Changelog

Source: NEWS.md

rmorie 1.4.5 - 2026-10-16

• When open.canada.ca is down, the survey PUMFs (CPADS, the Canadian Cannabis Survey, CSADS, CSUS) still load: the file URL of each resource is pinned, so the live file and then its Wayback Machine snapshot are tried without the portal’s metadata API, and after that the data.rmorie.com mirror, used only when it matches the SHA-256 recorded for the original (2026-10-16: the portal timed out and its datastore answered 504).

rmorie 1.4.4 - 2026-10-16

• A new data key needs the data access terms accepted first: rmorie login and morie_llm_login() (through rmoriebricklayer 0.5.14) show the terms, ask Do you accept ...? [y/N], and then ask for your name, organisation and why you need the data; the restricted tables also need a supervisor, department chair, director or coordinator. Give them as flags (--accept-terms, --name, --reason, …) or as arguments of morie_llm_login().

rmorie 1.4.3 - 2026-10-09

• Live progress in the long-running work, through rmoriebricklayer 0.5.13’s bricklayer_progress(): the bootstraps (morie_boot(), morie_boot_run(), morie_tsboot(), morie_bootstrap_sample()), the honest causal forest and its bootstrap, the random forests, the DML cross-fitting (morie_otis_irm_dml(), morie_dml_clustered(), morie_caus_dml_partial_lin()), super_learner(), morie_otis_aipw_superlearner() and the ten-estimator OTIS run. The bar shows only when someone is watching (an interactive session, or options(morie.progress = TRUE)), after the work has run for a moment; nested work shows one bar, and options(morie.quiet = TRUE) silences it.

• The vignettes were rewritten against the code: every chunk runs, none prints NULL, every count is recomputed, and the references to an “MRM framework paper”, a “companion paper” and a “package paper”, none of which exists, are replaced by the rmorie functions that do the work (citation("rmorie") for the software). Wrong field names (aipw$estimate and others), the IPW weights of ipw-deep-dive, and claims about functions (morie_estimate_ate() is one IPW estimate, not an ensemble) are corrected.

• morie_otis_analyze_a01_mrm() and the other Ruhela formulations say when they run on synthetic rows (the 5-row sample morie_otis_load() falls back to, or morie_synth_otis()) or on a treatment with an empty arm, and then give no interpretation; their printed results show every estimate table, and no result text claims a finding the data did not produce.

• The live OTIS loaders read CKAN’s datastore/dump with its column names (they came back as V1..V4 with the row id); the b05 and c11 samples use the real duration bins, and a table whose bins are unknown is reported, not summarised as nothing.

• Faster on the full OTIS a01 (65,467 person-years): g-computation’s bootstrap 55 s to 2 s, the SuperLearner’s forests 1,218 s to 314 s (exact node sort; rows that share a covariate pattern grow one weighted row), with the same estimates.

• morie_lmm() no longer stops on the boundary when nlminb reports singular convergence (Orthodont with a 1e-10 perturbation of y missed the REML maximum in 31 of 60 draws); native gbm propensities break an exact tie in split gain the same way on every platform.

• Fixed, each with a test: mrm_siu_outcome_classifier() on the shipped corpus; morie_siu_sanity_check() on a default read.csv(); TPS spatial measures drop the (0,0) placeholder and report how many; the Doob table summaries are computed; morie_e_value() returns e_value; morie_dag_estimate() has unique names; morie_dif_delta_mh() labels favours the right way round; morie_run_propensity_ipw_analysis() applies the survey weight; q75_days is named for what it is; morie_otis_irr_glmm_vm() uses the a01 definitions; morie_siu_compare() names unmatched fields.

• Needs rmoriebricklayer 0.5.13 (its live progress bars and the language-model layer this release is built on) and rmoriedata 0.3.6; CI installs rmoriebricklayer from its v0.5.13 tag, not a branch.

• Language models are not hosted any more: bring your own. morie_llm_ask(), morie_llm_ask_multi(), rmorie ask, rmorie models, rmorie config and rmorie doctor use the route rmoriebricklayer (>= 0.5.13) sets up: your own OpenAI-compatible endpoint, each provider whose key is set (OpenAI, Anthropic, Google Gemini, OpenRouter, Groq, Mistral and more), or a local Ollama. The model is given the help of rmorie, rmoriedata and rmoriebricklayer, with tools to search and read it. morie_llm_hosted_models() and morie_llm_probe_hosted() are gone, and route = "hosted" is an error that says so. With no model set up, the answer is the keyword fallback, marked as such.

• morie_siu_llm_extract() and the other SIU model calls take "route" (or a provider id such as "anthropic") for the model you set up; the default chain is c("ollama", "route", "gemini").

• rmorie login is for the curated tables at data.rmorie.com only (GitHub or an emailed code issues the key). A downloaded table is checked against the SHA-256 the manifest lists, and a restricted table (ICPSR, NCVS 38604, a2aj) answers with where to accept its terms.

• Research ledger, the third and last library-scan batch (20 Lean files, 84 audited theorems; 517 theorems over 103 files, 0 sorry) with R functions and tests in R/ledger_scan2.R: morie_nb_matches_zip(), morie_gamma_bound(), morie_blocking_crossover(), morie_adjustment_variances(), morie_budget_optimum(), morie_outcome_test(), morie_spatial_omission(), morie_iv_natural_bounds(), morie_design_sensitivity(), morie_moran_trend(), morie_exposure_positivity(), morie_permutation_valid(), morie_thinning_identification(), morie_poisson_chernoff(), morie_random_walk_exposure(), morie_extremal_index_clusters(), morie_tail_ratio(), morie_calibration_flip(), morie_bayes_risk_dpi(), morie_bellman_value(), morie_myopic_example(), morie_budget_rule() and morie_run_length_bound(). Every item of the 2026-09-21 library scan now has a machine-checked core.

• Research ledger, the second library-scan batch (12 Lean files, 57 audited theorems; 434 theorems over 83 files, 0 sorry) with R functions and tests in R/ledger_scan.R: morie_attenuation_two_regressor() (A4), morie_friedman_urn() (C2), morie_ruin_probability() (C4), morie_cox_overdispersion() and morie_zip_vuong_boundary() (C12, B24), morie_nonresponse_region() (B10), morie_instrumental_inequality() (B12), morie_misclassified_rd() and morie_misclassified_contrast() (B14), morie_sensitivity_amplification() (B18), morie_design_effect_direction() (A10, B9), morie_ht_elephant() (B8), morie_attractiveness_contraction() (C7) and morie_inspection_game() (A18, C24). Each function carries the theorem it computes from as its theorem attribute.

• Victimisation data and the applications that waited for it: the NCVS 1992-2022 concatenated file (ICPSR 38604; incident, person and household files, the person and household files in two year ranges each) is served at data.rmorie.com as ncvs_38604/* with the ICPSR codebooks, user guide, citation and terms of use beside it (rmorie pull ncvs_38604/incident). research/applications/ncvs_victimisation.R runs the ledger on it: the P1 survey/record pair (victimisations against the ones the police learned of, by year and class, with the partial-identification interval and the breakdown value of the 1993 to 2022 decline), the P7 concentration of victimisations on persons and households against the Poisson null, the P10 moments contagion and heterogeneity fit alike, and the A14 comparable-subset rate.

• Research ledger, fifteen more items closed with machine-checked proofs (research/lean, now 71 files, 377 audited theorems, 0 sorry): the recording rate is identified only on the survey-comparable offence set, with the dark-figure interval for the rest (morie_dark_figure_comparable(), library scan A14); the stochastic urn’s share converges almost surely and in mean, each step being the conditional expectation of the limit, so the long-run allocation is a random variable with the starting share as its mean (Lean Research.P4.Martingale, on Mathlib’s martingale convergence theorem); the hot-spot threshold of the Short-Brantingham reaction-diffusion model, the unique steady state and the dispersion relation with an explicit stability inequality (morie_hotspot_threshold()); the ring mis-specification bound for the spillover estimand (morie_spillover_ring_bound()); the hypergeometric rank-persistence null (morie_rank_persistence_null()); the stochastic urn’s share has variance at most a/(4(r0+b0)) at every horizon, so the loop drifts within a Chebyshev band fixed by the starting mass (morie_feedback_loop_envelope()); Albert-Anderson’s converse, overlap gives a finite logistic maximum-likelihood estimate (morie_logit_separation()$mle_exists); four ledger problems applied to public data (research/applications/: Toronto neighbourhood concentration against the Poisson null, break-and-enter over-dispersion and per-neighbourhood Hawkes fits, the OTIS gender gap in placement lengths split by DFL and RIF, the neighbourhood contiguity graph as the interference design), the rest listed as open with the reason; morie_weightit() for a continuous treatment follows WeightIt 2.x: the marginal density in the numerator is the sampling-weighted mixture of the conditional densities on a grid of the treatment, splined at the observed values (the 1.x normal marginal gave weights 7 to 10 percent larger); ten help pages carried “See Usage” placeholders beside the documented argument; morie_logit_separation() now settles complete versus quasi-complete separation by a second programme that maximises the smallest margin, so a vertex of the first programme sitting on a zero margin no longer reports completely separated data as quasi-complete; about 25 help pages had two @param tags on one line, folding the second parameter into the first description; the veil-of-darkness odds ratio is the relative visible-race factor and a sign test (morie_veil_of_darkness()); the three-list never-listed cell under no three-way interaction and its exp(-t) sensitivity (morie_dark_figure_no_interaction()); the error of the plug-in shrinkage factor (morie_shrinkage_factor_error()); the Gini is at most the coefficient of variation over sqrt(2), so a Poisson null caps it at about 1/sqrt(2 mean) (morie_concentration_null_bound()); the expected-cost deterrence response is linear in the log design with the discount as a coefficient ratio, and the Wald ratio identifies one dimension under exclusion and relevance (morie_deterrence_logcost(), morie_wald_iv()); the quantile RIF averages to the quantile exactly at the sample quantile and carries the gap (morie_rif_quantile(), morie_rif_gap()); two count laws with equal mean and variance but different zero mass, the second-moment core of Bartlett’s non-identification (morie_contagion_moment_witness()); the cross-age covariance of a type mixture that the aggregate age-crime curve cannot see (morie_age_crime_type_cov(), morie_age_crime_panel_cov()).

• morie_weightit(): balancing weights under WeightIt’s method names with no call into it, for binary, multi-category and continuous treatments: "glm" (any binomial link; maximum likelihood, Firth, ridge, lasso, elastic net; multinomial or cumulative-link; the density-ratio weight for a continuous treatment), "ebal", "ipt" and "cbps" (just- or over-identified), every get_w_from_ps estimand, sampling weights, stabilisation, balance moments and interactions. morie_causal_weighting() now runs "ebal" for every estimand and "ipt" natively. The engines had been in the package without an entry point; each is tested against the equation it solves and, where WeightIt is installed, against WeightIt.

• SIU: morie_siu_llm_extract() and morie_siu_anomaly_check() on a case with no harvester cache (the self-serve route) now carry the resolved drid and the cache row’s nrid instead of failing on the bookkeeping columns; a provider that answers with an empty candidates, content or choices array is reported as “returned empty text” rather than a subscript error; an empty verdict array from the auditing model is an empty audit table.

• OTIS: morie_otis_analyze_a01_ruhela_subgroup_male() and the b01 sibling refuse a subgroup under 100 cells with the same message as the female formulations.

• Dataset fetches: an HTTP error whose JSON body carries the message as the error string itself is reported with that message.

• morie_siu_reports(update = TRUE): a field name the parser does not emit (the pre-0.5.5 number_of_subject_officers) read as absent instead of a subscript error that abandoned every newer report, so the refresh now appends them.

• Tests cover the TPS hub wrappers without mocking, the advanced Hawkes drivers and their base-R fallbacks against the C++ paths, the statistical-physics routines on exact synthetic frames, morie_fetch_siu() and the SIU review side with the transport replaced by recorders, the dataset HTTP helpers on both backends, and the OTIS Mandela, chi-square, grid and master analyses with every reported number recomputed.

• Research ledger, three open items closed with machine-checked proofs (research/lean, 273 theorems and lemmas, 0 sorry): one recorded crime series cannot separate re-classification from cuffing and a second source’s true total can (morie_recording_regimes()); the Imbens-Manski cutoff exists and is unique (morie_bounds_confidence() cites im_cutoff_exists, im_cutoff_unique); across judges nested by leniency a release rule’s identified interval narrows and collapses exactly on containment (morie_selective_labels_judges()). The Lean library directory now holds all 56 proof files.

