Changelog

Source: NEWS.md

nomologR (development version)

• The content-review reader is tested against contentvalidR 0.10.0 output too (#53). Its five handoffs are identical to 0.9.0’s apart from the producer version and date, and the reader needed no change.

nomologR 0.9.0

nomologR 0.9.0 is the last minor release before 1.0.0, the stable release planned jointly with contentvalidR (#53). It carries the 1.0 scope selected so far (#113); what is added before the release candidate on 2026-10-17 is recorded there. It is about how the package presents itself and what it teaches:

• Output that reads well. Every print() and summary() is redesigned (#89): aligned tables with the columns that matter, explanations in full, APA-style numbers, and one flag wording (“review”, “concern”) in the console, plots, and reports.
• A record of how practice changed. nomo_methods() gives the year each method entered the literature and, for a historical method, the contemporary methods that took over its question. The research-basis article draws them into a timeline.
• Teaching, and the bridge from content review. “Teaching with nomologR” collects exercises whose answers are known from the population model. “From content review to empirical screening” carries a contentvalidR review’s items through the screen, on the joint walkthrough’s data, now shipped as nomo_demo_walkthrough (#60).
• Manuscript tables for validity evidence. nomo_apa_table() formats nomo_validity() results: each construct pair’s latent correlation with its interval beside HTMT2, and a table of AVE.
• The 1.0 contract. Every help page names the fields of the object it returns (#114).

No computed estimate changes. Compared value by value with 0.3.0 across 22 analyses spanning every stage, every estimate is identical. What differs:

• With two or more declared scales, nomo_screen() reviews each item within its own scale: its item-rest correlation against the rest of that scale (scale, scale_item_rest_r, scale_item_rest_n), and its negative inter-item correlations only with items of the same scale (scale_negative_interitem_n). Which items are flagged can change. corrected_item_rest_r and negative_interitem_n are unchanged.
• In a multi-group nomo_validity() result, latent_correlations$block names the groups by label rather than by number.
• A CFA that is not identified (negative degrees of freedom) is logged as a concern rather than a review.
• Wording: counts agree with their nouns, some decision-log observations read differently, and the invariance summary lists the largest diagnostics rather than the first.

Breaking changes.

• nomo_missing() takes the fitted model as fit rather than x (#114). Calls that pass the model first without naming it are unaffected. The function was marked experimental, so there is no deprecation period.
• nomo_screen() checks scales when effort = FALSE too, since the within-scale correlations use it. A scales that names items not being screened now stops with an error instead of being ignored.
• nomo_report(apa_tables = TRUE) adds the validity tables, and when the run has a validity stage, its construct-pair table takes the place of the CFA factor correlations.
• An invalid choice stops with a message that names the argument, such as “method must be one of …”, instead of base R’s “‘arg’ should be one of”. Only code that matches the old message text would notice.

nomologR 0.3.0 is still in CRAN’s queue for new submissions (#39); until CRAN accepts it, install from R-universe or GitHub.

