm familywise adjustments across item fit, DIF, equating and the application. BH remains available where false-discovery-rate screening is explicitly requested. BTL DIF uses HC3 covariance for unequal judge workloads and withholds omnibus probabilities below eight judges or eight effective judges in a factor cell.

• Item-fit documentation now distinguishes the principal item-trait test from the supplementary class-interval ANOVA and notes the limits of both in short administrations. HC3 was evaluated for item fit and was not adopted.

• dif_anova() fits several person factors jointly using Type II sums of squares. Repeated measurements use the person as the sampling unit and separate between- and within-person error strata. Multiplicity adjustment covers the complete family of uniform and non-uniform DIF tests rather than treating each term as a separate family; btl_dif() follows the same rule. Uniform between-person terms now use HC3 covariance. Class-interval interactions retain the residual-ANOVA reference used for non-uniform DIF.

• dif_contrasts() and dif_posthoc() provide planned and post-hoc logit contrasts, including simple effects and difference-in-differences for interactions. MFRM follow-ups pool the fitted facet cells of an underlying item; resolved EFRM follow-ups are withheld because an ordinary split would discard the frame units. The residual-mean Tukey table has been removed from dif_anova(); dif_posthoc() is the supported follow-up for multilevel terms.

• Repeated-measures DIF follow-ups use the full design-cell weights in their person-level tests. Reported resolved estimates and probabilities therefore address the same marginal contrast when nuisance factors are imbalanced.

• dif_size() reports resolved pairwise logit differences. Dichotomous items receive the itemwise ETS A/B/C classification. Polytomous items report the PCM signed expected-score area descriptively, without importing an incompatible score-metric classification.

• resolve_dif() splits confirmed DIF items iteratively while retaining a minimum anchor set. Automatic splitting is restricted to uniform DIF; non-uniform DIF remains visible for item review. MFRM residuals can be pooled to their source items, and EFRM factors that do not define frames can be tested.

• btl_dif() retains anchors and fitted dependence terms in its resolution refit. Resolved pairwise inference is withheld unless each factor cell has at least eight effective judges; pairwise degrees of freedom use the two cells’ effective counts, and the pairwise table reports the raw and effective support for both cells. BTL-EFRM fits require a frame-specific analysis rather than the equal-unit resolution model.

Diagnostics and model changes

• Identification checks now cover item, facet, frame and paired-comparison graphs, rank, separation and sparse categories. Unidentified estimates are refused; identified but weak estimates are marked or have inference withheld.
• dependence_magnitude() uses the joint covariance of resolved thresholds. Equating tests require independent calibrations and the covariance of banked locations.
• spread_test() applies the binomial least-upper-bound only to superitems formed entirely from dichotomous components. It now distinguishes a point estimate below the bound from adjusted one-sided evidence of dependence. Its significance level and multiplicity adjustment are available in the application. The component structure is retained through subsequent item splits and removals.
• The tailored-analysis bootstrap resamples complete persons, including all rows of a repeated-measures record.
• drop_items(), resolve_frames(), DIF splitting and superitem construction refit the active model and update downstream item and person estimates. Refit specifications retain anchors, keyed scoring, threshold constraints, factors and frame-linking controls; a non-converged downstream calibration is not returned as a completed analysis.
• Classical whole-test statistics and the Guttman scalogram are withheld when an item is represented by several facet or frame response cells. They remain available for a one-cell-per-item reduction.
• MFRM characteristic and information curves now combine facet conditions administered to the same person. Distinct rating designs receive separate curves rather than being added into a test no person received.
• Automatic model comparisons are available for the main model families. Structural changes are accompanied by before-and-after item and person summaries.

Application and documentation

• The Shiny application uses responsive control and result columns, compact explainers for outputs and options, scalable plots and downloadable tables. Plot controls sit below the plot, and related item curves may be overlaid.
• Analyses can be saved as .rasch projects and reopened. Reports can be produced as self-contained HTML, Word or PDF documents; the R code for each displayed result is available in the application.
• The application covers the extended model suite, including model comparison, DIF follow-ups, frame-invariance checks and refitted structural changes.
• The manuals and vignettes have been revised to state the fitted models, estimands, identification requirements and uncertainty methods directly.
• The shipped EFRM and BTL-EFRM case studies now use the current linking and uncertainty methods.
• plot_scree() and plot_btl_scree() label their component axes at whole components only, instead of overprinting the default axis.

rasch 1.11.7

• print() preserves the reference distribution used by saved BTL fits: current and transitional results are labelled t, while older results without cluster degrees of freedom retain their original z label.