• Every estimator now runs on the package’s own code; the reference packages are used only by tests that check the numbers (the opt-in backend = of morie_taphonomy_bhm() and morie_causal_weighting()‘s methods beyond glm/ps/cbps/ebal are the two exceptions you ask for by name). Newly native, each with tests against the package it replaces: morie_roc_auc_score() (pROC’s thresholds and AUC), morie_decision_tree_split() (rpart’s CART, surrogates included), anova_table() Type II/III (car::Anova), morie_grid_search_cv() and morie_random_search_cv() (caret’s folds, draws and scores, with native lm, glm, ridge and glmnet learners), morie_effects_emmeans(), _predictions(), _comparisons(), _slopes() and _tidy() (emmeans, marginaleffects and broom for lm and glm), morie_causal_impact() (a Bayesian structural time series by Gibbs sampling, with CausalImpact’s priors and summary), morie_eg_coint(), morie_johansen_cointegration() and morie_ts_stationarity() (urca’s ADF, Johansen and a new KPSS), morie_regime_switching() (MSwM’s EM), storey_q(), estimate_pi0() and the p-value combiners (qvalue, poolr), jackknife() and morie_rsample_bootstraps(), the superlearners’ random forest, the Aldrich-McKelvey and blackbox scalings (basicspace), morie_logit_separation()’s LP (lpSolve), morie_matching_estimate_propensity(model = "gbm") (gbm with bag.fraction = 1), DBSCAN in the TPS clustering and rendering functions (dbscan, including points exactly eps apart), gam_smoother() (mgcv), and the pseudoinverse in the geron learners (MASS). Before, several of these gave different results depending on which packages were installed, and some stopped when one was missing. The fairness GAN functions no longer require torch, which they never used.

• Changed results and shapes: morie_rsample_bootstraps() returns a morie_bootstraps data frame; morie_effects_emmeans() a data frame shaped like emmeans’ summary; morie_regime_switching() returns the log-likelihood (not its negative) and n smoothed probabilities; storey_q() estimates pi0 over a lambda grid by default, as qvalue does; n_effective_tests() rounds down as poolr does; the CV searches report the chosen candidate’s score (the regression score was the worst one before).

• DBSCAN’s neighbour search uses a grid of eps-sized cells in one to three dimensions: 30,000 points cluster in about a second instead of half a minute.

• New native estimators, written in R with no outside package: morie_lmm() (linear mixed models as lme4::lmer, with nlme::lme’s random = form and containment degrees of freedom), morie_glmm() (binomial and Poisson GLMMs by the Laplace approximation, as lme4::glmer), morie_gam() (thin plate and cubic regression splines with GCV, UBRE or REML smoothness selection, as mgcv::gam), morie_rq() (quantile regression by Frisch-Newton interior point or Barrodale-Roberts, with nid, iid, kernel and bootstrap standard errors, as quantreg::rq) and morie_nls() (Gauss-Newton and Levenberg-Marquardt, as stats::nls). Each has print, summary, coef, vcov, fitted, residuals and logLik methods, and tests that check it against the reference package when that is installed. Mixed models factor one small block per group level (a Schur complement handles a second grouping factor) instead of a dense n x n matrix, so a model that took 14 s with the previous dense code takes about 1 s; a two-smooth GAM on 10,000 rows takes under a second.

• Install routes: Remotes: no longer pins rmoriebricklayer@v0.5.5 and rmoriedata@v0.3.4, so remotes::install_github("rootcoder007/rmorie") builds the current companions instead of those old tags, and the floors are raised to rmoriebricklayer (>= 0.5.12) and rmoriedata (>= 0.3.6). The README install snippets name the companions and pass repos to every install.packages() call (under Rscript there is no mirror chooser, and a bare call stops with “trying to use CRAN without setting a mirror”), and the GitHub route upgrades "always" so an older companion already installed is replaced.

• The README’s R install routes lead with pak (download progress, compiler output hidden unless a build fails, named packages upgraded to the current release), with install.packages() and remotes as the fallback; a new install-routes workflow runs every README one-liner on Linux, macOS and Windows.

• Language models: a running Ollama server with no model pulled no longer counts as a usable route. morie_llm_detect_provider() picked it (the probe only asked whether /api/tags answered), so rmorie doctor reported “active provider: ollama” and ask first sent a request with no model name; the automatic route now skips such a server (unless OLLAMA_MODEL names a model) and goes on to the next route. New morie_llm_config() and rmorie config (show, help, get, set, unset, setup, path) save the route (auto, own, ollama or a provider id) and the address, key and model of each route in ~/.config/morie/llm.json (private, written only on request; the same file and keys as rmoriebricklayer’s rmbl config); an environment variable that is set still wins. A route other than auto is the only one asked, and says what is missing when it is not set up. rmorie ask --route ROUTE picks one for a single call, rmorie doctor ends with the route and model ask will use, and rmorie help start | llm | config | r are step-by-step guides. The Ollama route now sends OLLAMA_API_KEY (the probe and the chat request did not), and an endpoint given as .../v1 (as rmorie provider set documents) is no longer asked at .../v1/v1/chat/completions; Gemini’s endpoint had the same doubled path.

• CI: the daily r-cache-warmer pre-installs rmoriebricklayer and rmoriedata from r-universe, like the other jobs. pak runs vanilla R and only sees the CRAN mirror, where the companions are older than the floor, so the warmer had failed since 1.4.1.

rmorie 1.4.2 - 2026-10-08

• Atmospheric dispersion, after rmoriebricklayer 0.5.10’s stress test (the formulas are shared): AdvectionDiffusion2d() documented a stability criterion that was false (cfl below 1 and diffusion number below 1); the explicit upwind-plus-FTCS scheme needs |u dt/dx| + |v dt/dy| + 2 (kx dt/dx^2 + ky dt/dy^2) <= 1, and a run outside it reached 10^10 from a unit pulse while reporting itself stable. The bound is returned as stability_number and stable; a call outside it is taken in enough sub-steps of a smaller dt to satisfy it (reported as substeps and dt, with a warning), so the field is a solution over the same physical time rather than an explosion.
• Physical guards throughout the dispersion functions: emission rates, masses, heights and distances non-negative, wind speed, travel time and grid steps positive, n_images a whole number (2.5 was truncated silently), the stack hotter than the air for Briggs’ rise (a cooler stack returned NaN), x0 and y0 single numbers (a vector doubled mean_x), a warning when the release is above the mixing height. GaussianPlume(q = -100) returned a negative concentration, u = 0 returned Inf, GaussianPuff(t = -100) a negative one and PgSigmas(-500) a negative sigma.
• Emissions tracking places a run offline: after MORIE_COUNTRY_ISO and the geolocation lookup, and whenever the lookup is skipped (MORIE_EMISSIONS_OFFLINE) or fails, the system time zone is mapped to its country through the IANA zone tables (every zone of zone.tab plus every alias in backward, 549 names) and then the locale’s territory, before falling back to the world average. Two-letter codes are mapped to the energy table’s three-letter codes through the full ISO 3166 list instead of a hand-picked map of thirty countries.

rmorie 1.4.1 - 2026-10-07

• Tests: the Wayback fallback test pinned the reachability probe as well as the snapshot lookup. It asserted the “no snapshot” message and so failed whenever the machine running it could not reach archive.org (the 1.4.0 CI runners). Both messages are now asserted with the probe mocked; no test touches the network.
• pull --out DIR, a directory: says “cannot write DIR: it is a directory” without R’s “not a regular file” warning ahead of it; a file it cannot open is reported once, by the error that names the cause.

rmorie 1.4.0 - 2026-10-03

• Sign-in and errors, after rmoriebricklayer 0.5.9’s network review: the page the sign-in service names is handed to the browser only when it is an https address of a public host, and a gateway’s error text is redacted of the key it was sent (by value and by shape) before it becomes an R condition.

• The hosted MORIE tier is the last resort: morie_llm_detect_provider() and the fallback chain in morie_llm_ask() / morie_llm_ask_multi() now try a local Ollama, then your own Gemini / OpenAI-compatible / OpenAI keys, and only then the hosted tier. Its address, model and sign-in service come from the signed services document at rmorie.com (through rmoriebricklayer::bricklayer_services(); an older bricklayer keeps the previous defaults), so they can change without a release. Keys are personal and issued on request at https://rmorie.com/access; every hint says so (rmorie login --token, GitHub and emailed-code sign-ins keep working). The curated-data address follows the same document.

• The 19 OTIS Ruhela-formulation analyses that returned a “not yet ported” placeholder run: morie_otis_analyze_{a01,b01,b02}_ruhela_formulations() (the ten-estimator DLRM with the IRM-DML standard-error comparison and the Naive arm), the per-year driver and its a01/b01 forms, the a01, b01 and b02 alternative-treatment and gender-subgroup formulations, and morie_otis_analyze_a01_with_csi_context() (TPS tables passed as tps_data= or loaded). New estimators behind them: morie_otis_gcomputation(), morie_otis_atc(), morie_otis_balance() and morie_otis_per_year_irm_dml(). An aggregate analysis given a table without its columns says which columns are missing.

• Weak-instrument tests computed from the data: morie_iv_montiel_olea_pflueger() (effective F, robust to heteroskedasticity, clustering and serial correlation, with Patnaik critical values for the simplified, TSLS and LIML tests) and a native morie_iv_kleibergen_paap() (rk statistic and rk Wald F by the Kleibergen-Paap SVD, robust or clustered; it returned the Cragg-Donald statistic). morie_iv_2sls() reports the effective F and its critical value beside the iid quantities, and morie_iv_stock_yogo() outside its rows points to both instead of “TODO: extend”.

• The spatial-voting estimators that returned a substitute run their own models, each checked against its reference implementation: morie_spatial_voting_alpha_nominate() (Carroll et al. 2013 slice sampler; it returned an EM-IRT fit; agrees with anominate on alpha, beta and the ideal points, r = 0.9998), morie_spatial_voting_bayesian_am() (Hare et al. 2015; it returned basicspace’s deterministic AM; matches the authors’ JAGS model to the third decimal), morie_spatial_voting_bayesian_mds() and morie_spatial_voting_bayesian_unfolding() (Bakker and Poole 2013 lognormal model; they returned smacof’s stress solution), morie_spatial_voting_ordinal_irt() (Quinn 2004 with item cutpoints; it reported MCMCpack’s loadings as ideal points, and its fallback shared one set of cutpoints across items and rescaled the draws every sweep) and morie_spatial_voting_dynamic_irt() (Martin and Quinn 2002 Gibbs sampler). Without basicspace, morie_spatial_voting_aldrich_mckelvey() computes Aldrich and McKelvey’s closed form (it fitted a different regression) and morie_spatial_voting_blackbox() fits the observed cells only (it filled missing cells with zeros); both now match basicspace, which morie_spatial_voting_blackbox() never reached because it passed missing = NA.

• Every exported function has a runnable example, and a bad first argument (wrong type, shape or missing values) stops with a plain-language message naming the argument.

• Without DBI, a cache given as db_path= keeps its files beside that path; every such cache shared one directory, so a table cached for one database was read back for another.

• morie_estimate_ate() (and the MRM “ipw ate” row) reports the HC3 sandwich standard error of the weighted regression y ~ t whose coefficient its Hajek estimate is, as morie’s Python does. It used the Horvitz-Thompson influence function, which is not centred on the group means: the standard error ran 45-57% above Python’s and 69% above the sampling SD in simulation (95% intervals that covered the truth 100% of the time). The two arms now agree to 1e-12.

• morie_estimate_double_ml() cross-fits both nuisances on one sample split, its folds drawn from the splitmix64 uniforms rmoriebricklayer and morie’s Python share (R’s sample() and numpy’s permutation never agreed, so one seed gave different folds in the two languages).

• morie_kendall_tau() and morie_wilcoxon_signed_rank_test() use base R’s default p-values (exact below 50 pairs without ties), as morie’s Python does; they forced the normal approximation. exact = chooses.

• morie_hawkes_fit() fits 50 events or more with rmoriebricklayer::core_hawkes_fit() (bounded, analytic gradient): 1,083 Weibull or Lomax events take under a second instead of 32-38 s, and a Lomax fit that tends to its exponential limit stops at the bound with a note that its shape is not identified, instead of reporting alpha = 4.8e8. The result also carries at_bound.

• morie_tps_hawkes_advanced_fit() refuses a kernel/method pair no fit can serve (soe on a kernel that is not completely monotone, inar with a sinusoidal baseline) with an error, as Python does; it returned an empty result with the reason in $warnings.

• rmorie analyze tps runs: '{"datasets":["Assault"],"nrows":5000}' (names as morie_tps_load() takes) or '{"data":"FILE.csv"}'; without either it exits 1 with the usage (it exited 0 with “not_available”).

• morie_datasets_ckan_package() read no package (the resource list arrives as a data frame) and now returns every CSV, a bilingual package’s English and French copies under “(en)” and “(fr)”; CKAN CSVs in Windows-1252 (Ontario’s library statistics) are read instead of stopping on “invalid multibyte string”.

• A read-only or full home directory no longer stops pull --out or a loader: the cache is skipped with a message naming the directory, and the data are returned; a failed cache write names the path instead of a later “cannot open the connection”.

• pull cihi820b and the other CIHI workbooks: a jurisdiction merged down its block is carried to every row (112 of 133 rows were blank), Excel’s number noise (51.959413779999998) is written as the workbook’s digits, and pull writes CSV as morie’s Python does (quotes only where needed, an empty field for a missing value), so the two arms’ files compare equal. The progress line names the dataset, one line each.