• With two or more declared scales, the item audit reviews negative inter-item correlations only within a scale (#113), as it does item-rest correlations. Two items of one scale should correlate positively, so a negative pair there is a keying or wording clue. Items of different constructs need not correlate at all, and a correlation near zero is negative about half the time. In nomo_demo_network, social desirability is uncorrelated with the other two constructs by design, and its near-zero negative correlations with their items had flagged all eight Agency and Persistence items. relationship_summary gains scale_negative_interitem_n. negative_interitem_n and inter_item_correlations are unchanged. Negative correlations between scales are logged as information (negative_pairs_between_scales). The log entry for negative pairs now agrees in number: “1 estimable inter-item correlation is negative”.
• nomo_apa_table() formats a nomo_validity() result (#113). type = "discriminant", the default, gives each pair of constructs one row: its latent correlation with a 95% confidence interval (Rönkkö & Cho, 2022), beside HTMT2 (Roemer et al., 2021) and HTMT (Henseler et al., 2015) when they were computed. type = "convergent" gives each construct’s number of indicators and AVE. The pair table is deliberately not the Fornell-Larcker matrix with the square root of AVE on its diagonal, the comparison that nomo_methods() records as historical because it often misses constructs that are not distinct. Correlations and AVE lose their leading zero; the ratios, which can exceed 1, keep it. A multi-group model’s tables add a Group column. nomo_report(apa_tables = TRUE) adds both tables after reliability, and the pair table takes the place of the CFA factor correlations, which it repeats.
• In a multi-group nomo_validity() result, latent_correlations names its groups (block) by their labels, as ave does, rather than by number. lavaan::standardizedSolution() numbers the groups, so the two tables of one result had disagreed, for example “1” and “2” against “Pasteur” and “Grant-White”.
• When two or more scales are declared, through scales or a contentvalidR handoff, the item audit reads each item’s corrected item-rest correlation within its own scale (#113). An item is scored on its scale’s total, so the rest of that scale is what it has to agree with (Nunnally & Bernstein, 1994; Clark & Watson, 2019). Pooled across constructs, the items of a distinct scale look weak: in nomo_demo_network, the three social-desirability items correlate .15 to .17 with the rest of all eleven items and were flagged, and .52 to .57 with the rest of their own scale. The relationship summary gains scale, scale_item_rest_r, and scale_item_rest_n, and corrected_item_rest_r keeps the pooled value. The flag, the decision-log entry (which names the scale), the summary table, and the item-rest plot use the within-scale value. A guided run’s per-scale audits were already within-scale; the instrument-wide audit that settings$screen$effort = TRUE adds now agrees with them. scales is checked when effort = FALSE too, since it now matters there.
• The item-rest plot draws its values in a column past the bars and the reference line (#89). A flagged item sits just below the reference, so its value, drawn beside the bar, had crossed the dashed line.
• nomo_missing() takes the fitted model as fit, as nomo_reliability(), nomo_validity(), nomo_scores(), and nomo_hierarchical() do (#114). It had been x. The function is experimental until 1.0.0, so the name changes without a deprecation period. Calls that pass the model first, without naming it, are unaffected.
• A guided run or validity result with a single construct no longer warns “Unknown or uninitialised column” when printed. With one construct there are no pairs, and the pair tables have no columns; the ordering and key-evidence code read them with $, which warns on a tibble.
• A new article, “Teaching with nomologR”, collects exercises built on the teaching datasets’ known answers. Students find the weak and cross-loading items, check whether the factor-retention criteria agree, compare alpha with omega, handle ordered categories, and tell a structural path from a correlation (including why a path that is not significant does not show that it is negligible). They also find the non-invariant intercept, see where content review and the empirical screen disagree, and see why an item-total correlation depends on which total it is computed against. Each answer is stated from the population model, not from one run. The article also suggests how to use teaching and research modes, decision logs, reports, and nomo_methods() in a course.
• nomo_methods() records how practice changed (#113). introduced is the year a method entered the literature, taken from the publication that introduced it, and given only when the registry cites that publication (46 of 89 methods so far). contemporary_practice names, for a historical method, the contemporary methods that now answer its question: parallel analysis for the eigenvalue-greater-than-one rule, omega for coefficient alpha, HTMT2 and latent-correlation intervals for the Fornell-Larcker comparison, FIML for listwise deletion. The research-basis article gains a section, “How practice changed”, with a timeline from 1937 to 2022 and a table of what took over from each historical method, both drawn from the registry. The KMO entry now cites Kaiser (1970), which introduced the index. The methods table in nomo_report() shows the year each method was introduced, so an archived report carries the same record.
• Every function’s help page now names the fields of the object it returns (#114). The stability policy covers “the documented fields of the objects they return”, but the Value sections had described contents in prose without naming fields, so most fields users read, such as standardized_loadings and fit_evidence on a nomo_cfa, were not named in the contract. Each Value section now lists the fields to read, one line each, and says that the remaining fields (the call, settings, and engine intermediates) are not part of the stable interface. This changes documentation only. ?nomo_efa no longer explains factor_count with an internal milestone label.
• A new article, “From content review to empirical screening”, carries the items a contentvalidR review carried forward through the empirical screen, on the shared teaching data of the joint walkthrough (#60, #53). It shows the two stages disagreeing, which is the reason the pair exists. EF4 passes content review and carries almost no common variance. EF3 is flagged empirically but is the only item covering part of the domain. TF4 loads on both facets, and TF6 is not invariant across cohorts. The article builds without contentvalidR installed.
• New teaching data nomo_demo_walkthrough and nomo_demo_walkthrough_items: 400 simulated responses to twelve Study Persistence items in two cohorts, with each item’s built-in role (#60). They are contentvalidR’s walkthrough files, copied unchanged from its v0.9.0 release with their git blob hashes. contentvalidR’s generator is in data-raw/walkthrough/.
• The item audit explains a negative item-rest correlation for an item declared reverse-keyed, by a contentvalidR handoff or by reverse (#60). It says the sign is what such an item shows before recoding, and gives the item-rest correlation computed on a copy recoded as declared. If the recoded correlation is still negative, it says the keying does not explain it. nomo_run() passes a handoff’s declared keying to each scale’s audit for this. The data, the returned correlation, and the flag are unchanged; nomologR never recodes data.
• The measurement-invariance summary lists the ten largest equality-constraint diagnostics across all levels. It had listed the first ten in level order, so the metric level filled the table, and a scalar-level strain such as an intercept never appeared. On nomo_demo_network, the known ag3 intercept (score 61) had been hidden behind a metric loading (score 5).
• Three edge cases read better (#89). The CFA summary of a model that did not converge says “No fit index is available: the model did not converge.” It had listed the reference values with nothing beside them. The item-audit summary shows a column with one observed value as “constant” rather than “binary”. The returned item_type is unchanged. The missing-data print says why a strategy was not fitted, such as “FIML: Not fitted: no modeled variable has missing values.” for complete data, and shows its N as “-” rather than NA.
• An invalid choice names the argument it was given for (#89): nomo_scores(fit, method = "eap") now stops with “method must be one of”sum”, “mean”, “regression”, or “bartlett”, not “eap”.” instead of base R’s “‘arg’ should be one of …”. This applies to every argument with a fixed set of values, such as type in plot(), nomo_table(), and nomo_apa_table(). Partial matching (type = "load") works as before.
• nomo_cfa() says when its fit indices cannot test the model (#89). A one-factor model with three indicators is just identified (df = 0): it reproduces the covariances by construction, and its print had shown “RMSEA 0.000 | SRMR 0.000” like a well-fitting model. A model with negative degrees of freedom, such as a factor with two indicators, is not identified. The print now reads “Fit: not testable (df = 0, just identified)”, the summary explains why, and the guided run’s key evidence says the same. The decision log’s degrees-of-freedom entry, which was already recorded, now names the case, and a model that is not identified is a concern rather than a review. Warnings that lavaan raises while nomologR reads the results, such as “Could not compute standard errors”, are kept with the model’s engine warnings instead of printing at the console. The “NaNs produced” warnings that lavaan::fitMeasures() raises for negative degrees of freedom are no longer shown.
• Counts in printed output agree with their nouns (#89): “2 factors”, “1 pair”, “1 more degree of freedom”, rather than “2 factor(s)”. This covers the printed counts in scores, reliability, validity, hypotheses, and partial releases, and the notes in comparisons and bifactor models. Counted phrases in decision-log observations and report summaries agree too, with their verbs: “1 indicator was declared ordered”, “10 of 12 reviewed items were carried”.
• Plot legends use the display flag wording too (#89). The CFA loading and fit plots and the AVE and HTMT plots label their points “none”, “review”, or “concern” under “Flag”. The loading plot had shown the recorded values KEEP and REVIEW under “Review”, and the validity plots info and review under “Signal”. The item evidence map calls an informational signal a “note”, which is how the console shows it. Each flag level keeps the same point shape from plot to plot.
• The redesigned console output extends to the guided workflow: print() and summary() for nomo_run() (#89). The print opens with the run’s status, mode, sample design, and scales, then lists what the run found so far as “Key evidence” (item flags, suggested factors, CFA fit, the omega range, validity and invariance results). A request repeated for several scales, such as a factor-count decision for each of three scales, is shown once with the scales it applies to, rather than once per scale. Research mode leaves out the teaching-mode reason, options, and consequences, and keeps what each scale showed and an example. The summary is about a third of its former length: a stage table, recorded decisions as bullets, the component recipe, and a count of the methods used with the primary method of each stage, pointing to nomo_methods() and nomo_table() for the full entries. The gallery in dev/output-gallery.R now reports no presentation faults.
• nomo_report() tables are easier to read (#89). Column headings are words, not code names: “Omega CI lower” for omega_ci_lower, “Change in CFI” for delta_cfi, “Corrected item-rest r” for corrected_item_rest_r. Flags use the console’s wording (“review”, “concern”). A missing or empty cell is an em dash rather than NA or nothing, and TRUE/FALSE is yes/no. Item, scale, and factor names are left as written. The values are otherwise unchanged, and nomo_table() still returns the code’s column names and flag values. The citations table gives each package’s reference as text. It had held R’s whole print(citation()) output, including a BibTeX entry whose @Manual{ pandoc turned into a stray citation, and LaTeX such as \texttt{semTools}.
• The redesigned console output extends to comparisons, invariance, partial releases, hypotheses, networks, splits, and the notes of nomo_apa_table() (#89).
  • Wide tables are split: model fit apart from information criteria, difference tests apart from changes in fit, and fit by level apart from changes between levels.
  • Levels that failed or raised warnings are listed with what went wrong.
  • Measurement evidence in a comparison reads one column per model.
  • A hypothesis table states once that every relation is on the standardized scale, rather than repeating it in each row.
• The redesigned console output extends to reliability, validity, scores, missing-data sensitivity, and hierarchical models (#89). Tables replace wrapped tibbles. Severity-tagged notes print as review: and concern: bullets. Index names read as written in the literature (“omega hierarchical”, “ECV”). A column that holds one value throughout, such as a block column that always says “overall”, is left out. The note on correlational accuracy names the pair of factors whose score correlation strays furthest from the factor correlation. It had named one of the two, and which one depended on rounding, so it differed between platforms.
• The content-review reader is now tested against contentvalidR 0.8.0 and 0.9.0 output as well as 0.6.0 and 0.7.0 (#53). The added fixtures include a nine-expert panel whose carry decision changed between producers. Seven of nine experts rate item N7 relevant, which is what Lynn (1986) requires. contentvalidR 0.7.0 compared the I-CVI with a rounded .78 and held N7 back; 0.8.0 compares counts and carries it. The test confirms that nomo_screen() follows the decision the handoff records rather than inferring one from the producer version. The reader needed no change: 0.8.0 and 0.9.0 keep schema version 1 with the same fields and columns.
• The redesigned console output extends to the item audit, factor retention, and EFA (#89): print() and summary() for nomo_screen(), nomo_factors(), and nomo_efa(). The screen summary lists each flagged item with the decision log’s own explanation rather than internal metric names. The factor-retention summary prints the reason a criterion was not run, and the Bartlett result, in full rather than cut off in a table cell. The EFA summary no longer hides the item flags behind “2 more variables”.
• Plot titles, subtitles, and captions are wrapped to fit the plot (#89). ggplot2 does not wrap them, and ten plots had text long enough to run off the edge at 7 × 5 inches: four CFA plots, three EFA plots, two network plots, and the hierarchical variance plot. The CFA fit plot now draws CFI and TLI, which sit near 1, and RMSEA and SRMR, which sit near 0, in separate panels with their own scales, and a legend with one entry is no longer drawn.
• Console output begins a redesign for readability (#89), starting with nomo_cfa()’s print() and summary(). Tables are aligned text with the columns that matter and no type rows. A flagged loading’s explanation is printed in full rather than cut off inside a table cell. Numbers have fixed decimals, p-values follow APA style, and prose wraps to the console width. Flags are shown in one wording across the package: nothing for no flag, then “review”, then “concern”. This is display only: returned objects, nomo_table() types, and decision-log values keep their values. dev/output-gallery.R regenerates every print, summary, and plot and reports the remaining presentation faults.
• nomo_cfa(), nomo_invariance(), and nomo_network() name any variable the model uses that the data lack, before fitting (#89): “model names variable(s) not found in data: agX. Check the spelling against names(data).” A misspelled item had reached lavaan, and the error named lavaan’s internal step (lavaan->lav_step02_options()) rather than the item. The more specific argument checks, such as for ordered, still come first, and model syntax that lavaan cannot parse is still reported by lavaan.
• nomo_table() works on every measurement-stage result: nomo_factors, nomo_efa, nomo_cfa, nomo_reliability, and nomo_validity (#89). They had no method, so the documented way to extract a table failed for five of the package’s main results. Each returns the tables its report section shows. nomo_report() now gets them through nomo_table(), so the two cannot drift apart. ?nomo_table lists every supported object and type, including nomo_screen and nomo_scores, which were missing from it.
• plot(<nomo_screen>, type = "responses") draws a histogram for a continuous item (#89). It drew a bar for every distinct value, which for continuous responses means one bar per respondent: an unreadable plot, with a percentage label on each bar. Categorical items keep their bars. When the selected items mix the two, the categorical items are drawn and the caption names the continuous ones and how to plot them.
• nomo_validity() lists each construct pair once in its print, summary, and report section, with the latent correlation and HTMT-family values in the same row (#89). Latent correlations name a pair in model order and HTMT names it the other way round, so the two never merged. Each pair appeared twice, once with its correlation but a signal of “unavailable” and once with its HTMT values, and the print counted twice as many pairs as there are. Pairs are now shown in model order, including in the HTMT plot. The returned tables and every computed value are unchanged.
• nomo_report() headings are rendered for every scale of a multi-scale run (#89). A heading that followed a plot, which is every scale after the first in the item audit and factor retention, plus “Careless responding”, was run into the plot’s paragraph. It appeared as literal ## Scale: ... text in HTML and Word, and the table of contents filed every scale’s tables under the first. The report tests had rendered only single-scale runs or had turned plots off; a two-scale report with plots is now tested in both formats.
• The reproducible-report article’s overwrite example shows the refusal message with its temporary directory as <report_dir>. It had printed the full temporary path of the computer that built the article, including a user name, and the path changed on every build. The message still comes from nomo_report(), and the example still shows that an existing report is refused unless overwrite = TRUE is given.