rasch 1.11.6

• BTL print methods read both current and earlier dependence-statistic columns. Current clustered statistics are labelled t, in accordance with their t reference distribution.

rasch 1.11.5

• Ordered paired-comparison margins must be numeric or an ordered factor; unordered categorical margins are rejected.
• Invalid graded responses and malformed secondary-dimension specifications now produce errors rather than being coerced or dropped.
• Clustered dependence and position statistics are labelled t.

rasch 1.11.4

• Score validation reads factor labels and rejects non-integer, non-numeric, or non-finite values.
• Ordered BTL responses require an ordered factor or integer scores.
• Clustered dependence and position tests use a t reference with G - 1 degrees of freedom.
• simulate_rasch() validates secondary-trait correlations and item sets.

rasch 1.11.3

• All estimators share the same integer-score validation.
• item_moments() uses a log-sum-exp calculation for wide category ranges.
• BTL object-separation reliability is withheld when the clustered covariance is rank deficient.
• btl_efrm() requires each judge to belong to one panel.
• Undefined reliability and omnibus statistics are reported as NA.
• Secondary-trait simulation retains its requested mean and standard deviation, and MLE score-table calculations use the correct common-unit score equation.

rasch 1.11.2

• The BTL dimensionality reference simulates count-weighted data at the unordered-pair level and returns the leading strength from each replicate.
• Judge-bootstrap EFRM inference requires more than one judge per panel and notes panels with fewer than five judges.
• BTL fits report when the number of judge clusters cannot support a full-rank clustered covariance.

rasch 1.11.1

• BTL dimensionality reference draws now reproduce count-weighted binomial or multinomial sampling.
• Pairwise chi-square degrees of freedom include position and dependence parameters; untestable designs return NA.
• Judge-clustered covariance uses the CR1 small-sample factor.
• btl_efrm(se_method = "judge_bootstrap") resamples judges within panels and refits both stages.
• The interpretation of btl_next_pairs() as a ranking rule is stated in its documentation.

rasch 1.11.0

• Threshold and item-location covariance is transformed consistently after recentering items with different maximum scores.
• Warm WLE uses the common-discrimination score equation in which the common discrimination cancels.
• Untestable item-trait statistics return NA, and the total test includes testable items only.
• Equating drift tests account for the estimated origin shift and joint covariance of common-item locations.
• Judge-clustered inference requires more than one judge and reports a caution for fewer than ten clusters.
• btl_dif() uses the judge as the sampling unit and count-weighted opponent bands.
• rasch() and btl() warn when estimation has not converged.

rasch 1.10.3

• rasch() reports unknown id, factors, and items columns as errors.
• Fractional scores are rejected rather than truncated.
• MFRM rows with missing design identifiers are omitted with a note.
• equate_tests() excludes common items without usable locations or standard errors from weighted linking and drift inference.
• report_html() escapes data-derived labels and notes.

rasch 1.10.2

• The Shiny application adds consistent hover labels to scalograms, residual heatmaps, equating plots, tailored analysis, and paired-comparison displays.

rasch 1.10.1

• Person- and item-fit plots in the Shiny application show identifiers, locations, and fit residuals on hover.

rasch 1.10.0

• compare_fits() adds composite-likelihood AIC and BIC based on the Godambe effective parameter count.
• Model comparison now covers BTL position, threshold, and dependence specifications. The Shiny comparison page supports these fits.
• Optional cross-package tests compare results with eRm, sirt, and psychotools.

rasch 1.9.3

• Pairwise estimation now checks that the observed item graph, together with any anchors, identifies a common scale.
• Thresholds adjoining critically sparse categories are marked weak. Their threshold and item-location standard errors are withheld.

rasch 1.9.2

• Case study: inst/casestudies/party_blocs_crisis.R applies btl_efrm() to the Tuebingen 2009 party-preference data.
• Sets without stable panel-ratio information are excluded from the unit reconciliation and refitted at the reconciled panel units.
• Boundary-unstable bootstrap parameters receive NA standard errors with the number of boundary replicates reported.
• BTL-EFRM convergence is assessed by the gradient per comparison.

rasch 1.9.1

• The Shiny Frames page supports BTL-EFRM panel and object-set definitions, unit tables, frame fit, and unit plots.

rasch 1.9.0

• btl_efrm() fits the paired-comparison extension of the extended frame of reference model, with panel units, object-set units, and set origins.
• Bootstrap standard errors refit both stages. Conditional standard errors remain available for descriptive work.
• plot_btl_units() and simulate_btl_efrm() support display and simulation.