• download-bootstrap keeps a table that is already cached (--refresh downloads again), keeps one copy instead of two 230 MB files, and says “the CKAN datastore answered HTTP 500” rather than “unreachable” when the server answered.

• spwkth() computes its integrated density in milliseconds (the omega integral is exact, not a 20,001-point quadrature: 37 s before). Slow examples were trimmed and the morie_entheo_* examples say how to get the DMT_Imaging data instead of printing an error.

• DESCRIPTION requires rmoriedata (>= 0.3.4) and pins Remotes to rmoriebricklayer@v0.5.5 and rmoriedata@v0.3.4; the README installs both from r-universe (CRAN carries older versions) and describes the 0.3.4 SIU corpus (4,613 reports, 66 columns, English and French).

• Security: morie_crypto_hybrid_encrypt() and rmorie crypto encrypt now keep files private. The R arm still wrote the 1.3.x container, whose wrapping key was derived from the KEM ciphertext and the public key alone (the shared secret was computed and discarded), so anyone holding the file and the PUBLIC key could open it. The wrapping key now comes from HKDF(shared_secret || kem_ct || pk) behind the 1.4.0 marker MORIEHYB 0x02, the same bytes morie writes: rmorie and morie open each other’s files, and a 1.3.x file still opens, with a warning to encrypt it again. Files encrypted with rmorie 1.3.x or an earlier 1.4.0 build should be encrypted again. morie_crypto_hybrid_container_version() tells the two apart.

• rmorie crypto keygen --output DIR writes marked key files (MORIEPK/MORIESK, as morie) and the secret key owner-only (mode 600; it was world-readable), and refuses to replace an existing secret key without --force. decrypt --key takes a .moriesk file as well as a keystore name (a key pair from keygen --output could not decrypt anything); encrypt and decrypt take --out and refuse to overwrite an existing file without --force; the wrong kind of key file, a missing one and a 1.3.x public key are each refused with their own message.

• morie_mrm_estimate_causal_effect()‘s consensus standard error is the weighted mean of the estimators’ standard errors. It pooled them as independent studies (sqrt(1 / sum(w))), but all four run on the same rows, so the interval was about half as wide as it should be: in 300 simulated data sets (n = 500, true effect 0.8) its 95% interval covered the truth 70% of the time (morie’s run); it now covers 96%.

• morie_matching_att_matched() counts a control matched to several treated units once per pair squared (Abadie and Imbens 2006), not once per pair: with matching with replacement its interval covered the truth 83% of the time in the same simulation; it now covers 93% (95% in morie). With every control used once the standard error is unchanged.

• Citations without a title now carry the publisher’s title, and the citation gate also catches a reference written “Author (year), …”, which it had missed.

• morie_mrm_estimate_causal_effect()’s matching estimate matches with replacement and takes the ATT from the matched pairs. Without replacement, when controls were no more numerous than treated units, nearly every control was used and nothing was balanced: on the package’s own simulated design (true effect 0.8) it returned the unadjusted difference, 1.13, while IPW, AIPW and DML gave 0.85. It now gives 0.92, the same ATT and standard error as morie to ten decimals.

• SIU police_service is the service of the subject officials, read from the director’s analysis (rmorie’s own copy of the SIU core, as in rmoriebricklayer 0.5.5), not the force that notified the SIU.

• Without jsonlite, JSON is read by rmoriebricklayer’s port of jsonlite’s reader, so every caller sees the same object either way: the Socrata by-id loaders returned list() on the default install, because rows carrying a nested location object were left a bare list of records.

• HQC (morie_crypto_hqc_keygen() / _encaps() / _decaps()) runs without liboqs through rmoriebricklayer’s native HQC-1 (v5.0.0, a 32-byte secret; liboqs builds use round 4, 64 bytes); the ML-KEM and ML-DSA wrappers check key and ciphertext sizes before either backend, with the same words. The post-quantum and hybrid tests run on every install, and the network-gated tests probe a web host on port 443 (port 53 answered only on DNS servers, so 13 of them could never run).

• morie_granger_test() and morie_transfer_entropy_gaussian() refuse a constant or perfectly predictable response (an exact fit leaves round-off, not 0, in the residual sum, so the test passed and returned noise), with the same values as morie’s Python on every other series.

• Hawkes fits in seconds instead of tens of minutes. morie_tps_hawkes_advanced_fit(), morie_tps_compare_hawkes_kernels() and morie_tps_hawkes_markovian_vs_nonmarkovian() fit through rmoriebricklayer’s core_hawkes_fit() (analytic gradient, projected BFGS in C++, the routine morie’s Python calls) and take a method: "exact" (Ozaki’s O(n) recursion for the exponential kernel), "soe" (Lomax, and gamma with shape < 1, as sums of exponentials with relative error eps; Beylkin & Monzon 2010), "truncate" (kernel tail mass below eps left out), "em" (Veen & Schoenberg 2008) and "inar" (Kirchner 2017); "auto" is exact for the exponential kernel, truncate for Weibull and soe for Lomax and gamma. The reported nll is always the exact likelihood, so AIC compares across routes. They fit every event by default (max_n = NULL; a requested subsample is deterministic), prepare the event times as morie does, and the KS p-value is exact up to n = 10,000, so the two arms return the same estimate on the same data.

• Native engines everywhere a result could depend on what is installed. Survey designs are native (morie_survey_design() takes strata, clusters, nest and a finite-population correction; the mean and GLM standard errors are the Taylor linearisation of survey::svyrecvar, equal to survey to 1e-12; a survey::svydesign object is read into the same form). Before, an installed survey changed the object type and stratified or clustered designs fell back to unclustered SEs. morie_matching_full() is exact optimal full matching (a minimum-weight edge cover; its total equals optmatch’s at a tight tolerance and is below it at the default one), morie_matching_subclassify() and morie_matching_variable_ratio() follow MatchIt’s rules with identical subclasses, pairs and weights, and the OTIS match_first step uses the same matcher whether or not MatchIt is installed (it used a random-order logit match without it). morie_estimate_g_computation() and estimate_ate_gcomputation() report stdReg’s sandwich SE, computed natively (equal to stdReg::stdGlm to 1e-16); before, without stdReg the SE was sd(mu1 - mu0) / sqrt(n) or a bootstrap, which leaves out the outcome model’s uncertainty. morie_causal_weighting() is native for “glm”, “cbps” and ATT entropy balancing and uses WeightIt only for other methods. Changed results: morie_matching_nearest_neighbor() with n_neighbors > 1 now matches in rounds as MatchIt does (every treated unit gets its first control before any gets a second), pair-identical to MatchIt; it gave each treated unit all its controls in turn before. Bad input: BayesOutbreak(), DlaAggregate(), morie_dsp_ruler_fd() and two prime helpers no longer hang, DiffEnt() no longer opens a PDF device, and morie_siu_refresh_manifest() checks out_path before its 6,000-request crawl. All six CIHI workbooks read without readxl (only a binary .xls needs it), with the real header row even when a note widens the sheet. Two-proportion effect sizes add the risk difference and odds ratio beside Cohen’s h, and the frequentist test table adds Bonferroni and Benjamini-Hochberg p-values over the family. The key store no longer needs the sodium package and its files are now morie’s: scrypt and ChaCha20-Poly1305 are native (RFC 7914 and RFC 8439 vectors), so a store written by the Python arm opens here and the other way round; entries sealed by rmorie 1.3.x still open. Without libsodium, ChaCha20-Poly1305, HKDF and the random bytes run on the same native code (identical output), so the hybrid ML-KEM envelope and rmorie selftest work on every build.

• Fresh-user test, fourth pass. morie_psymet_omega() now factors by principal axes and takes hier from the Schmid-Leiman transformation (promax-rotated factors, one general factor from their correlations), so a multi-factor scale no longer reports omega hierarchical near 1. Items that load negatively on the general factor are reverse-keyed and scored the other way round first, as psych::omega() does (a message names them); on the psych::bfi items and psych::sim.hierarchical() data it agrees with psych::omega(fm = "pa", rotate = "Promax") within 0.01 (the rotations differ in detail). The two-proportion power table applies Kish’s design effect of the weights (n_eq_eff, power_deff) instead of placeholders, and power_srs uses the observed group sizes. rmorie verify reports a declared placeholder table and a header-only table as such; rmorie analyze exits 1 when every analysis of the subject failed; a small stratified total names the strata it leaves empty; ask --model names the model no provider answered for, and the SIU extraction chain ends with the same cause line; connection warnings no longer leak before a download error; login says it is still waiting; run-module prints no completion line when nothing was written; the OTIS grid names the columns a pair lacks; both spellings of a NAPS key resolve; emissions --country with an unknown code says the location is detected; morie_siu_resolve_so() accepts a case number. The REML heterogeneity estimate is the converged optimum (metafor’s default stopping threshold leaves it 0.3% off). The native SIU parser reads ordinal dates (“August 3rd, 2017”) and French months, and names the police service from the sentence that notified the SIU (17-OVI-201 now reads Guelph Police Service, 2017-08-03); SIU page text decodes in a C locale; ebac-core and ebac-integrations no longer print the expected design-weight glm warning (other glm warnings still show); the causal-effects table drops an always-empty column.

• Research: four new problems join the Lean-backed programme, each with its R function and tests. morie_meta_random_effects() and morie_meta_dl_bias() (pooling evaluations: the DerSimonian-Laird truncation is biased upward under homogeneity and the random-effects variance is never below the fixed-effect one, Research.P13); morie_logit_separation() (complete or quasi-complete separation makes the logistic likelihood climb without a maximiser, the Baldus proportionality-review logit, Research.P5); morie_bounds_confidence() (Imbens-Manski intervals for partially identified sentencing effects, Research.P11); morie_cheeger_bound() (the conductance of a hot-spot set bounds the spectral gap of the street graph, Research.P3). The Python package carries every research function at parity.

• Research: seven more problems join the Lean-backed programme, each with its R function and tests: Duncan-Davis bounds (morie_ecological_bounds(), Research.P12), monotone treatment selection (morie_sentence_effect_mts(), Research.P11), the extinction probability of a near-repeat chain (morie_contagion_extinction(), Research.P10), judge-leniency designs and what the Wald ratio identifies (morie_judge_iv(), morie_judge_iv_population(), Research.P14), the Oaxaca-Blinder decomposition of a sentencing gap with both references and the interaction (morie_disparity_decomposition(), Research.P15), Little’s law on a court docket (morie_court_backlog(), Research.P16) and the incapacitation identity with its marginal year (morie_incapacitation(), Research.P17), and selective labels for release rules (morie_selective_labels(), Research.P18), and regression to the mean at selected hot spots (morie_regression_to_mean(), Research.P19).

• Research: seven continuations close the ledger’s open items, each with a Lean theorem, an R function and tests, at parity with Python: morie_two_point_bound() (Le Cam’s two-point lower bound for the dark figure, Research.P1LeCam), morie_meta_hksj() (the Hartung-Knapp-Sidik-Jonkman interval with DerSimonian-Laird or REML heterogeneity, Research.P13HKSJ), morie_judge_slope_test() (the many-judge slope test of monotonicity, Research.P14Slope), morie_dfl_reweight() (DiNardo-Fortin-Lemieux reweighting, Research.P15Reweight), morie_backlog_censoring() (the disposed-cases mean as a bound, Research.P16Censoring), morie_incapacitation_career() (desistance and replacement, Research.P17Replacement) and morie_hotspot_shrinkage() with morie_shrinkage_loss() (empirical-Bayes shrinkage and the size of the regression-to-the-mean fall, Research.P19Shrinkage). The Lean audit now covers 264 theorems in 58 files, 0 sorry, standard axioms only.

• The pollution concentration-response coefficients match their sources (both arms): NO2 all-cause mortality RR 1.02 per 10 micrograms per cubic metre (Huangfu and Atkinson 2020, the WHO 2021 review) instead of 1.04, and PM2.5 all-cause mortality log-linear at RR 1.08 per 10 (Chen and Hoek 2020, WHO 2021) instead of an IER triple with no source; the Burnett IER stays for IHD and stroke, so verify-pollution burdens are smaller than 1.3.x reported.