nomologR 0.3.0

nomologR 0.3.0 is the first release submitted to CRAN (#39), and its scope was kept deliberately lean for that reason (#38). It does three things:

• Connects content review to empirical validation. A handoff from contentvalidR now flows into nomo_screen() and nomo_run(), and the reasons for each item’s carry decision travel with it (#46).
• Folds the v0.2.1 tools into the guided workflow and its report. These are careless-responding screens, scores, missing-data sensitivity, and manuscript tables (#73).
• Starts the public interface’s stability promise. From this release the interface changes only after a deprecation period, under the policy written on ?nomologR (#74).

Test coverage is back to every executable line (#72). Reviewing each uncovered line found four defects, now fixed. Certification is recorded in #75. Distribution is R-universe until CRAN accepts the submission. The joint 1.0.0 release with contentvalidR is tracked in #53.

• CRAN preparation for the first submission (#39).

  • Spelling. DESCRIPTION declares Language: en-US, and the package’s text uses American spelling throughout. About 40 British spellings (analysed, modelled, behaviour) are changed in messages, help pages, articles, and NEWS; only wording changes, not code. spelling::spell_check_package() now finds nothing, with author names and technical terms listed in inst/WORDLIST.
  • URLs. urlchecker::url_check() finds all 131 URLs correct.
  • Build. .git and .gitignore are excluded from the build. A git worktree has a .git file, and a hidden file in the tarball is what CRAN returned contentvalidR 0.3.1 for.
  • Test time. Under CRAN conditions the tests took about 15 minutes on a slow Windows machine; they now take about 3. The 117 tests that took 2 seconds or more are skipped on CRAN: simulations, full guided runs, and rendered reports. Two stay on CRAN because they check that nomologR reproduces lavaan’s own estimates, so a lavaan change that alters them is caught there. Continuous integration still runs every test, and coverage is unchanged.
  • Examples. nomo_scores() and nomo_apa_table() gain runnable examples, so every exported function now has one.
  • Description. The DESCRIPTION cites the two works that frame the package, Cronbach and Meehl (1955) and Flake, Pek, and Hehman (2017), with their DOIs.
  • References. The research-basis article’s reference list gains the APA Publication Manual, which ?nomo_apa_table already cited. Its “Planned” notes now say that longitudinal invariance and multiple imputation are candidates for v0.4, not v0.3.

• nomo_report(apa_tables = TRUE) appends a Manuscript tables appendix (#73, last of three parts). It holds the nomo_apa_table() tables for the results a run holds: CFA loadings, fit, and factor correlations; reliability; and, when present, invariance and the network’s hypotheses and fit. They are numbered in the order the report presents them. A table that does not apply is left out rather than failing the report. The default, FALSE, leaves reports unchanged.

• nomo_run() can score the measurement model and compare missing-data strategies, and its report shows both (#73, second of three parts).

  • Scores. settings = list(scores = list(method = "sum")) scores the model with nomo_scores(). The researcher must name the method; a request without one, including scores = list(), is refused, because nomologR does not choose a scoring method.
  • Missing-data sensitivity. settings = list(missing = list()) compares strategies with nomo_missing() for the measurement model, and for the network too when one is requested.
  • When they run. Both run after convergent and discriminant evidence, so they are in view when the researcher decides whether to carry the model forward. Each matches the standalone call.
  • Not stages. Neither is a stage, so the stage table keeps its shape, and a run that requests neither is unchanged.
  • If they cannot be computed. Requested evidence that cannot be computed is recorded in the design log and the run continues, since no later stage depends on it.
  • Resuming. Both can be requested when resuming before the CFA, and are locked once computed.
  • Where they appear. Their evidence joins the component log, their methods are credited to the run, and the report gains “Scores” and “Missing-data sensitivity” sections.

• A written stability and deprecation policy, on the package help page (?nomologR) and in the README (#74). From 0.3.0, the first CRAN release, the public interface changes only after a deprecation period.

  • Covered: exported functions, documented arguments and defaults, documented fields of returned objects, nomo_table() types, and decision-log columns.
  • Breaking changes: any change that breaks the interface, including a changed default that changes results, is deprecated for at least one minor release first.
  • Experimental until 1.0.0: nomo_missing(), whose flagging rule may be refined, and the layout of nomo_apa_table() tables. Both are marked in their help pages.
  • Required by the joint release: #53 requires this policy for the joint 1.0 release with contentvalidR.

• nomo_run() computes careless-responding indices when settings = list(screen = list(effort = TRUE)) is given (#73, first of three parts).

  • Once, across the instrument. The indices describe a respondent across the whole instrument, and several cannot be computed within one scale. Even-odd consistency correlates across at least three scales, and psychometric pairs can span scales. So they are computed once, over every item in the run with the run’s scales, and the per-scale item audits are unchanged. The result matches a standalone nomo_screen(effort = TRUE) call exactly.
  • Keying. reverse, scale_range, pair_magnitude, and scales can be set alongside effort. When the scales came from a contentvalidR handoff that declares keying, that keying is used unless the settings give their own, and the design log records which applied.
  • Where it appears. The indices’ log rows join the run’s component log, and the methods are credited to the run. The report gains a “Careless responding” section after the item audits: a summary, one row per index with the rule a source states for it or “none stated”, and the evidence rows.
  • Refusals. An unreadable effort value is refused before any stage runs.

• nomo_screen() and nomo_run() accept a handoff from contentvalidR’s content_handoff(), so items move from content review to empirical screening with their reasons intact (#46). The interface is schema version 1, agreed with the contentvalidR maintainer and documented identically in both packages.

  • What is analyzed. nomo_screen(data, items = handoff) screens only the items content review carried. Every held-back item is listed in the decision log with its status and recommendation quoted in contentvalidR’s own words, and is never analyzed or reinstated.
  • Guided runs. nomo_run(data, scales = handoff) takes its scales from the review, and its design log records that item membership came from content review rather than from these data. A review with no construct mapping, such as an expert relevance panel, is refused by nomo_run() with the carried items listed, because a guided run needs scales and nomologR does not invent them. It can still be screened.
  • Keying. Declared keying and the response scale fill reverse and scale_range for the careless-responding indices, exactly as agreed. An undeclared item is never treated as forward keyed, a response scale is never inferred from the data, and arguments given in the call take precedence, with a note in the log.
  • Refusals. A carried item missing from the data is refused, never dropped. A handoff with a schema version this release does not read is refused, naming both package versions. Fields this release does not know are ignored, so contentvalidR can add fields within a schema version. An object content_handoff() could not have produced, such as keying missing for some items but not others, is refused as malformed.
  • Report. nomo_report() gains a content-review section for runs that came from a handoff, so the archive starts where the validity argument starts. Reports of other runs are unchanged.
  • Testing. contentvalidR is not a dependency. The reader is tested against genuine output from contentvalidR 0.6.0 and 0.7.0, stored in tests/testthat/fixtures with their generator and checksums. The guided-workflow article gains a runnable section built on the example handoff in inst/extdata.