rasch 1.8.0

• btl() adds anchored estimation and a first-position effect.
• btl_equate() and plot_btl_equate() provide common-object linking and drift tests for paired-comparison calibrations.
• btl_information(), plot_btl_targeting(), and btl_next_pairs() provide design information and greedy next-pair selection.
• Count-weighted BTL sandwich covariance now reproduces expanded-data covariance. Equating uses each calibration’s stored covariance.
• The Shiny Targeting and Equating pages support paired comparisons.

rasch 1.7.1

• Rasch simulation layers now retain all previously specified DIF, dimensionality, response-style, dependence, and guessing terms.
• PCM simulation uses item-specific threshold structures and applies stricter validation to DIF, guessing, and disordered-threshold specifications.
• BTL dimensionality references include fitted within-judge dependence.
• MFRM simulation and recovery use the recorded item and rater parameters consistently.
• The Simulate page links recovery output to the current generated dataset.
• A simulation vignette and package logo were added.

rasch 1.7.0

• Simulation functions add population-distribution controls, response styles, speededness, and MFRM halo effects.
• sim_replicate(), sim_recovery(), and plot_recovery() support repeated simulation and parameter-recovery summaries.
• The Shiny Simulate page exposes the population controls and recovery output.

rasch 1.6.1

• The Shiny application adds a Simulate page for the four data layouts and their model-departure controls.

rasch 1.6.0

• simulate_efrm() gains n_categories for partial credit items within frames, with planted thresholds recorded in the truth attribute.
• simulate_rasch(), simulate_btl(), simulate_mfrm(), and simulate_efrm() generate data from the package’s model families and can introduce nominated departures. Generating parameters are stored in the returned data.

rasch 1.5.1

• plot_btl_judge_map() now displays individual matchups. judge_pair_surprise() returns the corresponding residuals.

rasch 1.5.0

• judge_surprise() and plot_btl_judge_map() compare a judge’s object-level preferences with the consensus object scale. The display is available from the Shiny Judge fit page.

rasch 1.4.0

• btl_transitivity() reports circular triads and Kendall’s consistency coefficient for suitable paired-comparison designs.
• btl_dimensionality() decomposes the skew-symmetric residual preference matrix and compares its leading component with a model-based reference.
• New plots display BTL transitivity, scree, and residual maps.

rasch 1.3.1

• The package was renamed from its development name, rmt, to rasch. Result classes use the rasch_ prefix.

rasch 1.3.0

• btl() adds count-weighted exposure and carry-over effects, separation handling, and plot_btl_dependence().
• btl_dif() carries fitted dependence effects into its residual analysis and handles aggregated comparison counts.
• Mixed-design DIF terms are evaluated in their corresponding error strata. Repeated-person logit contrasts are provided by dif_contrasts().
• Residual parallel analysis uses data simulated from the fitted model.
• The Shiny application retains the settings used by each DIF analysis and applies consistent fit flags.

rasch 1.2.0

• plot_pca_biplot() draws the item loadings on the first two residual principal components on equal axes.
• residual_correlations() now also returns the adjusted-Q3 star_matrix and plot_resid_cor() can draw raw Q3 or adjusted Q3*.
• The Shiny trait and local-dependence pages pair tables with their plots and allow the original data to be restored after restructuring.
• dif_anova() is now the single DIF analysis-of-variance function. One factor is analysed one-way; several factors are fitted jointly. It supports repeated-measures and mixed designs.
• resolve_dif() resolves DIF iteratively by item splitting.

rasch 1.0.0

First stable release.

Models

• rasch() fits dichotomous, partial credit, and rating scale models by pairwise conditional maximum likelihood, with Warm WLE person estimates.
• rasch_mfrm() fits additive and item-by-facet many-facet models.
• rasch_efrm() fits the extended frame of reference model.
• btl() fits dichotomous and ordered paired-comparison models.

Diagnostics

• Item and person fit, category functioning, targeting, reliability, test information, residual dimensionality, and local dependence.
• DIF analysis, item splitting, tailored analysis, common-item equating, and anchored calibration.
• BTL judge fit, invariance, and paired-comparison diagnostics.

Display and reporting

• Base-graphics functions cover item, person, threshold, targeting, information, residual, and paired-comparison displays.
• fit_summary_table() and targeting_table() return the headline statistics; save_outputs() and report_html() export results.
• run_app() launches the Shiny interface and shows the R call corresponding to each analysis.