• Fixes from the 1.4.0 fresh-user test agents (tests in test-agent-round-1_4_0.R): the dataset catalog carries the 24 NAPS air-quality keys and the CCHS 2022 PUMF, so both arms list the same 71 keys (morie_fetch_naps() fetches ECCC’s hourly files; a zip with no member named yields its first CSV); rmorie list-datasets prints a route for every key and the footer; rmorie pull, inspect and verify without a path exit 2; exec sees the package’s functions; edit refuses a terminal editor when stdin is not a terminal; generate-template takes the module as a positional argument, fills the module description, rejects an unknown module and never overwrites without --force; emissions --seconds Inf is refused; verify-pollution validates its numeric flags, uses one reference concentration throughout and counts avoided deaths over the exposed share; the ArcGIS downloads show their progress; the synthetic-CPADS notice is given once per session; the launcher written by install_cli() pins its library with .libPaths() so an R_LIBS in ~/.Renviron cannot replace it; the liboqs message is one sentence; a login instruction always names the email route. Round three from the same agent: verify fails p-values outside [0, 1] and reversed confidence intervals; --module selects the module’s own tables (directory or name prefix) and an unknown module is refused; a stratum shorter than its allocation carries weight 1 under --per-stratum; run-modules/pipeline reject an unknown module before loading data; run-module fails when a module wrote nothing and says what the artifact modules collect; otis-analysis and mapq-psychometrics say they run on synthetic frames; morie_run_pipeline() refuses a missing project_root; morie_bricklayer(check = TRUE) no longer mistakes the R launcher for the rmorie-cli binary; morie_siu_sanity_check() accepts the reviewed corpus’s bare officer counts and reasoning text; HTML entities in SIU text are decoded (', accented names); emissions --country names an unknown code; the OTIS causal grid returns an empty table with a message; the fallback cause names a rejected GEMINI_API_KEY; a failed read says why and what the Wayback Machine said; pull --out names a file it cannot write; sample --seed must be a number; the morie_write_audit_markdown() example runs; hyphenated catalog keys (the NAPS keys) resolve as written; emissions --seconds tops out at a day; the first-paper template reads as a sentence and names the rmorie command; verify-pollution defaults the counterfactual to 10 for NO2 and 5.8 for PM2.5; the morie_cluster() examples label their cases (the UL1.2 warning stays, by the standard); morie_bricklayer() and agent() no longer speak of a separate, proprietary rmorie-cli: the command line is this package’s launcher (install_cli()).

• Fixes from the 1.3.9 stress test (every one with a test in test-stress-1_4_0.R): rmorie verify no longer fails a table whose text column is blank throughout; inspect/verify --module match the hyphenated module name against the tables it wrote; verify-pollution refuses --exposure-mean without --exposure-prevalence instead of assuming prevalence 0; emissions --seconds validates its value and the sampler thread is joined at exit (no more core dump after an R error); sample --n is the total for stratified draws (--per-stratum for N each) and the .weight column is announced and droppable (--no-weight); run-module rejects an unknown name before loading data; exec prints what the code printed with a final newline; agent --help and perseus --help describe the verb; login --to-email needs --email; download-bootstrap asks for --survey instead of starting a 376 MB download; analyze and explain use the same usage exit code; the tutorial says so when stdin is closed; ask/percy name why no model answered (rejected hosted key, unreachable endpoint) and the local fallback text carries a fallback attribute; install_cli() writes a launcher pinned to the library it was installed from and honours ~/.Renviron; the duckdb banner is silenced on listing verbs; morie_load_dataset(), morie_spillover_ht() and morie_run_pipeline() give worded errors; morie_otis_causal_grid() skips a pair whose columns the frame lacks; morie_install_extras() looks packages up without loading them; morie_emissions_track() writes nothing unless output_dir is given; morie_matching_multi_treatment() skips the reference level for integer treatments; morie_siu_sanity_check() accepts the corpus’s yes/no, true/false and lower-case gender values; the SIU LLM chain reads GEMINI_API_KEY and knows the hosted tier; the README SIU calls match the functions; the crypto message names rmorie and SystemRequirements lists liboqs. The 116 example topics that emitted R warnings now run clean (corrected examples, and informational notes demoted from warnings to messages; morie_cluster() keeps its UL1.2 warning for unlabelled cases, by the standard).

• Datasets: the fourteen Health Infobase tables (CPADS, CSADS, CSUS) download from the portal and fall back to the copy at data.rmorie.com (hib/... keys); the three OTIS research environments are fetched from data.rmorie.com as R objects (morie_load_dataset("otisfin") returns the environment); the Ontario correctional institution locations join the catalog as otisloc. The message for a missing own file now says what it means.

rmorie 1.3.9 - 2026-10-01

• The causal-estimators module runs again on the synthetic CPADS frame: its AIPW line now lines the propensity rows up with the outcome rows by name instead of predicting from a frame without the label covariates.

• The hosted tier lists additional AI models (kimi-k2.6:cf, kimi-k2.7-code:cf, deepseek-v4-pro:cf, deepseek-v4-flash:cf, glm-5.2:cf, glm-5.3:cf, glm-5.3-flash:cf, gpt-oss-120b:cf, gpt-oss-20b:cf, llama-4-scout:cf, qwen3.8-27b:cf, nemotron-3-120b:cf and gemma-4-26b:cf) beside the ollama.com ones, and falls back to them when a cloud model is rate limited; rmorie ask --model gpt-oss-120b:cf picks one.

• Downloads that drop part-way (a reset connection, a server that closes early) are started again, up to three attempts, with the reason on screen; a short body no longer passes as a complete file.

• Command line, from a hostile QA pass: the launcher works under R 4.6 (which keeps --args in commandArgs()); rmorie pull keeps what it fetches under the user cache directory, so later run-module, pipeline and selftest runs use the real PUMF; run-module without --output-dir writes under morie-output/NAME; VERB --help describes the verb instead of running it; verify checks CSV tables; ask exits 1 when no provider answered and names a model nobody serves; login --token refuses a key the gateway rejects; doctor distinguishes a rejected key from an unreachable gateway; analyze exits 1 on an error payload; sample --n and missing files are reported in words; edit no longer creates a file when there is no editor.

• Statistics: morie_chi_square_test() gains correct = (Yates is now a documented choice; Cramer’s V uses the uncorrected statistic); morie_linear_regression_ols() treats a constant column you supply as the intercept and keeps standard errors aligned; morie_matching_nearest_neighbor() warns that 1:1 matching without replacement estimates the ATT on the matchable subset when controls are scarce (pass replace = TRUE); morie_estimate_aipw() defaults to outcome_model = "auto" (logistic for a binary outcome) in both arms; the causal-estimators module reports influence-function and bootstrap standard errors instead of values about 165 times too small; AIPW, Kaplan-Meier and the E-value refuse missing or invalid input in words. The package’s JSON class is morie_json, so loading jsonlite no longer prints an S3 override notice.

• Every download draws a live progress bar (percent, size, rate) on a terminal and prints milestone lines in a log: the catalog files, bootstrap weights, StatCan tables, Wayback copies and the curated tables at data.rmorie.com. options(morie.quiet = TRUE) silences it; the command line turns it on, R sessions show it when interactive.

• morie_hosted_manifest(), morie_hosted_datasets(), morie_load_hosted_dataset(): the curated tables at data.rmorie.com, opened by the stored MORIE key; morie_load_dataset(), morie_list_datasets() and rmorie pull accept their db/table keys. rmorie pull --all downloads the whole catalog; --dataset KEY on run-module, run-modules and pipeline.

• .cpads_default_csv() falls back to rmoriedata’s synthetic CPADS PUMF when the real one is not checked out; selftest skips only when neither exists; the first-paper template ships; ASCII-only sources; lint-clean lines.

rmorie 1.3.8 - 2026-10-01

Command line: every Python verb has an R twin

• explain, inspect, verify, profile-dataset, sample, run-modules, pipeline, emissions, verify-pollution, percy / agent, chat, selftest, tutorial, generate-template, update, crypto, ingest, download-bootstrap, exec, edit, percysuits. verify-earth-engine says it lives on the Python side.

Pollution -> health

• morie_envhealth_crf_pm25() / _no2(), morie_envhealth_attributable_fraction(), morie_envhealth_mortality_displaced(), morie_envhealth_burden(), morie_envhealth_sensitivity(), morie_envhealth_equity(), morie_envhealth_burden_by_fsa() and morie_verify_pollution(): the Python morie.envhealth module, formula for formula, with the citations.

Compute emissions with provenance

• morie_emissions_start() / _stop() / _track(): the CodeCarbon method (TDP x utilisation + RAM, IEA grid intensity via morie_emissions_carbon_intensity()), a C++ background thread sampling the kernel’s CPU counters, the same emissions.csv layout as the Python package, and a bricklayer capsule per run (emissions_manifest.json plus a signed capsule_bundle.json) that morie_emissions_verify() checks. rmorie pipeline tracks and seals its runs.

rmorie 1.3.7 - 2026-10-01

Command line

• New verbs: list-modules, run-module NAME [--output-dir DIR] [--cpads FILE], list-datasets, pull KEY [--out FILE.csv], cheatsheet, and provider set --base-url URL --key KEY [--model NAME] | show | unset. morie_module_names() lists the CPADS modules.

Your own model

• morie_llm_provider_set(), morie_llm_provider_show() and morie_llm_provider_unset() attach any OpenAI-compatible endpoint; the setting lives in the credentials file shared with the Python package and the provider chain and models use it. LLM_API_BASE_URL, LLM_API_KEY and MORIE_API_MODEL take precedence.

rmorie 1.3.6 - 2026-10-01

• Version lockstep with the Python package’s 1.3.6 hotfix. No change to the R package.

rmorie 1.3.5 - 2026-10-01

CRAN follow-ups

• No example line renders wider than 100 characters; the morie_load_dataset() example reads the built-in copy first, and the Hawkes examples fit 400 and 500 events. research/ stays in the repository and out of the tarball.

Continuous integration

• The command-line tests run under covr. testthat runs in one process under covr and on Windows, where a parallel worker died with an access violation. On R-universe’s R-oldrel macOS x86_64 runner the test files run inside a 10-minute budget (MORIE_TEST_BUDGET_SECONDS) so the check fits its 60-minute cap; every other platform runs the whole suite.

rmorie 1.3.4 - 2026-09-30

See what you can ask

• morie_llm_hosted_models() returns the models the hosted tier offers your key, with the default morie_llm_ask() would use as an attribute.
• rmorie models prints that list (default marked) and the local Ollama server’s models; rmorie doctor names them on its hosted line; rmorie ask --model NAME PROMPT picks one for a single call.

Bayesian optimisation

• morie_bayoptr() could stop with “non-finite value supplied by optim”: far from any improvement the expected-improvement closed form rounds to a subnormal of either sign and optim’s L-BFGS-B line search choked on it. EI and PI are now exactly zero below the double minimum, and a polish that fails leaves the unpolished candidate in the race instead of ending the run. The documented example returns the same optimum as the Python arm.

SIU manifest and README

• morie_siu_index(lang = "en") and lang = "fr" returned an error: the shipped siu_drid_manifest.csv.gz carried the column as X_language (written through check.names = TRUE in an earlier build) while the code reads _language. The shipped file is regenerated with the rmoriedata header and the reader normalises the old name; the manifest holds 2,532 English and 2,217 French reports.
• The README’s SIU examples define df (morie_siu_reports(), the reviewed corpus from rmoriedata) before using it, quote the real language counts, and say what the outputs-manifest example expects.
• morie_siu_audit_columns() drops blank and NA case numbers before auditing: 628 of the 5,157 reviewed reports carry no parsed case number, so sampling the corpus used to spend audit calls on empty strings.
• morie_llm_ask() without a reachable model now explains how to enable one, including the hosted tier (morie_llm_login(), rmorie login), instead of listing Python commands.

rmorie 1.3.3 - 2026-09-30

Release hygiene

• The research vignette now attaches the package; it had only ever been rendered in a session that already had rmorie loaded, so R CMD check could not build it. The hosted-model fallback test allows its (mocked) probe under R CMD check, which parks network probes by default. 1.3.2 was tagged before its check finished and was never submitted to CRAN; 1.3.3 is that release with the two fixes.

rmorie 1.3.2 - 2026-09-30

Hosted LLM tier

• morie_llm_ask() falls back to the MORIE hosted tier at llm.rmorie.com (since retired) when no local Ollama answers. Sign in once with morie_llm_login() (GitHub device flow), morie_llm_login(email = ) (a 6-digit code sent to the address), morie_llm_login(email = , to_email = TRUE) (the key itself mailed to the verified inbox) or morie_llm_login(token = ) (a key obtained elsewhere, probed once). morie_llm_logout() forgets it. The key lives owner-only in $XDG_CONFIG_HOME/morie/credentials.json, shared with the Python package; MORIE_HOSTED_KEY, MORIE_HOSTED_BASE_URL ("off" disables the tier; an empty value also does on POSIX) and MORIE_HOSTED_MODEL override it. When the gateway no longer lists the requested model the first model it does list is used, so a model retired upstream cannot turn every request into an error. morie_llm_probe_hosted() reports the tier; the probes consult their cache before touching the network.

Command line inside the package

• morie_cli() dispatches login, logout, doctor, ask, analyze, version and help; inst/bin/rmorie is the launcher and install_cli() links it onto your PATH (an rmorie.cmd wrapper on Windows). rmorie login --email you@example.com --to-email and rmorie login --token mirror the R functions above.