• The research-basis article describes the careless-responding indices that shipped in 0.2.1. It had still listed them as planned.

• Test coverage is back to full executable-line coverage after v0.2.1 (#72). Reviewing each uncovered line found four defects, now fixed:

  • Network fit in nomo_missing(). For a nomo_network, the fit comparison listed every fit index as missing. A network’s fit evidence is one wide row, while a CFA’s is a long table, and only the long form was read. Every strategy’s chi-square, CFI, TLI, RMSEA, and SRMR now appear.
  • nomo_scores() with ordered indicators. The parallel-model test that unit weighting implies was written for continuous indicators. With ordered indicators it was fitted by maximum likelihood against a categorical model, failed, and reported only that the comparison “could not be computed”. It is now not run for ordered indicators, and the note says why.
  • RMSEA interval in the network fit table. The APA fit table for a nomo_network had an “RMSEA [90% CI]” column but printed no interval. The interval is now read from the fit, using the robust, scaled, or plain RMSEA that the evidence reports.
  • Long-string with no usable scale. The mean within-scale long-string returned NaN when no scale had two items. It now returns NA, as the inter-item SD already did.

• Lines that no input can reach were removed rather than tested. Examples are the empty-row checks in the careless-responding indices and the guards for fitted models that are always present. The remaining guards are tested against real fitted objects:

  • models that did not converge, have several groups, or were fitted from a covariance matrix;
  • single-indicator and cross-loaded factors;
  • small and degenerate response sets.

  Failures that only lavaan can produce are simulated.

nomologR 0.2.1

nomologR 0.2.1 completes the six workstreams that moved out of v0.2.0 at certification (#37). It also carries three workstreams opened during the cycle (#54, #56, #62). Together they cover what happens after a measurement model is established:

• scoring it, with the evidence for the scoring choice (#33);
• screening for careless responding (#34);
• checking whether results depend on missing-data handling (#32);
• judging factor-score quality (#56);
• carrying scores into a network honestly (#62);
• reporting in manuscript-ready form (#35), including from inside a thesis document (#40).

Where a published worked example exists, the package reproduces it exactly: Curran’s (2016) examples for the careless-responding indices, and Rodriguez, Reise, and Haviland’s (2016) MASC values for factor determinacy and construct replicability. Skrondal and Laake’s (2001) factor-score regression design and the missing-data comparison are checked against known population values. Every source cited is in the research-basis reference list, with its DOI checked against the registry. Certification is recorded in #70. Distribution remains R-universe; the first CRAN submission is targeted for v0.3.0 (#39).

• nomo_report() now works from inside a knitted R Markdown or Quarto document, such as a thesis chapter, with no workaround in the calling document (#40). A nested render shares knitr’s state with the document that started it, which previously stopped the report with a duplicate chunk label, and which silently dropped every figure from the report when the calling document set a graphics device its template did not expect. The report is now rendered with knitr’s default chunk options, which its own template then sets for itself, and the calling document’s chunk options and duplicate-label setting are restored afterwards. A report therefore renders the same from the console, a script, or a chunk, and rendering one does not change the document that asked for it.

• Restored full executable-line coverage by testing every nomo_compare() failure guard (#54). These branches handle a failure inside lavaan or semTools rather than a researcher-facing input, and several of them produce the explanation a researcher reads when a check cannot be run, such as why a nesting check was skipped or why a difference test is unavailable. Wording that has never executed can be wrong without anyone noticing, so those sentences are now asserted against real fitted objects.

• nomo_hierarchical() now reports factor determinacy and construct replicability for the general factor and each group factor, in a new factors table reachable with nomo_table(x, "factors") (#56). Both reproduce the published values for the MASC example in Rodriguez, Reise, and Haviland (2016) exactly. The two indices are the same quantity only when a construct is unidimensional: under a bifactor model, determinacy uses the whole reproduced correlation matrix while H sees one factor’s loadings alone, and Rodriguez et al. decline to prefer either, so both are reported and each is labeled with the question it answers. Gorsuch’s (1983) and Hancock and Mueller’s (2001) thresholds appear as their authors’ recommendations where a value falls below them, never as rules.

• Added nomo_scores(), which computes sum, mean, regression, or Bartlett scores from a fitted measurement model and reports what those scores are and are not (#33). Unit weighting is treated as the model it is: McNeish and Wolf

  2020. show that adding items assumes a parallel model, with equal unstandardized loadings and equal residual variances, so for "sum" and "mean" that constrained model is fitted and compared with the model supplied. Every method reports Grice’s (2001) three criteria, computed from the fitted model: validity, univocality, and correlational accuracy. Validity for regression scores is the factor determinacy coefficient reported by nomo_hierarchical(), since that method maximizes it. Correlational accuracy is the one to read before using scores in a later analysis: correlations among scores do not reproduce correlations among the factors, the discrepancy is substantial, and its direction depends on the method and the model rather than being a constant that could be corrected for, so it is reported and left visible. Gorsuch’s (1983) thresholds appear as his recommendations where a value falls below them, never as rules. The supplied data is never modified, and only the cases the model used are scored. Includes the Scoring a measurement model article.

• nomo_screen() gains effort = TRUE, which adds case-level indices of careless or insufficient-effort responding (#34): long-string, inter-item standard deviation, Mahalanobis distance, even-odd consistency, and psychometric antonym and synonym correlations, following Meade and Craig (2012), Huang et al. (2012), Curran (2016), and Marjanovic et al. (2015). Curran’s published worked examples are reproduced exactly and used as tests. The indices are reported side by side rather than combined, because they detect different failures: inter-item standard deviation detects random responding and gives a respondent who answers every item identically the best possible score, which is exactly the case long-string flags. The decision log reports that disagreement when it occurs. A case is flagged only where a source states a rule, with the source’s own qualification attached; the rest are reported without a flag. Inter-item standard deviation is averaged within scales, as Marjanovic et al. computed it, since one value across every item rates a respondent who is high on one construct and low on another as more erratic than a random responder. Within-person correlations over fewer than three pairs or scales are never reported, because with two every value is exactly +1 or -1. Reverse keying and the response range are declared, never inferred, and recoding happens on an internal copy only. Cases are flagged, never removed, and a screen that does not request these indices is unchanged. nomo_table() now supports nomo_screen objects.

• nomo_network() now discloses relationships estimated between observed variables (#62). When an endpoint is observed rather than latent, its measurement error enters unmodeled, and if it is a composite of several items (a sum, mean, or factor score) the relationship carries the discrepancy nomo_scores() reports as correlational accuracy. The network cannot tell a composite from a single measured variable, so it does not guess: it classifies each hypothesis by its endpoints and discloses observed ones in the decision log, for review when both ends are observed and for information when one is, saying plainly that a single measured variable is not a composite. The disclosure points to modeling the items as indicators, to lavaan::sam() (Rosseel & Loh, 2024), and, for a linear regression among factor scores, to the design Skrondal and Laake (2001) proved consistent: regression-method scores for the predictors and Bartlett scores for the outcome, each block scored from a measurement model of its own. The second condition is easy to miss. In reproducing their result, scoring both factors from one joint model biased the estimate by about +.10 in the same check that recovered the latent value with separate models. No correction is applied automatically.

• nomo_scores() and the scoring article now describe that design (#62). The note on correlational accuracy says that scores from one model containing every factor are not it, and the article applies it to its own data next to the two ways of getting it wrong: the same method for both blocks falls well short of the latent slope, and scoring both factors jointly overshoots it.

• nomo_reliability() gains ci_ncpus, which runs the bootstrap on several worker processes through lavaan’s snow backend (#42). The default, 1, keeps the bootstrap serial and unchanged. snow is used everywhere rather than forking, which is unavailable on Windows, so results are comparable across platforms. The draws depend on the worker count as well as the seed, so a result is reproducible for a given seed and worker count: both are recorded in the bootstrap status table and, newly, in the decision log that reaches reports, which previously carried neither. An unseeded bootstrap is flagged for review as not reproducible. Verified against an installed package: the same seed and worker count give identical intervals, and serial and two-worker runs differ by under .01, as the draws should. More workers are not always faster, since each must start and load its packages; in the evaluation recorded on the issue, eight workers were slower than four. Tests use at most two workers, following CRAN policy.

• Added nomo_apa_table(), which formats the evidence in a result as an APA 7 table ready for a thesis, dissertation, or manuscript (#35): standardized loadings, factor correlations, and model fit from nomo_cfa(); reliability from nomo_reliability(); the invariance sequence from nomo_invariance(); and hypothesis evidence and model fit from nomo_network(). Tables carry a bold number, an italic title, no vertical rules, and notes below in APA’s order, and they knit directly into R Markdown or Quarto. Leading zeros follow the statistic rather than its value, as APA 7 specifies: statistics that cannot exceed 1 (reliability, correlations, CFI, p) lose the zero, and those that can (TLI, RMSEA, SRMR, standardized loadings) keep it. A value that rounds to zero is never printed with a sign, and an estimate the model could not produce is shown as an em dash. The hypotheses table reports each estimate’s evidence against its prediction, and marks a hypothesis specified after the data were seen as exploratory. No cell reads pass or fail. Word output for nomo_report() follows separately.

• nomo_report() now writes Word documents: a file ending in .docx produces one, alongside the existing HTML report (#35). Pandoc drops raw HTML when it writes Word, and the report template wrote its tables, notes, and its interpretation contract as raw HTML, so a Word report would have lost all of them, including the statement that the report does not turn review references into pass/fail rules. Each part of the template now writes HTML for an HTML report and markdown for a Word one, and the HTML report is unchanged. The Word report was checked against the HTML one: the same 37 data tables with the same header rows in the same order, the interpretation contract, the figures, and no raw HTML or escaped entities in its text. Collapsible sections are shown expanded, and the document uses Word’s default styles. Section numbering is used where the installed rmarkdown supports it for Word.

• Added nomo_missing(), which refits a nomo_cfa or nomo_network under alternative missing-data strategies and reports whether the cases used, the estimates, fit, reliability, or theory evidence change (#32).

  • Strategies compared. Continuous indicators are compared under listwise deletion and FIML, with FIML as the reference. Ordered indicators, for which lavaan offers no FIML, are compared under listwise and pairwise deletion.
  • Assumptions and the flag. Each strategy is labeled with the mechanism it requires, following Enders and Bandalos (2001). A difference larger than half the reference standard error is flagged for review, following Schafer and Graham’s (2002) rule for when bias becomes practically important. The flag says that the difference estimates bias only if the data are missing at random and the model is correct, and how many cases listwise deletion discarded. The count matters because, in checks under MCAR, differences from sampling alone grew with the share of cases discarded.
  • Hypotheses. A hypothesis whose concordance changes between strategies is flagged for review.
  • What agreement means. Whether data are missing at random cannot in general be tested from the data at hand, so agreement is reported as insensitivity to the choice, not as evidence that either strategy is unbiased.
  • What lavaan actually did. Each strategy records what lavaan estimated, not only what was requested. When FIML is requested with ULS, lavaan 0.7 runs its two-stage method instead (as it does with GLS), while lavaan 0.6-21 refuses. With MLM, FIML is refused. Each of these is reported, not hidden, and the tests accept either lavaan behavior. The data supplied must reproduce the fitted model when refitted with its original strategy, so a comparison cannot silently run on different data.
  • Validation. Tests reproduce lavaan’s estimates for each strategy and use simulations with a known factor correlation of .50. Under MAR, FIML is within sampling error of it while listwise deletion is attenuated and flagged. Under MCAR, both are unbiased, and FIML’s standard errors are smaller where it keeps cases listwise deletion discards.
  • Not offered. Mean substitution is explained and not implemented.
  • Documentation. The measurement-model article gains a section applying the comparison to nomo_demo_continuous, whose missing values are MCAR by construction, and to a simulated MAR sample.

nomologR 0.2.0

nomologR 0.2.0 makes the workflow research-backed from historical to contemporary practice and more useful to graduate students and researchers. Every method is placed in its literature and cited in reports; researchers can compare competing models, revise a model with its lineage recorded, and evaluate total and subscale scores; and replication language no longer reads noise as a finding. Scope was selected in #22.

Six Planned workstreams moved to v0.2.1 at release certification (#37), each with its reason recorded: missing-data sensitivity (#32), score guidance (#33), insufficient-effort responding screens (#34), manuscript-ready tables (#35), rendering reports inside R Markdown or Quarto (#40), and optional parallel bootstrap intervals (#42). The license is GPL-3.0-only; the published 0.1.0 release keeps its original MIT license. Distribution remains R-universe; the first CRAN submission is targeted for v0.3.0 (#39).

Release certification (#37)

• README links to files excluded from the built package (ROADMAP.md, LICENSE.md) now use absolute URLs, so they resolve for readers of the package tarball as well as on GitHub (#48).
• The default-run part of the nomo_compare() example now skips the side-by-side evidence, which moved to \donttest{}, keeping every default-run example within CRAN’s time expectations.
• Added the R-hub v2 GitHub Actions workflow (run on demand) for multi-platform CRAN-style checks.
• Certification tests of every nomo_methods() crediting rule against real objects found, and fixed, methods that reports would have mis-cited: WLSMV comparisons were credited with the Satorra-Bentler scaled test instead of the scaled-and-shifted test; AIC and BIC were never credited; FIML requested as missing = "ml" or "direct" was not credited; a requested Fornell-Larcker comparison was not credited; and a nomo_partial() release specification was credited with a multiple-group fit it does not perform. A registry entry for a fixed change-in-CFI rule that no function displays was removed.
• nomo_revise() now refuses an item revision whose measurement model disagrees with the revised items. Previously, dropping an item without supplying a revised cfa_model screened and explored the reduced item set but still fitted the dropped item in the CFA, while the lineage recorded it as removed.

Bifactor and higher-order models (#29)

• nomo_model() gains structure = c("correlated", "higher_order", "bifactor") and general. Bifactor syntax writes its identification and orthogonality explicitly, so it is identified the same way whatever std.lv is used. Higher-order models need at least three first-order factors; with exactly three the model is noted as fitting exactly as well as correlated factors. Impossible configurations are refused and fragile ones noted.
• Added nomo_hierarchical(), which reads a fitted bifactor or higher-order model and reports omega total, omega hierarchical, their ratio, explained common variance (ECV), the percentage of uncontaminated correlations (PUC), and, per subscale, omega and omega hierarchical subscale, with item-level general and group loadings. For higher-order models these are the exact Schmid-Leiman decomposition.
• Indices match analytic population values and semTools::compRelSEM() under both observed and model-implied denominators. Ordered indicators report the latent-response estimand, labeled as an upper bound on observed ordinal sum-score reliability.
• No index is treated as a pass/fail threshold, following Reise (2012), who notes that no benchmark value of ECV establishes unidimensionality. The output discloses that a bifactor model usually fits at least as well as the alternatives even when it did not generate the data.
• Models whose general and group factors correlate, multi-group and multilevel models, and non-hierarchical models are refused with an explanation. The higher-order refusal shared by nomo_reliability() and nomo_validity(), and nomo_reliability()’s cross-loading refusal, now point to nomo_hierarchical().
• Added print(), summary(), plot(), and nomo_table() methods, seven methods-registry entries, and the Total and subscale scores article.

Replication-status language (#30)

• A change in sign between primary and validation estimates is no longer labeled "sign_reversal" from the point estimates alone. Estimates scattered around a null relation differ in sign about half the time, and the previous label called that “a substantively important replication discrepancy.”
• For directional predictions, a sign change is now classified from the two 95 percent confidence intervals: "sign_reversal" when both exclude zero on opposite sides, "direction_not_replicated" when exactly one excludes zero, and "sign_change_within_uncertainty" when neither does or an interval is unavailable. The first two are recorded as concerns and the third for review.
• The rule and its rationale are documented in ?nomo_network, the nomological-network vignette, and the research-basis article. It does not treat a non-significant path as evidence of no relation; that still requires a negligible(within = ...) prediction.