Policing, public safety and search

• Police operations research: the hypercube queueing model (HypercubeQueue()) and the square-root law of patrol travel distance (SquareRootLaw()).
• Search theory for search and rescue: lateral range curves, probability of detection under the inverse-cube and other sweep models, and the optimal circular search (LateralRange(), SearchPod(), OptimalCircularSearch()).
• Calls-for-service staffing: Erlang C, staffing requirements at a target service level, shift schedules, relief factors and the Wilson-Weiss obligated-time model (ErlangC(), StaffingRequirement(), ShiftSchedule(), ReliefFactor()).
• CrimeStat journey-to-crime, the circle hypothesis, risk terrain modelling, the Crime Severity Index, combat models, Massey-Denton segregation indices and 2SFCA / Gaussian accessibility.
• The native SIU core with the model-agnostic reading panel and the siu CLI; SIU case-to-decision summaries are Kaplan-Meier; the Lakner stock/flow measures in the MRM.

Spatial statistics

• Weights and lattice autocorrelation: contiguity, regime and block weights, lag and error operators, rho bounds, Geary’s C, Getis-Ord G and local G*, local and bivariate Moran, Lee’s L, join counts, Moran eigenvector filtering, correlograms, permutation tests and semivariograms (the earlier placeholder versions of these families were replaced by the real methods).
• Spatial regression: SAR, SEM, SDM, SAC and SLX diagnostic front ends (Jacobians, LM and Rao score tests, impacts, bootstraps), spatial Durbin ML, S2SLS, GS2SLS and KP-HET, the spatial J-test, spatial quantile regression, spatial probit and logit by GHK simulated ML, spatial-lag Poisson, NB, ZIP and ZINB with impacts and overdispersion tests, fixed- and random-effects spatial panels with a dynamic ARX, Bayesian SAR samplers, GWR, multiscale GWR with corrected inference, GTWR and Gaussian hidden Markov models.
• Kriging and simulation: Bayesian, area-to-point, nonstationary, GA, empirical Bayesian, probability, filtered and indicator kriging; cokriging through the linear model of coregionalisation in any dimension; Vecchia, NNGP, tapered, sparse-GP and fixed-rank kriging; space-time variograms, kriging and cokriging including the Porcu quasi-arithmetic and diffusion covariances; sequential Gaussian and indicator simulation, turning bands, LU/Cholesky, spectral and circulant-embedding fields, p-field, SNESIM, FILTERSIM and direct sampling; nested covariance model selection.
• Point processes: log-Gaussian Cox, Thomas and area-interaction processes, Berman-Turner fits, Abramson intensity, space-time K, L, G and J functions, mark correlation and variogram tests, Ripley’s translation correction, Voronoi residuals, STARMA space-time models with forecasts and portmanteau tests, space-time G*, tensor and wavelet decompositions.
• Spatial machine learning and epidemiology: spatial cross-validation, penalised, quantile, forest, boosting, SVM and stacked spatial learners; GCN, SGC and GAT layers; FPCA and persistence landscapes; spatial GLMMs by Laplace; BYM2, Leroux, CUSUM, prospective scan, ecological regression, wombling, kernel relative risk, spatial SIR, frailty and cure models; Fay-Herriot and BHF small-area EBLUPs.
• Geometry, grids and sampling: SKATER, REDCAP, AZP and max-p regionalisation, Delaunay and 3-D Voronoi, alpha shapes, natural neighbour interpolation, marching-squares contours and bands, polygon overlay, projections and datum transformations, square, hexagonal and Coons grids, random, hexagonal, adaptive-cluster and declustered sampling designs, spatial block and stationary bootstraps.

Political science

• R arms for the fifteen spatial voting models; roll-call cohesion and party influence; logistic and EM ideal points (emIRT::binIRT) and a Philox Gibbs sampler for the CJR model; the W-NOMINATE fit and the BLACKBOXT transpose scaling; majority tournaments (Condorcet, top cycle, uncovered set, Banks set, Copeland, Borda), the exact yolk, agenda setting, power indices, nucleolus and kernel; party competition (Hotelling-Downs, valence, entry, bargaining, committees, PUNE), voter utility and vote probability models, polarisation measures, optimal cutting lines, the Gelman-King incumbency advantage, seat allocation, issue salience and panel binary choice.

Optimisation and operations research

• BFGS, L-BFGS-B (Byrd, Lu, Nocedal and Zhu), Nelder-Mead, sequential quadratic and linear programming, and twelve population metaheuristics on Philox streams; linear assignment, TSP, p-median, p-centre, maximal and set covering, facility location, transportation, Clarke-Wright and time-window, multi-depot, pickup-delivery, periodic and stochastic-demand routing, the Chinese postman, constrained quadratic utility maximisation and the VCG mechanism.

Statistics from the bookshelf

• Hastie, Tibshirani and Friedman: ridge, lasso, LAR and forward stagewise, PCR, regression trees, gradient and K-class boosting, AdaBoost, additive models, MARS, projection pursuit, hierarchical mixtures of experts, flexible and mixture discriminant analysis, curds and whey, supervised principal components, archetypes, principal curves, product-density ICA, graphical models and the Ising MLE, with the model-assessment chapters.
• Wilcox robust methods, the Bilder categorical-data book (complete), Hedderich, the Morin probability results, MacKay’s coding results, Rangayyan’s physiological signal processing, Schabenberger-Gotway, Wasserman, Gibbons, twenty-three closed-form distribution families, the power-analysis modules rebuilt on exact power functions, acceptance sampling, Bonett-Wright sample sizes, Kenward-Roger, parametric AFT models, RMST and weighted log-rank tests, panel diagnostics (Pesaran CD, Wooldridge, Baltagi-Li, Conley HAC), the serial-correlation linear mixed model, robust LDA, Rosenbaum amplification and design sensitivity, the truncated-product and submax tests.

Causal inference against the reference packages

• Matching balance equals cobalt and the Abadie-Imbens SE equals Matching; sensitivity analysis equals EValue, sensemakr, konfound and tipr; TWFE and event-study DiD equal fixest; staggered DiD aggregation is did::aggte; synthetic DiD is synthdid’s estimator; fuzzy DiD is the Wald-DiD 2SLS; the wild cluster bootstrap follows boottest; RD bandwidths, bias correction and the density test equal rdrobust and rddensity; IV standard errors, GMM, Anderson-Rubin and IV probit match ivreg; honest sensitivity is the Rambachan-Roth set; the Kitagawa instrument-validity test and the CLR precision correction are real.

Other estimators verified against R

• GEE, ICC, repeated-measures and unbalanced two-way ANOVA, Lin’s CCC, Dunnett, Fleiss and partial correlations, survey estimators and the design-based GLM, JK1 and JKn replicate weights, raking calibration, boot::boot.ci intervals, DeLong’s AUC interval, cox.zph, concordance tie rules, Hommel and q-value pi0, influence measures, PPS sampling, SMO support-vector machines, Little’s MCAR test and Rubin pooling, and upper-tail p-values taken from lower.tail = FALSE throughout.

Clustering, networks and symbolic computation

• Affinity propagation, CLARANS, CURE, DIANA, CHAMELEON, DENCLUE, FLAME, growing neural gas and consensus clustering; Leiden rewritten to Traag et al. (2019), Girvan-Newman, Infomap, walktrap and label propagation; a CDCL SAT solver, DPLL(T) and linear-chain CRFs; symbolic regression with a Pareto front, symbolic calculus and the Risch algorithm, Jordan form and Galois groups.

Environmental, earth and life sciences

• Climate indices and atmospheric dispersion, hydrology, forestry, FRAGSTATS landscape metrics, soil processes, the marine module, multispectral imaging, Berkeley Earth averaging, species distribution and home range, wildlife ecology and animal movement; SMILES and SMARTS chemistry with drug-likeness and PAINS filters, DSSP, MUSCLE alignment, read assembly and disorder prediction.

Criminology research program

• The open-problems ledger (research/LEDGER.md: the dark figure of crime, contagion and branching ratios of self-exciting crime processes, ecological correlation, the recording map behind downgrade-and-caution rises, sentencing bounds under contamination, concentration, fairness and deterrence) is implemented as morie_* functions with Philox-seeded simulations that the Python arm replays exactly, and the identification results are machine-checked in Lean 4 under research/lean/ (176 theorems), next to the MRP proofs of the VSR paper (38 theorems).

Tests behind every export

• Every one of the 5,031 documented exports has a test that recomputes its value in the test body; placeholder implementations that returned fabricated numbers were replaced with the real methods, family by family, and the few that remain raise an error naming the missing method. The coverage job runs the tests and the examples on separate runners.

Windows

• Replaypack() writes its canonical text through a binary connection so the bytes on disk equal the digested text on every platform. The hosted-tier off switch works on Windows ("off"), where an empty environment variable is dropped.

Security

• The OllamaFreeAPI provider tier was removed: it sent prompts to anonymous volunteer Ollama servers found through a public registry. Its place in the chain is taken by the hosted tier above, which is authenticated and rate-limited, and the chain is now Ollama, hosted, Gemini, an OpenAI-compatible endpoint, OpenAI, then the local fallback.

Parameters that were accepted and not read

An AST scan over the package found defaulted arguments never referenced in their bodies. Those documented as no-ops are unchanged; the rest now act: cramers_v() carries a noncentral chi-square interval at confidence, random_effects_meta() implements method = "PM" and "REML", link_test() fits model_type = "logistic", morie_rdd_plot_data() returns the local fits (cutoff, p_local, bandwidth, kernel), morie_rdd_sharp() bootstraps by cluster, morie_rdd_bandwidth_rot() reports both sides of cutoff, morie_weights_greg() iterates to max_iter, morie_matching_att_matched() weights by the named column, morie_matching_abadie_imbens_se() divides the reuse count by n_matches, morie_brunner_munzel(), morie_iv_anderson_rubin(), Landmark() and Aftfit() use alpha, the ingest helpers send user_agent, morie_bayes_lm() honours stop_on_convergence, and morie_spatial_voting_aldrich_mckelvey() refuses n_dims != 1 instead of ignoring it.

Two-proportion test

mrm_twoprop_test() with no events (or nothing but events) in both arms returns chi-square 0 with p = 1 and warns, instead of NaN from chisq.test(). This is morie’s own convention (the limit of two identical arms), the same one the Python arm adopted in 1.3.2, so the two arms agree again on degenerate tables.

Missing values

morie_fairness_gini() is NA with a warning and an “undefined” interpretation when a value is missing or non-finite (it returned 0, “relatively evenly spread”). morie_tox_left_censor_impute() no longer counts NA as below the detection limit: it stays NA, is left out of fraction_censored, and is reported as n_missing. Same change in the Python arm.

Spatial voting

Every basicspace call (morie_spatial_voting_aldrich_mckelvey(), morie_spatial_voting_blackbox(), morie_spatial_voting_bayesian_am()) is gated on a double matrix with at least two rows and two columns: basicspace 0.25’s Fortran writes past its arrays on a single-row or single-column input and the process dies later (valgrind). Thin inputs take the native path. morie_spatial_voting_aldrich_mckelvey() reaches basicspace as documented: it passed missing = NA, which basicspace rejects, so every call had taken the fallback. Its numbers change: the stimulus positions are now basicspace’s, standardised to mean 0 and sd 1. On data with structure to recover the two engines agree closely (stimuli correlate about 0.97, intercepts 1.00) but not exactly; on structureless noise they do not agree at all (0.04), as nothing constrains either. The respondent intercepts and slopes are the per-respondent regression of the placements on those positions on either engine, so they are defined for every respondent; basicspace reports them only against a self-placement, which this function does not take, and a 1-7 scale had given all-NA respondent parameters. Non-finite stimuli from basicspace fall back to the native solver.

Clustering

predict.morie_cluster() computes the n x k distances to the centroids directly. It built the full (n + k)^2 distance matrix and kept n x k of it: 6 s and 2.8 GB at n = 10,000 for a 4 ms job, and out of memory at n = 20,000. morie_cluster() gains iter_max for the k-means iteration cap. One example and one test used R 4.4 idioms (%||%, psmirnov(alternative =)) on a package that declares R >= 4.3.0.

rmorie 1.3.1 - 2026-09-21

Canadian legal data: A2AJ and CanLII clients

Seven morie_ingest_a2aj_*() functions cover the A2AJ Canadian Legal Data corpus (about 226,000 court and tribunal decisions and 27,000 statutes and regulations, maintained by Access to Algorithmic Justice at Osgoode Hall and Toronto Metropolitan University): _coverage() lists what the corpus holds, _search() and _fetch() wrap the REST API, _download() and _load() bring one court’s Parquet file from Hugging Face into a data.frame, _citation_edges() turns the cited-cases field into an edge list, and _gaps() tabulates the courts the corpus does not carry. Every record carries its upstream_license.

Eight morie_ingest_canlii_*() functions wrap CanLII’s REST API for those gaps (the Human Rights Tribunal of Ontario, the BC Human Rights Tribunal, the Ontario Superior Court, Alberta, Quebec and the remaining provinces): databases, per-court decision lists with date filters, one decision’s metadata, the citator (cited cases, citing cases, cited legislation), and the legislation browse; _case_id() derives CanLII’s identifiers from a neutral citation. A free key is read from CANLII_API_KEY.

The native Parquet reader now decodes three-level LIST columns (one vector per row, NULL for a null list), which is how the A2AJ citation network fields are stored; other nesting is still refused rather than silently mis-counted. The mrm-dataset-fetchers vignette gains a section on both sources.