Methods registry (#26)

• Added nomo_methods(), a machine-readable registry of every method the package implements. Each entry records its workflow stage; its lineage (historical, contemporary, or emerging); its role (primary, supporting, or context); its estimand and key assumptions; the function and computational engine that implement it; and its references.
• role = "context" marks methods displayed only so readers can recognize them in published work, such as the eigenvalue-greater-than-one rule and fixed fit-index cutoffs. They take no part in any synthesis or decision.
• The registry lists only what the package computes. Planned methods stay in the research-basis article and the issue tracker.
• nomo_methods(x) returns only the methods a result object actually used, read from what each component recorded: a skipped retention criterion, alpha that was not computed, or a comparison that was not run is not credited. A one-factor solution is credited with no rotation method, since none was applied.
• references = TRUE expands the result into a reference list with full citations and DOIs.
• The “Methods and citations” section of nomo_report() now lists the methods the run used and a deduplicated reference list for them, ahead of the software citations it already contained. summary() of a nomo_run and nomo_table(run, "methods") show the same methods.
• Every reference is in a single bibliography that the tests keep consistent with the registry in both directions. dev/verify-method-dois.R resolves each DOI and checks that the registered first author, year, and title match the citation; all 79 DOIs pass.

Revision lineage (#28)

• Added nomo_revise(), which creates a child workflow from a parent nomo_run() with a revised measurement model, a revised item set, or both. The parent workflow is never modified.
• Each revision records what changed (model, items, or both), the required researcher rationale, and whether the change was prespecified or post hoc. Post-hoc revisions are labeled, and the decision log recommends confirming the revised model in independent data.
• Parent and revised measurement models are compared with nomo_compare() (#27), so a revision carries its own evidence. Item changes alter the observed variables, so those comparisons are descriptive, as documented.
• The factor-count decision is inherited from the parent unless the researcher supplies a new one, so a revision changes only what was intended.
• Revisions chain: $lineage accumulates one row per revision and is available through nomo_table(run, "lineage"), in print() output, and in a new “Revision lineage” section of nomo_report().
• The measurement_model = "revise" decision now points to nomo_revise() instead of only advising a fresh workflow.

Model comparison (#27)

• Added nomo_compare() for comparing two or more fitted nomo_cfa() models. A researcher rationale is required and recorded, the comparison can be labeled a priori or post hoc, and no model is ever selected automatically.
• Nested models receive the difference test that matches the estimator, computed by lavaan::lavTestLRT(): the ordinary chi-square difference test for ML, the scaled difference test for robust ML (Satorra & Bentler, 2001), and the scaled-and-shifted test for WLSMV (Satorra, 2000). A test lavaan cannot compute, or a negative scaled statistic, is reported as unavailable with lavaan’s message rather than replaced by another method.
• Nesting is checked from the models’ implied moments with semTools::net() (Bentler & Satorra, 2010). Researchers can declare nesting when the check cannot run, and a declaration the check contradicts is recorded as a concern.
• Each comparison reports changes in CFI, TLI, RMSEA, and SRMR without cutoffs, AIC and BIC when they are defined, and a plain-language interpretation.
• Side-by-side standardized loadings, reliability, AVE, and HTMT2 are reported for each model. Reliability for a model with loadings fixed to zero is labeled as describing a composite that still includes those items.
• Comparisons are refused, with an explanation, when models use different estimators, missing-data handling, cases, or data. Models with different observed variables receive descriptive evidence only.
• Added print(), summary(), plot(), and nomo_table() methods.

Maintenance (#36)

• Split the guided-workflow implementation (R/nomo_run.R, 2,066 lines) into nomo_run_state.R (state and input validation), nomo_run_decisions.R (researcher decisions), nomo_run_stages.R (stage execution and blocking), nomo_run_provenance.R (provenance, recipe, and settings tables), and nomo_run.R (entry point). Every function body and signature is unchanged, verified by comparing the parsed functions with the previous source, and the generated documentation is identical.
• Redistributed the 102 tests in tests/testthat/test-regressions.R, which were consolidated during pre-v0.1 hardening, into the per-module test files. Test names, the number of tests (464), and expectation calls are unchanged.
• Raised the declared minimum for EFAtools from 0.8.0 to 1.0.0. With EFAtools 0.8.0, nomo_factors() could not run the empirical Kaiser criterion, NEST, Hull, or comparison-data criteria, and 22 factor-retention tests failed. The full test suite passes with the declared minimums lavaan 0.6-21, semTools 0.5-9, and EFAtools 1.0.0.
• Added a CI job that installs the minimum versions declared in DESCRIPTION Imports, confirms they are the versions installed, and runs R CMD check, so a declared minimum is never an untested promise.
• Evaluated parallel execution for bootstrap reliability intervals. Four snow workers were about 3.5 times faster than serial refitting, and results are reproducible for a given seed and worker count but differ across worker counts. Because an opt-in option adds a user-facing argument and provenance fields, it is tracked separately (#42); bootstrap behavior is unchanged here.
• Recorded the distribution plan: R-universe through v0.2.x, a CRAN-readiness gate in v0.2.0 certification (#37), and the first CRAN submission targeted for v0.3.0 (#39).

Learning foundations toward v0.2.0

• Set the v0.2.0 direction: research-backed, usable scale-development and construct-validation workflows for graduate students and researchers, spanning historical to contemporary methods. Scope, priorities, and exit criteria were recorded in #22 and the linked workstream issues (#25–#37); v0.3 candidates are tracked in #38. Distribution remains R-universe; CRAN submission is deferred (#39).
• Added three simulated teaching datasets with documented population models and fixed-seed generation code in data-raw/nomo_demo.R: nomo_demo_continuous (two correlated factors with a cross-loading item, a weak item, and MCAR missingness), nomo_demo_ordinal (five-category ordered version), and nomo_demo_network (three constructs, an observed outcome, and two administration groups with a known source of scalar non-invariance) (#25).
• All workflow vignettes now evaluate against the teaching datasets, so the guided workflow, measurement-invariance, nomological-network, and reproducible-report articles show real output (#25).
• Replaced the stale overview article with Get started, including a learning path for graduate students and researchers, and added a Research basis article mapping each workflow stage from historical to contemporary practice (#25, #26).
• Added DOI-verified references to every exported analysis function (#26).
• Replaced \dontrun{} examples with runnable examples and added examples to nomo_network(), nomo_invariance(), nomo_table(), and nomo_defaults() (#25).
• nomo_table(<nomo_invariance>, "local_strain") now adds a constraint_display column with human-readable parameter labels (for example Intercept: ag3 (online vs. paper)); the raw lavaan constraint label is retained (#31). No statistical behavior changed.
• Fixed invariance figures and constraint labels that used non-ASCII symbols (Greek capital delta and arrows). On the default pdf() device these symbols were dropped or mangled, and running the vignette code under R CMD check failed. Labels now read Delta CFI, Delta RMSEA, Delta SRMR, ->, and <->, matching the hypothesis syntax.
• Documented the table types available from nomo_table().
• Grouped the pkgdown reference index by workflow stage and articles by learning path, with a redirect from the former overview URL (#25).
• Archived the completed v0.1 roadmap specification in dev/roadmap-v0.1-record.md and refocused ROADMAP.md on v0.2 and later.

Post-v0.1 development

• Changed the current development source to GNU GPL version 3 only (GPL-3.0-only; GPL-3 in R metadata), retaining the historical MIT notice. Previously published releases retain their original licenses.
• Aligned source citation metadata with development version 0.1.0.9000; release tags retain their release-specific citation metadata.
• Reconciled README, roadmap, issue navigation, historical checkpoint wording, and the documentation/planning workflow. Candidate methods remain proposals until their scope is accepted; no analytical behavior changes in this update.
• Opened development toward v0.2.0.
• v0.1.0 remains the current stable public release. # nomologR 0.1.0

First stable public release

• First stable public release of the guided empirical scale-development and construct-validation workflow.
• Includes the complete v0.1 workflow from item screening through reproducible reporting.
• Release certification includes a clean test suite, R CMD check, coverage audit, installed-package smoke test, and public pkgdown documentation.

Milestone 9 closeout — reproducible reporting and v0.1 release hardening

• Completed nomo_report() as the archival reporting layer for guided nomo_run() workflows.

• Added self-contained HTML reporting across item screening, factor retention, EFA, CFA, reliability, convergent/discriminant evidence, invariance, and theory-specified nomological networks.

• Added researcher inputs, data characteristics, decision provenance, deviations/post-hoc decisions, methods/citations, reproducibility information, and a full evidence trace linking recommendations back to their source evidence.

• Added report-render regression tests covering the core workflow and optional invariance/network branches.

• Completed pre-v0.1 engineering hardening with the full test suite passing, R CMD check at 0 errors / 0 warnings / 0 notes, executable-line coverage at 100%, and an empty covr::zero_coverage() result.

• Finalized v0.1.0 as the first stable public release after release-readiness review and certification. ## Milestone 8 closeout — guided workflow and decision provenance

• Replaced the nomo_run() development stub with a resumable guided workflow spanning screening, factor-retention evidence, EFA, CFA, reliability, convergent/discriminant evidence, optional invariance, and optional theory-specified nomological-network analysis.

• Added explicit researcher handoffs for EFA factor count, CFA model specification, and downstream continuation after measurement evidence.

• Preserved completed component results across resume calls; future-stage settings can be added without silently recomputing completed work.

• Added teaching and research presentation modes that change presentation, not statistical behavior.

• Added nomo_split-aware sample roles and exact-model network replication.

• Added stage maps, decision/rationale provenance, component evidence logs, reproducibility recipes, and the guided-workflow vignette.

• Optional invariance/network branches remain researcher configured; no automatic parameter freeing, respecification, item deletion, or one-number validity score is introduced.

• M8 closeout coverage reached 94.07% package-wide and 93.16% for R/nomo_run.R; the full test suite and R CMD check were clean.

• Milestone 8 is complete. Milestone 9 was subsequently completed in this development version; see the closeout above.

Milestones 6–7 closeout — theory-specified networks, invariance, and Checkpoint C

• Added positive(), negative(), and negligible() expectation helpers plus nomo_hypotheses() for machine-readable a-priori and post-hoc theory specifications. Quantified negligible predictions use researcher-specified SESOI/equivalence regions; bare negligible predictions remain qualitative and are never confirmed merely because p > .05.
• Added nomo_network() for theory-specified latent SEM with transparent addition of missing prespecified theory paths, observed external criteria/outcomes, standardized or explicitly unstandardized hypothesis evaluation, measurement-context propagation, and calibration/validation replication of the exact same prespecified fitted model.
• Kept theory concordance, uncertainty, measurement quality, confirmatory status, and replication as distinct evidence streams rather than collapsing them into a single validity score.
• Added nomo_invariance() using current semTools::measEq.syntax() and lavaan infrastructure with identification-aware sequences for continuous, ordered 4+ category, three-category, and binary indicators.
• Added diagnostic-only localized equality-constraint score tests and nomo_partial() for explicitly researcher-specified partial-invariance releases with required rationale and cumulative provenance. The package never searches until it finds a partial-invariance model that passes.
• Added nomo_table() methods plus researcher-facing network and invariance figures, including theory-compatible regions, replication comparison, fit/change evidence, and human-readable localized equality constraints.
• Added the “Nomological network” and “Measurement invariance” vignettes plus truth simulations, edge/failure hardening, and presentation regression tests.
• Final Checkpoint C coverage audit reached 94.19% package-wide. Core M6/M7 modules: R/nomo_hypotheses.R 94.33%, R/nomo_network.R 89.95%, R/nomo_network_presentation.R 91.49%, R/nomo_invariance.R 92.04%, R/nomo_invariance_presentation.R 93.27%, R/nomo_partial.R 90.57%, and R/nomo_table.R 100.00%.
• Advanced the development version to 0.1.0.9003, marking Checkpoint C: Generalizability and Nomological Evidence Complete. Milestone 8, nomo_run(), is the next active development target.

Milestone 5 closeout — reliability, convergent/discriminant evidence, and Checkpoint B

• Replaced the nomo_reliability() development stub with a model-based reliability workflow using current semTools::compRelSEM() infrastructure.
• Made omega/composite reliability the primary coefficient for the general congeneric CFA workflow while retaining coefficient alpha as an explicitly qualified secondary statistic.
• Added ordered-indicator reliability guidance that distinguishes observed ordinal-score reliability from latent-response-scale coefficients and refuses to silently substitute a different alpha estimand.
• Added optional nonparametric bootstrap confidence intervals for reliability estimates through repeated refitting of the same lavaan CFA. Point estimates remain unchanged; uncertainty is additive and explicit.
• Replaced the nomo_validity() development stub with convergent and construct-separation evidence including standardized loadings, AVE, latent correlations, HTMT2, original HTMT, and optional legacy/supporting Fornell-Larcker output.
• Kept AVE in the validity layer rather than mislabeling it as reliability.
• Prioritized HTMT2 for congeneric measurement models while retaining original HTMT as comparison evidence.
• Added explicit guards for ambiguous or unsupported simple-structure reliability/validity estimands, including mixed ordered/continuous composites, cross-loaded first-order structures, and silent multi-group/multilevel HTMT pooling.
• Added compact print()/summary() methods and nonredundant reliability, AVE, and construct-separation plots. Coefficient plots use the conceptual 0-to-1 scale by default and expand only when empirical estimates or intervals require it.
• User-facing audits confirmed two key teaching cases: excellent global CFA fit can coexist with weak reliability/AVE, and strong reliability/AVE can coexist with poor construct separation.
• Added the Checkpoint B vignette, “From CFA to a defensible measurement model,” integrating CFA, reliability, convergent evidence, discriminant evidence, uncertainty, and decision guidance.
• Added direct regression tests against semTools::compRelSEM(), semTools::AVE(), and semTools::htmt(), plus ordinal, redundant-construct, weak-measurement, bootstrap-uncertainty, failure-disclosure, and presentation hardening tests.
• Final M5 coverage audit reached 95.13% package-wide. Key M5 modules: R/nomo_reliability.R 90.26%, R/nomo_validity.R 92.56%, R/nomo_measurement_helpers.R 92.70%, R/nomo_reliability_uncertainty.R 89.33%, R/nomo_reliability_presentation.R 88.93%, and R/nomo_validity_presentation.R 81.20%. Coverage hardening targeted consequential branches rather than artificial 100% execution of low-value presentation/failure paths.
• Advanced the development version to 0.1.0.9002, marking Checkpoint B: Measurement Model Complete. Milestone 6, theory-specified nomological networks, is next.

Milestone 4 closeout — confirmatory factor analysis

• Replaced the nomo_cfa() development stub with a production CFA workflow around lavaan::cfa() while retaining the underlying lavaan fit unchanged.
• Added nomo_model() for simple reflective CFA syntax generation without automatically adding cross-loadings, residual covariances, or other post-hoc parameters.
• Added explicit estimator provenance and guidance for continuous ML, researcher-selected robust ML estimators such as MLR, and declared ordered-indicator WLSMV workflows.
• Added early guardrails for estimator/missing-data combinations incompatible with declared ordered indicators.
• Added convergence status, captured engine warnings, case-retention reporting, standardized loadings with uncertainty, latent-factor correlations, and global-fit evidence including chi-square, CFI, TLI, RMSEA with confidence interval, and SRMR.
• Added localized residual-correlation diagnostics and Heywood/improper-solution diagnostics.
• Added modification indices as quarantined post-hoc diagnostics; they never free parameters or trigger automatic model respecification.
• Added print(), summary(), and four CFA plot views for standardized loadings, global fit, localized residuals, and modification indices.
• Added nomo_split() for reproducible calibration/validation splitting, including the independence-versus-precision tradeoff and restoration of the caller RNG state.
• Added direct regression tests against lavaan::cfa() plus stress tests for correctly specified and misspecified CFA, omitted cross-loading, omitted correlated residual, ordinal CFA, robust ML, FIML, nonconvergence, improper solutions, modification-index quarantine, and split-sample reproducibility.
• Final M4 coverage audit reached 96.69% package-wide: R/nomo_cfa.R 97.37%, R/nomo_cfa_presentation.R 99.58%, R/nomo_model.R 97.73%, and R/nomo_split.R 98.31%.
• User-facing audit confirmed that well-specified, deliberately poor, and independent holdout CFA results are clearly differentiated without pass/fail validity language or hidden respecification.
• Local tests/checks and GitHub Actions passed; the Milestone 4 PR was squash merged before Milestone 5 development began.