Fixes

• Two CodeQL findings in the recursive least-squares filter (src/morie_dsp.cpp): the row offset into the covariance matrix was an int product widened after the multiplication; it is now computed in size_t.
• Twelve roxygen blocks in the R sources carried two @examples sections, so their Rd pages ran every example twice.

rmorie 1.3.0 - 2026-09-19

Design and measurement for small or under-enumerated subpopulations

Fifteen functions for studies whose target group is a small share of the frame, is reached by a design that is not simple random sampling, or is imperfectly identified by the frame itself. The three problems compound, and the package previously covered only the estimators, not the planning.

Sample size and design: morie_sample_size_proportion() composes the finite population correction, a design effect and a non-response inflation in that order; morie_sample_size_domain() answers the question the estimators could not, namely how large the whole sample must be for the subgroup to reach a stated precision, dividing by both the subgroup share of the frame and the rate at which the frame identifies it. morie_deff_cluster() and morie_neff_cluster() give the clustering component of the design effect, which multiplies with the weight component from morie_design_effect(). morie_oversample_factor() returns the selection multiplier for a target composition together with the weight-induced design effect it creates, so the two are planned together rather than in sequence. morie_screen_design() costs a two-phase screen for a rare group. morie_alloc_optimal() allocates across strata with unequal per-unit costs, reducing to Neyman() when costs are equal.

Weighting: morie_rake() performs iterative proportional fitting to known margins, reporting convergence rather than assuming it.

Identification error: morie_misclass_correct() and morie_misclass_count() implement the Rogan-Gladen correction, which recovers a prevalence from an observed one when the classifier has known sensitivity and specificity. Where group membership is recorded administratively rather than self-reported, sensitivity below one biases every rate computed on the recorded counts.

Measurement equivalence: morie_dif_sample_size() sizes a differential item functioning comparison, whose power is governed by the smaller group; morie_dif_delta_mh() puts a Mantel-Haenszel odds ratio on the ETS delta scale with its A/B/C classification; morie_invariance_compare() runs the nested configural/metric/scalar sequence on chi-square, CFI and RMSEA; morie_irt_theta_se() and morie_irt_marginal_reliability() report score precision, which is worth reporting by group for the same reason DIF is.

New vignette subpopulation-design, including the governance framework the arithmetic sits inside.

rmorie 1.2.5 - 2026-09-18

• agent() quotes every argument it hands to the CLI. Unquoted, a parenthesis in the task was a shell syntax error, a space split the task in two and a ; ran the rest as a command. task, model and backend are validated, and a stub binary on PATH now exercises the branch the tests never reached.

• morie_safe_relabel(), morie_decode_labelled(), morie_transfer_verify() and morie_relabel_forensics(): relabel by name only, decode labelled imports by code, verify an SPSS/Stata/SAS import against the source program’s code book and frequency table, and name the mechanical step (alphabetical positional relabel, rotation, …) that reproduces an observed permutation. The OHRC 2023 correction’s four-way rotation is the documented case and the test fixture.

• morie_marginals_verify() and morie_transfer_verify() gain strict; with strict = FALSE a mismatch comes back as ok = FALSE with the permutation and message instead of an error, so morie_relabel_forensics() is reachable from the public path. morie_audit_categories() now flags leading/trailing (incl. non-breaking) whitespace, whitespace-variant duplicates, empty-string labels, missing-value sentinels stored as labels, and a reference level that is any of those. The identity permutation is reported as no permutation. The shared C++ cor_pearson is centred and clamped to [-1, 1].

• The greedy one-dimensional matcher refuses NA or NaN scores, a NaN caliper and a ratio below one instead of handing them to the sort, where a NaN comparator is undefined behaviour and corrupted the heap under the degenerate-input sweep.

• Two more process-killers from the degenerate-input sweep are R errors now: morie_kernel_pca() and morie_spectral_cluster() check the feature matrix before kernlab’s C code sees it (an empty list or a zero-row matrix used to segfault), and morie_install_extras() refuses empty or missing package names instead of handing them to install.packages().

Twenty-one modules from morie’s R arm, three-way verified

Sixteen modules that morie added in September and rmorie lacked now ship here too, each checked against the Python arm and morie’s R arm on identical inputs to twelve digits with falsifiable anchors (ledger/wave3/run3.sh): the bound families Bndcvr, Bndlgt, Bndnmt, Bndsdo, Bnsipv; the clustered-design family Ccdsgn, Ccmem, Ccngg; Cluseq, Cnsint, Crrcim, Cumcif, DepthP, DepthS, Difpst and Rdmcbw. Five more come across without a Python twin: morie_fairness_spatial_gan() and its family, the BPE tokenizer (morie_tokenizer_new() and friends), the OTIS MAPQ study module, the plot() method for morie_rich_result objects, and Qrf (quantile forest pinball loss). torch and reticulate join Suggests as the optional backends the fairness GAN and tokenizer can use.

Functions that seed the RNG now leave the caller’s stream alone

Two hundred and thirty-six functions called set.seed() internally and left the session’s random stream replaced, so a user who had seeded their session for reproducibility got identical downstream draws whatever seed they had chosen. Every one of those sites now seeds for the rest of that call only and restores the caller’s stream (or its absence) on exit; the seeded results themselves are unchanged. morie_det_rng() still seeds the session, which is its documented purpose, but the twenty-six functions that call it on their deterministic_seed path now restore the caller’s stream as well.

The compiled mean no longer overflows where base R does not

morie::core::mean() in src/morie_core.h, reached by morie_mean_cpp() and by everything built on it (variance, standard deviation, moments, bootstrap means), summed naively: rep(1e308, 3) gave Inf and rep(1e120, 3) was off by 1.4e104, so the variance of three identical values came back as 3e208. It now uses base R’s algorithm (extended precision sum plus one corrective pass) with a running mean as the fallback when the sum overflows although every input is finite. Results on ordinary data are bit-identical to mean(). morie_mean_cpp() and morie_var_cpp() gain a shared argument: TRUE (the default) uses the rmoriebricklayer kernel and FALSE the identical kernel rmorie vendors; morie_mean() and morie_var() pick the vendored one whenever the installed rmoriebricklayer predates 0.5.1, so the answer never depends on which bricklayer build resolved.

Non-finite propensity scores no longer abort optimal matching

The exact 1-D optimal matching kernel sorted its scores with <, which is not a strict weak ordering when a score is NaN or Inf; under the libstdc++ assertions that Fedora and Debian compile R packages with, that aborted the whole R session. The kernel now refuses non-finite scores with an error naming the element.

Undefined effect sizes and bootstraps warn

morie_cohens_d() warns when the denominator standard deviation is zero or missing instead of returning Inf or NaN silently, and morie_bootstrap_ci() warns when x has fewer than two observations.

Fixed

• The release notes named the Lakner decomposition morie_otis_stock_flow(); the exported name is mrm_otis_stock_flow().
• The README said “over 2,000 exported functions”; there are over 11,000.

Installation is about twice as fast

ByteCompile: no. Byte-compiling the package at install time took half of the install (about four minutes on a laptop); R compiles functions on first use anyway, so nothing is lost at run time.

Eighty-eight documented functions are now exported

Eighty-eight functions carried full roxygen documentation, references and examples but no @export, so their examples reached them through rmorie:::. Most are the genomic-prediction and kernel-method family (morie_gblup(), morie_kernel_matrix(), morie_msm_*(), morie_fda_*()), the AlphaFold loss functions (Alfviol() and its siblings) and morie_laplc(). All are exported now and their examples call them directly.

pkgdown

DESCRIPTION lists the pkgdown site URL, which pkgdown::check_pkgdown() requires.

mrm_otis_stock_flow, and the Lakner stock/flow measures

The MRM OTIS family gained the Lakner (1976) decomposition: average daily population as a stock (sum(x_i)/t, days per day), average length of stay as a flow (sum(x_i)/N, days per person), and the identity that ties them (adp = N_a * alos / t). Flow and stock rates can carry opposite signs over the same period, so reporting one as if it were the other reverses the finding – it did for an OTIS headline before this landed.

No test in the package had ever skipped

tests/testthat/setup.R decided whether it was running under R CMD check by reading R_TESTS. R sources that file at startup and then unsets it, so the variable is always empty by the time the setup file runs: every skip_on_cran() in the package was inert, on CRAN included. The discriminator is now _R_CHECK_PACKAGE_NAME_, which survives. With the gates actually firing, the tests stage dropped from 956s to 192s and the whole check from over an hour to about 39 minutes – which is what brought it back inside r-universe’s sixty-minute budget.

The SIU index is no longer swept under R CMD check

.siu_resolve_drid() and its fetch path reached the live index on every check run. They now return NA_integer_ unless live fetches are explicitly allowed, so a check no longer depends on a third party being up.

Kernel symmetry is imposed, not assumed

.gh_pairwise_sq() built its Gram matrix from a BLAS call whose result is only symmetric up to rounding, and the Cholesky downstream rejected it on some platforms. The matrix is now symmetrized explicitly. morie_ghgpm() also checks finiteness first and preserves the real chol() error instead of replacing it.

rmoriebricklayer floor

rmoriebricklayer (>= 0.4.8) is required for the native decimal conversion; older copies are skipped rather than failed.

rmorie 1.2.3 - 2026-09-09

morie_digest2int matches digest on arm

digest::digest2int() hashes a plain char, whose signedness is the platform’s: signed on x86, unsigned on arm. The native kernel forced signed char, so it agreed with digest on x86 and diverged on arm for any byte above 0x7f – the r-universe arm64 builds failed on "é", expecting -1366949649 and getting 26298542. The cast now follows the platform, as digest does.

rmorie 1.2.2 - 2026-09-08

The Rd manual builds again

R CMD Rd2pdf failed on 1.2.1’s sources, which is what the CRAN check reported as “checking PDF version of manual without index … ERROR”. Three causes, none of them visible without running the build:

• Inside \code{} Rd reads an apostrophe as opening an R-style quoted string, so a transpose or a derivative (V', g') swallowed the closing brace and \code never terminated. 90 spans in man/ and 42 in roxygen are now escaped as \', which renders identically.
• \[ and \] are not Rd escapes; inside \eqn/\deqn they reached LaTeX as display-math delimiters nested in display math, and elsewhere as undefined control sequences in the HTML manual. 661 occurrences.
• The Hawkes documentation wrote \u(t) for \mu(t); \u is LaTeX’s breve accent. RFC 7464’s record separator was a literal control byte that had decayed to U+FFFD, which pdflatex cannot typeset.

Also removed: a data() call that read as a load into the global environment, a stray top-level 5.rds, and literal angle-bracket placeholders that reached the HTML manual as unrecognised tags.

Each fixture ships once

inst/ went from 2.18 MB to 0.25 MB, and the tarball from 11.66 MB to 9.73 MB: 54 fixtures byte-identical to rmoriedata’s are resolved from there at runtime, the generated dataset catalogue is built on demand, and describe_corpus.Rds moved to rmoriedata with the rest of the data.

rmorie 1.2.1 - 2026-09-08

CRAN compliance: becomes

CRAN’s rule is that \dontrun marks an example the user cannot run – missing software, an API key – and that everything else belongs in \donttest, because \dontrun prints a “# Not run:” warning that misleads the reader. 157 example blocks were wrapped in \dontrun purely because they reach the network, which is not a reason CRAN accepts: they are runnable, they are already guarded with try(), and they degrade cleanly offline. They are now \donttest.

\dontrun is kept in the handful of places it is correct: the Vertex token/Gemini calls, the StatCan and NIBRS ingests that need an API key, and install_extras(), which installs packages.

CI keeps its speed through _R_CHECK_DONTTEST_EXAMPLES_: false rather than through misleading markup – --as-cran executes \donttest, and on a GitHub runner the network examples crawl against throttled services until the job times out. CRAN’s own machines decide for themselves; the full example sweep runs on L14.

rmorie 1.2.0 - 2026-09-08

Native specializations: the dependency-light release

This release merges the feat/native-specializations line (650 commits). Every estimator that previously delegated to a heavy runtime dependency now has a native implementation, cross-validated against the package it replaces and benchmarked: double-ML runs 40-61x faster than DoubleML, the R-learner causal forest 2.9x faster than grf, optimal matching beats optmatch’s own optimum, and design-based GLM matches svyglm to 1e-6. The full matrix is green on Windows, macOS, Ubuntu (release, devel, oldrel-1) and a true Debian bookworm container.

Wave 3: every ledger module implemented three-way

• The 567-module wave-3 ledger reports zero template stubs: each module has a Python arm, a morie R arm and an rmorie R arm, executed on identical inputs and compared key by key at twelve digits, with falsifiable anchors checked in every arm. 49 suites, 23,270 quantities, 0 differing (2026-08-16).
• Nineteen modules are deliberately held with a written reason each (architecture-with-trained-weights, fabricated-by-construction, or a primary source that does not state the rule); none is shipped as a stub.