Milestone 3 closeout — exploratory factor analysis and Checkpoint A

• Replaced the nomo_efa() development stub with a production common-factor EFA workflow using psych::fa() as the statistical engine and nomologR as the guidance, diagnostics, logging, and presentation layer.
• Added researcher-controlled factor counts and direct handoff from nomo_factors(), including inherited item sets, modeling types, correlation models, missing-data decisions, and explicit smoothing choices.
• Preserved M2 retention ambiguity in the EFA decision log so a handoff is not misrepresented as proof of dimensionality.
• Added MINRES + oblimin defaults, while keeping extraction and rotation choices explicit and validating supported extraction methods before calling the engine.
• Added tidy pattern/structure matrices, communalities, uniquenesses, loading complexity, factor correlations, reproduced correlations, residual matrices, localized residual pairs, and off-diagonal RMSR.
• Added KEEP, REVIEW, and STRONG REVIEW item guidance based on configurable loading, cross-loading, and communality teaching references without automatic deletion, reverse scoring, factor-count changes, rotation hunting, or hidden model refitting.
• Added KMO/Bartlett supporting evidence, descriptive sample-adequacy context, small-sample review behavior, and explicit non-positive-definite/smoothing handling.
• Added neutral public factor labels (F1, F2, …) while retaining the full underlying psych fit object for advanced inspection.
• Added summary.nomo_efa() and four teaching-oriented plot views for pattern loadings, primary/secondary loadings, unique residual pairs, and unique interfactor correlations.
• Corrected modeling-decision provenance so types inherited from a nomo_factors() object are distinguished from new researcher overrides made at the EFA stage.
• Added known-structure, cross-loading, weak-item, ordinal, redundancy, missingness, orthogonal-rotation, smoothing, validation, handoff, provenance, presentation, and failure-mode regression tests.
• Added the Checkpoint A vignette, “From item audit to exploratory structure,” demonstrating nomo_screen() -> nomo_factors() -> nomo_efa().
• Milestone 3 coverage closeout reached 97.78% for R/nomo_efa.R, 98.70% for R/nomo_efa_presentation.R, and 96.52% package-wide in the final pre-closeout audit.
• Advanced the development version to 0.1.0.9001, marking Checkpoint A: the exploratory measurement workflow is complete and Milestone 4 (CFA) is next.

Milestone 2 closeout — researcher control and hardening

• Changed modeling-type override precedence so an explicit, valid types declaration can rescue otherwise ambiguous storage before default-type rejection, while storage-compatibility checks still block unsafe coercions.
• Added decision-log entries for researcher-specified modeling-type overrides.
• Improved error ordering so constant and all-missing items are diagnosed as hard data failures before modeling-type inference.
• Expanded README guidance and examples showing when and how to use types, including ordered interpretation of factor levels and the limits of overrides.
• Added ordinal one- and two-factor regression simulations plus targeted coverage hardening for validation, criterion, presentation, and failure paths.
• Milestone 2 core computational files exceed the v0.1 >=90% coverage gate; presentation coverage reached 100% in the closeout audit.

Milestone 2B — retention triangulation and sensitivity

• Refined concordance to group closely related methods into criterion families, so original/revised MAP variants do not behave like independent votes. Internally split families remain explicit and are not forced into one bar.
• Added convenient correlation and modeling_types fields alongside the existing internal-detail names for easier inspection of automatic choices.
• Surface EKC’s non-Pearson approximation directly in criterion status and summary output rather than leaving that qualification only in the decision log.
• Polished evidence/concordance plots for longer criterion sets and wrapped captions to remain readable at ordinary RStudio plot-device sizes.
• Expanded nomo_factors() from a two-criterion workflow into configurable retention bundles: minimal, core, extended, and all.
• Added revised Velicer MAP (TR4) alongside original MAP (TR2), with both minima and criterion curves retained as sensitivity evidence.
• Added the empirical Kaiser criterion (EKC) to the default core bundle.
• Added optional NEST and Hull (CAF) criteria for supported continuous/Pearson analyses, with explicit skip reasons when their reference machinery is not compatible with ordinal, binary, mixed, or varying-N pairwise analyses.
• Added optional comparison-data retention and a clearly labeled legacy Kaiser-Guttman (> 1) rule in the all bundle; the legacy rule is excluded from the evidence synthesis.
• Parallel analysis now reports percentile, mean, and Crawford stopping-rule suggestions from the same null simulations while preserving one explicit selected rule for the primary recommendation.
• Added criterion-status and concordance tables so unavailable methods are never silently substituted and agreement is summarized without majority-vote logic.
• Expanded plot.nomo_factors() with parallel-rule sensitivity, MAP curves, multi-method evidence, and retention-concordance views.
• Added EFAtools (>= 0.8.0) as the runtime engine for modern optional retention criteria while preserving nomologR’s interpretation and decision-log layer.

Milestone 2A — factor-retention evidence

• Replaced the nomo_factors() development stub with a production factor-retention engine.
• Added explicit correlation-model selection for Pearson, polychoric, tetrachoric, and mixed indicator sets, with conservative handling of numeric-discrete items and user overrides for intended measurement level.
• Added reproducible common-factor parallel analysis using independently permuted null data and a configurable null-eigenvalue quantile.
• Added Velicer MAP as complementary retention evidence, plus observed scree information.
• Added KMO/item-level MSA and Bartlett diagnostics as supporting evidence rather than factor-count decision rules.
• Added explicit non-positive-definite matrix handling: no silent smoothing; optional smoothing must be user-requested and is recorded in the decision log.
• Added summary.nomo_factors() and plot.nomo_factors() views for retention, scree, method-concordance, and KMO evidence.
• Promoted psych to a runtime dependency because factor-retention methods now use its established factor/correlation engines directly.

Milestone 1C — integrated review and visualization

• Added summary.nomo_screen() with an integrated item-review table that combines descriptive, relationship, response-category, and decision-log evidence.
• Added plot.nomo_screen() with evidence-map, item-rest, inter-item, response-profile, and missingness views.
• Added explicit visibility for declared-but-unused ordered response categories without silently collapsing or recoding them.
• Added an item-level attention summary (none, review, concern) that remains deliberately non-prescriptive about item retention.
• Added ggplot2 as a core visualization dependency and expanded regression tests for presentation behavior.

Milestone 1B — psychometric screening

• Added corrected item-rest and inter-item relationship diagnostics for explicitly scored numeric/logical candidate items.
• Added review guidance for negative and weak item-rest relationships without automatic reverse-scoring or deletion.
• Added configurable response-concentration and near-zero-variance screening.
• Added descriptive floor/ceiling concentration for ordered and numeric-discrete items where those boundaries are meaningful.
• Added descriptive skewness and excess kurtosis for continuous-like numeric indicators without normality pass/fail declarations.
• Expanded nomo_screen() printing and tests for the psychometric screening layer.
• Refocused the README on the contentvalidR -> nomologR measurement workflow and the active v0.1 development path.
• Added a GitHub issue form for roadmap-milestone tracking.

Milestone 1 — item/data screening

• Implemented the first production slice of nomo_screen().
• Added conservative item-type descriptions, item-level missingness and response summaries, response distributions, case-level completeness diagnostics, and explicit zero-variance/all-missing flags.
• Added an evidence-guided decision log for screening observations without automatically deleting items or cases.
• Added a concise print.nomo_screen() method.
• Added focused tests for input validation, type classification, missingness, response distributions, case completeness, decision logging, and the non-destructive workflow.
• Updated GitHub Actions checkout steps to actions/checkout@v5 for the Node 24 runtime.

Foundation milestone

• Defined nomologR as a guided workflow for empirical scale development and construct-validity evidence rather than a replacement for psych, lavaan, or semTools.
• Clarified the package boundary relative to contentvalidR and solomonR.
• Standardized package naming and the planned public API on the nomo_* prefix.
• Reframed common numerical cutoffs as teaching/reference values rather than automatic pass/fail rules.
• Established the principle that the package flags, explains, and documents but never silently deletes items or respecifies models.
• Updated planned reliability methods to current semTools infrastructure such as compRelSEM().
• Updated planned measurement-invariance methods to current semTools infrastructure such as measEq.syntax().
• Separated AVE/convergent evidence from reliability.
• Added a testable internal decision-log scaffold.
• Cleaned development and continuous-integration scaffolding for the public GitHub repository.