Native JSON codec, complete (2026-09-02)

• morie_jsonlt_* is now jsonlite’s whole public surface, written in base R and pinned to jsonlite byte-for-byte by tests/testthat/test-jsonlt-parity.R whenever jsonlite is installed (1,742 grid cases, 0 differences): every toJSON() option including rownames, keep_vec_names, json_verbatim, always_decimal, time_format, UTC, no_dots; jsonlite’s number formatting (num_to_char / modp_dtoa2 rounding, digits = NA and I(n)); the fromJSON() simplifier (record lists, matrices and higher arrays, $date, _row, "NA"-string vectors, bigint_as_char); yajl-layout prettify() / minify(); validate(); serializeJSON() / unserializeJSON() for every storage mode; base64_enc/dec (+url); read_json / write_json / parse_json; ndjson stream_in / stream_out; rbind_pages.
• The package’s own readers route through this codec (.morie_from_json, .s03json_*), so no code path needs jsonlite at run time; jsonlite stays in Suggests for the parity test only.

Native hash family (digest’s surface, 2026-09-02)

• morie_digest(), morie_hmac(), morie_make_raw(), morie_digest2int(), morie_sha1() (with every class method) and morie_aes() reproduce the digest package’s API and bytes: MD5, SHA-1, SHA-224/256/384/512, CRC-32, CRC-32C, xxHash32/64, MurmurHash3 x86_32 and Jenkins one-at-a-time in src/morie_digest.cpp, written from their specifications; AES-128/192/256 in ECB/CBC/CFB/CTR. serialize() header skipping, length/skip windows, raw = TRUE, seed, file = TRUE and the error modes follow digest exactly. tests/testthat/test-digest-parity.R checks the published test vectors always and every routine against digest itself when it is installed. spookyhash, blake3 and xxh3 are not provided (no specification to implement from) and say so when requested.

Documentation reconciliation (2026-09-01)

• man/ and NAMESPACE are generated from the R/ roxygen again. The docs marathon had written about 1,950 \examples{} blocks straight into .Rd files and hand-appended exports to NAMESPACE; both were lifted back into the sources (@examples on the function’s block, exports only from @export), so roxygenise() reproduces the tree exactly.
• 25 human-written roxygen blocks had been orphaned above helper functions and were silently merged into the next helper’s page (roxygen joins every #' line between two expressions). They are re-attached to the functions they document (morie_estimate_aipw(), morie_mtr2sx(), morie_haldane(), .fit_propensity() and others); the stray export names that the merge produced are gone.
• Documentation for the compiled entry points lives in R/rcpp_docs.R (documentation-only stubs); R/RcppExports.R is pure generated glue again, so Rcpp::compileAttributes() no longer deletes any docs.
• rmorie no longer pulls in ‘jsonlite’ or ‘digest’ through rmoriebricklayer: rmoriebricklayer 0.3.8 hashes with its compiled SHA-256 core and reads/writes JSON with a native codec. The stale requireNamespace("jsonlite") guard in morie_fetch_tps() is removed.

Documentation marathon

• Every function definition in R/ now has roxygen, an .Rd page and a \value; 0 unparseable pages, 0 checkRd findings. @param text is derived from the body and quotes its evidence. Executable \examples were added in execution-gated batches; network loaders use \donttest{} with real arguments. Gaps that remain are counted, not hidden.

Fixed (found by executing paths that had never run)

• changepoint_rjmcmc(): the uniform stream returned the same number forever (rmorie), the birth/death slices used descending colons at k = 0, the interval search stopped one edge short, and the reverse-move probability indexed one entry low (log 0, births never accepted). The sampler is now byte-identical across the two R arms and mirrors the Python arm: same stream, same k posterior.

• morie_survnnr_fit(): backprop took an outer product against a nested list; the input activation was stored as list(list(a)).

• mqtmpl method = "em": was a Haley-Knott regression on a genotype probability that ignored each individual’s flanking markers (every row identical, LOD all NA). Now the Lander-Botstein EM the Python arm implements, per-individual backcross probabilities, cM positions; the three arms agree to 1e-10 on a shared anchor.

• Native Parquet reader: an OPTIONAL BYTE_ARRAY column read its own definition levels as its first string ("d\001") and dropped its last row; the schema’s repetition type now decides and 1-bit RLE/bit-packed levels are decoded and spliced (nulls become NA).

• VERSION_INVENTORY.csv regenerated (the drift check would have failed).

• morie_estimate_plr() reported its method as “(morie native)” inside rmorie; it is “(rmorie native)”.

rmorie 1.1.6 - 2026-07-22

Patch release: 07-22 hardening + CRAN prep (lockstep with morie 1.1.6)

• StatCan WDS loaders: guard malformed/HTML responses, surface embargo messages, accept v-prefixed vector IDs with a clean error for non-numeric input; morie_datasets_statcan_full_csv_url() reports status=NULL on HTML outage pages instead of crashing.
• Shared TPS datetime parser .morie_tps_parse_datetime(): explicit per-format passes (ISO 8601 incl. T/millis/Z, US m/d/Y with and without AM/PM). Fixes mixed-format vectors losing ISO entries (as.POSIXct with no format throws when any element fails, R >= 4.3) and the sample-CSV crash on Debian.
• .morie_dataset_http_post_json(): non-JSON service error pages raise a clear error with a response snippet.
• CRAN prep: optmatch added to Suggests (used via @examplesIf in the full-matching example), network examples wrapped in try(), long example lines wrapped, non-standard top-level files .Rbuildignored.
• New tests cover every hardening branch plus the Bayesian MDS initialiser and criminal-network node sizing.

rmorie 1.1.4 (development, feat/native-specializations)

Modules 14-17 (quasi-experimental front-ends, weighting, modern DiD, crim methods)

• morie_did(), morie_iv_2sls(), morie_rdd() – unified quasi-experimental entry points routing to the native toolboxes: auto-staggered CS dispatch with a Goodman-Bacon contamination warning, a Staiger-Stock weak-instrument refusal gate with the Anderson-Rubin set, and RDD with the manipulation test and placebo cutoffs bundled.
• morie_weight_* – the full propensity-weighting family (logistic/entropy/CBPS/overlap/stabilized/trimming/SuperLearner stack/diagnostics) as a composable morie_weight class; weights match WeightIt to 1e-8 and CBPS balance moments to < 1e-6.
• morie_did_sun_abraham(), morie_did_borusyak(), morie_did_did2s() – the three modern staggered-DiD estimators on the native TWFE machinery; within 0.02 of the did2s package.
• morie_crim_etas(), morie_crim_hawkes_multivariate(), morie_crim_near_repeat(), morie_crim_risk_terrain() – the criminology methods absent from the toolbox after the variant audit.
• Canonical replication fixtures (LaLonde, Basque, Lee 2008, CigarettesSW) bundled under inst/extdata/quasiex/ with literature-replication tests, plus the phase-17 composition test (DAG -> identify -> match -> weight -> estimate -> refute -> report on real data).

Major changes

• All 25 native-specialization modules are complete. Every statistical algorithm now runs on rmorie’s own R/C++ engines, each cross-validated in tests/cross/ against its reference package and benchmarked in inst/benchmarks/: matching (7 methods; modules 1-7), design-based IPW/svyglm to 1e-6 (8-9), double machine learning 40-60x faster than DoubleML (10), R-learner causal forest 1.5-2.9x faster than grf (11), all four meta-learners (12), the causal DAG toolkit with Bayes-Ball identification and DoWhy-style refutation (13), the complete DiD family — TWFE and event studies reproducing fixest to 1e-10, Callaway-Sant’Anna with DRDID-identical influence functions, Goodman-Bacon, DID-M, feTR weights, honest (Rambachan-Roth style) sensitivity (14), Abadie synthetic control + synthetic DiD (15), RDD with a native Imbens-Kalyanaraman bandwidth and McCrary test (16), k-class IV (2SLS/LIML), two-step/CUE GMM, and NEW morie_its() interrupted time series with native Newey-West (17), NEW IRT: morie_irt_2pl(), morie_irt_grm(), morie_irt_eap() (18), geostatistics: variogram ML + ordinary kriging equal to gstat to 1e-6 (19), DSP: Butterworth/filtfilt/FIR/Savitzky-Golay/Welch/ DWT (20), Hawkes MLE (21), standalone SHA-256/HMAC/PBKDF2 pinned by NIST/RFC vectors (22), native JSON/XML/HTML/Parquet parsers with optional accelerators (23), and the MRM flagship: morie_mrm_load_si_dataset(), morie_mrm_reconcile(), morie_mrm_estimate_causal_effect(), morie_mrm_report() (24).
• Full three-family post-quantum coverage: lattice (ML-KEM-768, ML-DSA-65), hash-based (SLH-DSA via liboqs + an always-available native Lamport one-time signature), and code-based (HQC-128), with morie_crypto_pqc_inventory().
• NEW module 25 — categorical-integrity guards: morie_safe_recode(), morie_safe_factor(), morie_audit_categories(), morie_crosstab_verify(), and a binary-treatment guard wired into the MRM pipeline, making the silent category-mapping corruption class (numeric codes imported without labels, positional recodes, reference releveling) structurally impossible.
• Composition test (phase 17): DAG -> identification -> matching -> DML estimate -> refutation suite -> MRM report, one test, all native.

Breaking changes

• None intended. Public signatures and result shapes are preserved; method labels now read “(rmorie native)”. morie_did_group_time_att and morie_did_doubly_robust gain an se_convention argument (“reference” reproduces did/DRDID exactly; “bessel” keeps Bessel’s correction). morie_did_twoway_fe_weights supports type = “feTR” only (the canonical diagnostic).

Bug fixes

• Cliff-Ord Moran’s I variance used a wrong combining formula (off by ~n); it now matches spdep::moran.test(randomisation = FALSE).
• The event-study wrapper treated a 0-coded never-treated group as a finite onset.
• The MRM citation string now derives from inst/CITATION at runtime.

Dependencies

• Removed from Imports: digest (native C++ SHA-256).
• Removed from Suggests (now native): fixest, did, DRDID, bacondecomp, DIDmultiplegt, TwoWayFEWeights, HonestDiD, coresynth, rdrobust, AER, ivreg, gmm, plm, psych, spdep, signal, wavelets, hawkes, emhawkes, openssl.
• jsonlite/xml2/arrow remain in Suggests as OPTIONAL parser accelerators with pure-R fallbacks; reference packages remain only for tests/cross/.
• Version locked in step with morie 1.1.4.

rmorie 1.1.3 - 2026-07-15

• Test suite runs clean on a minimal (hard-deps-only) install: every optional-package test declares its requirement with skip_if_not_installed(); live SIU network tests are opt-in (RMORIE_NETWORK_TESTS=1) and skipped on CRAN.
• New r-minimal-deps CI job permanently guards the Suggests-less install path.
• All eight df-taking sampling/weighting functions give a clear error for non-data.frame input.
• Version locked in step with morie 1.1.3.

rmorie 1.1.2

CLI analysis entry point + NYPD/CPD backends

• cli_main() — single R-side dispatcher for the rmorie-cli analyze <subject> verb (otis/siu/nypd/cpd run real analyses; tps returns a guided note). Forwards CLI flags as JSON with formal-intersection.
• morie_cpd_all_analyses() and morie_nypd_all_analyses() — new single-call analysis suites for the Chicago/NYC policing subjects (crimes by type, predpol area concentration, temporal, and race disparate-impact), mirroring morie_otis_all_analyses. Offline on bundled fixtures.

Security hardening (parity with the morie ecosystem)

Non-breaking trust-boundary guards on the process-execution and deserialization sinks. Defaults are unchanged; these add a kill-switch and a single choke point (see the new internal R/aaa_exec_guard.R).

• MORIE_NO_EXEC=1 kill-switch — now honoured by every sink that spawns an external process: morie_agent(), morie_ask_percy() (Python), morie_vertex_access_token() (gcloud), the morie_bricklayer() pip install, and the DMT_Imaging git clone. Set it on shared/CI/locked machines to disable all process execution.
• Deserialization choke point — the three readRDS() sites that take a user-supplied path (morie_cache_file(), morie_inspect_output(), morie_ml_load()) now route through .morie_safe_readRDS(), which also honours MORIE_NO_EXEC (an .rds file can execute code while loading). First-party cache/bundled loads are unchanged.
• git-ref validation — morie_entheo_clone_dmt_imaging() now rejects branch names outside ^[A-Za-z0-9][A-Za-z0-9._/-]*$, closing an option-injection vector (e.g. --upload-pack=...) into git clone.

All system2() sinks continue to use argv lists (no shell-string interpolation); that invariant is now documented in the guard file.

rmorie 1.1.0

CI hardening (docs-site build) + version re-lock with morie

• pkgdown and goodpractice dependency builds now run serial make (MAKEFLAGS=-j1), matching r-cmd-check/r-coverage: the ECOSolveR SuiteSparse/AMD vendored Makefile races under parallel make on every cache miss, which killed two consecutive main-branch pkgdown runs as silent 80-140 minute runner hangs (2026-07-09).
• pkgdown build job gained timeout-minutes: 120 so a wedged runner fails fast instead of holding the run open for hours.
• No R code changes since 1.0.5; version locked with morie 1.1.0.

rmorie 1.0.5

Figures actually write now

• Statistical-physics suite: the eight morie_tps_* statphysics functions previously accepted save_fig = TRUE and reported a written figure while writing nothing (there was no png() call in the file). Each now takes fig_dir = and writes a real PNG (SDB attractiveness/density fields, Levy CCDF + Hill fit, urban-scaling log-log fit, Lotka-Volterra trajectories, Turing lattice, inspection-game phase heatmap, premise co-occurrence network) – or says plainly that it skipped.
• New dataset bridges morie_tps_load_tps_dataset() and morie_tps_load_tps() replace the never-shipped stubs the statphysics suite guarded on (“NotYetPorted”): local project export when present, otherwise a live fetch from the public TPS ArcGIS portal. The data-seeded statphysics analyses now run anywhere.
• morie_tps_figures() gains "langevin" and "fokker_planck" families (OU path fan + evolved-vs-stationary density).
• New morie_otis_figures(): five OTIS exploratory figures (placement-length survival tail with the 15-day Mandela threshold, alert-prevalence trends, alert co-occurrence Cramer’s V, age-by-year, regional Pareto) rendered offline from the bundled b01 sample.

rmorie 1.0.4

• New morie_tps_figures(): portable batch export of the TPS stochastic diagnostic figures (Hawkes fit panel with time-rescaling residuals + KS diagnostic; SARIMA hold-out panel), fetching live from the public TPS ArcGIS portal and writing only under the caller’s out_dir. OTIS/SIU figure exports follow in the next patch release.
• .tps_stoch_date_series() now parses TPS month-name OCC_MONTH and ArcGIS epoch-millisecond OCC_DATE values (previously every live-fetched tps_stochastic analysis errored).
• morie_tps_hawkes_temporal_fit() deterministically spreads same-day tied events within the day, fixing the degenerate omega -> Inf decay on daily-resolution sources.

rmorie 1.0.3

• CRAN-submission housekeeping: refreshed cran-comments.md, added the CRAN version/downloads badges and the hex logo (man/figures/logo.png, generated reproducibly by data-raw/hex_logo.R).
• Version-locked with morie 1.0.3.

rmorie 1.0.2

• All DoubleML fits now run under a shared sequential-plan guard that also restores the caller’s {future} plan on exit, fixing FutureLaunchError and the R-devel connection-diagnostic crash introduced by {future} 1.70.0.
• future declared in Suggests (fixes the '::' import not declared NOTE).

rmorie 1.0.1

• Resolved all crypto codoc WARNINGs and the Rcpp wrapper name-collision class (12 internally-renamed exports); CI now guards against new collisions.
• TPS data directory now resolves from the project root instead of a fixed parent-hop count.
• Live CKAN generic-dispatch tests are skipped when the portal returns 0 rows (oldrel flake).

rmorie 1.0.0

• Version 1.0.0 (first alpha), version-locked with morie 1.0.0.
• New modules: forensic toxicology (morie_tox_*), cluster-robust double machine learning (morie_dml_clustered), and spatiotemporal Hawkes (morie_hawkes_st_*).

rmorie 0.9.9

• Version 0.9.9 (version-locked with morie 0.9.9).
• Citation no longer pins a package version (prevents stale “(v…)” in citations).

rmorie 0.9.8

• Ecosystem connect: fast stats via the shared rmoriebricklayer C core; morie_bricklayer() assembler + morie_core.h sync-check.
• Version 0.9.8 to match the morie family; source/UA URLs -> github.com/rootcoder007.

rmorie 0.9.6 (in development)

Bugfix: cluster-robust SE with empty cluster levels

• .otis_cluster_se(): use sum(grp^2, na.rm = TRUE). A factor cluster carrying levels absent from a subset (e.g. per-year analysis of an individual-clustered panel) made tapply() emit NA for the empty clusters, yielding an NA SE that errored downstream (if (se > 0): “missing value where TRUE/FALSE needed”). Empty clusters contribute nothing, so they are now dropped. Fixes morie_otis_irm_dml() on per-year clustered runs. Regression test added.

Citation cleanup: remove Zenodo references

• README.md: removed the “(DOIs will be re-added once we re-deposit on Zenodo.)” promise. The Zenodo deposits for the morie publication set were taken down; we are not committing to re-depositing.
• inst/CITATION: removed the stale “also the R package source on Zenodo” comment.
• NEWS.md: removed Zenodo DOI references from the historical 0.9.5.7 changelog entry.
• No live DOI strings ever shipped in rmorie; this just removes the language that implied otherwise.

Phase 1 hotfix: drop wrappers for CRAN-archived packages

• Removed morie_anchors_analyze (Phase 1.l) – upstream anchors was archived from CRAN on 2022-03-06 (check problems not corrected). pak resolver could not solve the dependency, blocking CI.
• Removed morie_causal_mediation (Phase 1.h) – upstream causalweight was archived from CRAN on 2026-05-18 because its dependency LARF was archived. Same pak resolver failure.
• DESCRIPTION: Suggests -= anchors, causalweight (35 wrapper extenders remain across Phases 1.k-1.n).
• Both wrappers can be restored from git history (a9469ec, 4d78188) if the upstream packages return to CRAN.

Phase 1.n: FDR/nonparam extenders – locfdr / fdrtool / quantreg / np / dirichletprocess / lcmm

• New file R/extenders_nonparam.R with 6 wrappers (morie_locfdr_estimate, morie_fdr_qvalues, morie_quantile_reg, morie_np_kernel_reg, morie_dp_gaussian_mixture, morie_lcmm_latent_class)
• DESCRIPTION: Suggests += dirichletprocess, fdrtool, lcmm, locfdr, np, quantreg
• Tests in tests/testthat/test-extenders-nonparam.R

Phase 1.m: spatial/multivariate extenders – gstat / copula / kernlab / metafor / mvtnorm

• New file R/extenders_spatial.R with 9 wrappers covering variograms, kriging, copulas, kernel PCA, spectral clustering, meta-analysis, multivariate normal sampling and CDF (morie_geostat_variogram, morie_geostat_krige, morie_copula_fit, morie_copula_sample, morie_kernel_pca, morie_spectral_cluster, morie_meta_rma, morie_mvnorm_sample, morie_mvnorm_pmv)
• DESCRIPTION: Suggests += copula, gstat, kernlab, metafor, mvtnorm
• Tests in tests/testthat/test-extenders-spatial.R

Phase 1.l: RDD/IRT extenders – rddensity / rdlocrand / rdpower / anchors / anominate

• New file R/extenders_rdd.R with 5 wrappers (morie_rdd_density_test, morie_rdd_local_randinf, morie_rdd_power_calc, morie_anchors_analyze, morie_anominate_ideal_points)
• DESCRIPTION: Suggests += anchors, anominate, rdlocrand, rdpower (rddensity was already listed)
• Tests in tests/testthat/test-extenders-rdd.R
• Note: the rdpower wrapper is exported as morie_rdd_power_calc because morie_rdd_power already exists in R/rdd.R as a closed-form analytical power formula taking (n, tau, sigma); the simulation-based rdpower::rdpower surface is preserved alongside.

Phase 1.k: stats extenders – DescTools / performance / ppcor / coin / randtests

A new file R/extenders_stats.R adds 17 wrapper-as-extender entry points under the canonical morie_<pkg>_* prefix that delegate to five CRAN statistics packages. Each function follows the requireNamespace-guarded hard-error pattern used by the other 1.g/1.h/1.i/1.j extenders and returns a thin two-slot list with $method (qualified upstream name) and $raw (upstream object):

• morie_desc_cramers_v, morie_desc_kappa, morie_desc_winsorize, morie_desc_gini, morie_desc_atkinson -> DescTools::{CramerV, CohenKappa/KappaM, Winsorize, Gini, Atkinson}.
• morie_performance_check_model, morie_performance_r2, morie_performance_check_collinearity, morie_performance_check_outliers -> performance::{check_model, r2, check_collinearity, check_outliers}.
• morie_ppcor_partial, morie_ppcor_semipartial -> ppcor::{pcor, pcor.test, spcor, spcor.test} (matrix-wise when y/z are omitted, single-triple test otherwise).
• morie_coin_independence, morie_coin_wilcoxon, morie_coin_oneway -> coin::{independence_test, wilcox_test, oneway_test}.
• morie_randtests_runs, morie_randtests_turning_point, morie_randtests_bartels -> randtests::{runs.test, turning.point.test, bartels.rank.test}.

DESCRIPTION: adds DescTools, ppcor, randtests to Suggests (alphabetised). coin and performance were already listed.

Tests: tests/testthat/test-extenders-stats.R covers one happy path per function, each gated by skip_if_not_installed().

Phase 1.g gap: TwoWayFEWeights + synthdid extender

Two new wrapper-as-extender entry points have been added to R/did.R to close the Phase 1.e gap. Both follow the requireNamespace-guarded hard-error pattern of the existing DiD wrappers and ship under the canonical morie_did_* namespace:

• morie_did_twoway_fe_weights(panel, group, time, treatment, outcome, type = "feTR", ...) – thin interface to TwoWayFEWeights::twowayfeweights (de Chaisemartin & D’Haultfoeuille, 2020). Returns an morie_did_twfe_diagnostics S3 list with the negative-weight count, the sum of weights, and the full twowayfeweights object as $raw. Complements morie_did_panel_fe (which estimates the TWFE coefficient itself) and morie_did_chaisemartin_dhaultfoeuille (which uses the same identification argument to deliver the DID-M estimator).
• morie_did_synthdid_estimate(panel, unit, time, treatment, outcome, vcov_method = "placebo", ...) – thin interface to synthdid::synthdid_estimate (Arkhangelsky et al., 2021) under the synthdid-canonical name. Parallel to the existing morie_did_synthetic, which keeps the rmorie result-list shape; this extender surfaces the full synthdid object and the requested variance estimator (placebo, bootstrap, jackknife).

DESCRIPTION: adds TwoWayFEWeights to Suggests. synthdid was already listed in 0.9.5.12.

Tests: tests/testthat/test-did-extender.R covers both the happy path (skipped when the optional package is not installed) and the missing-package error path.

Phase 1.g sensitivity rewrite - extend EValue / tipr / sensemakr / konfound

The sensitivity subsystem (R/sensitivity.R, ~733 LOC) keeps its existing inline math as a fallback arm but cross-references the canonical CRAN packages (rbounds, tipr, sensemakr, specr, episensr) in the Rd files, and four new wrapper-as-extender entry points have been added so MRM / paper callers can reach the full surface of these packages from inside rmorie:

• morie_sensitivity_evalue(estimate, se, sd, type = "OLS", ...) – thin interface to the EValue::evalues.* dispatch family (evalues.OLS, evalues.RR, evalues.OR, evalues.HR, evalues.MD). Pairs with the typed e_value_rr/_or/_hr/_d wrappers, which call the same backend but with a fixed scale.
• morie_sensitivity_tipping_point(estimate, smd, r2, ...) – thin interface to tipr::tip for unmeasured-confounder tipping points. Pairs with tipping_point_analysis (which targets missing-data sensitivity rather than unmeasured confounders).
• morie_sensitivity_omitted_var_bias(model, treatment, benchmark_covariates, ...) – thin interface to sensemakr::sensemakr on a fitted lm. Pairs with omitted_variable_bias (closed-form Cinelli-Hazlett robustness value when only estimate + se + dof are available).
• morie_sensitivity_konfound(estimate, se, n, n_covariates, ...) – thin interface to konfound::pkonfound for the Frank et al.
  2013. percent-bias-to-invalidate and impact-threshold-of-a- confounding-variable (ITCV).

All four hard-error with a clear install.packages(...) message if the optional dependency is missing, matching the Phase 1.e / 1.f pattern.

rosenbaum_bounds, tipping_point_analysis, omitted_variable_bias, specification_curve, and probabilistic_bias_analysis are kept in-house (the rmorie result shapes are part of the public API and the inline math remains the reference fallback). manski_bounds, bias_adjusted_estimate, and sensitivity_summary are novel / aggregator code with no clean CRAN counterpart.

DESCRIPTION: adds tipr, sensemakr, konfound to Suggests.

Tests: tests/testthat/test-sensitivity.R extended with eight new test_that() blocks (one happy path + one missing-package error path per new extender, all gated by skip_if_not_installed). ## R/causal.R rewrite + new causal extenders

The causal-inference subsystem (~876 LOC) has been thin-wrapped over the canonical CRAN causal-inference packages while preserving every inline math fallback for CRAN-only installs (dual-arm pattern).

• morie_estimate_propensity_scores() delegates to WeightIt::weightit(method = "glm", estimand = "ATE") when available; falls back to stats::glm(family = binomial()).
• morie_estimate_ate() / morie_estimate_att() / morie_estimate_atc() inherit the WeightIt delegation through the propensity-score helper; Hajek difference and influence-function SE remain inline to preserve the closed-form result shape.
• morie_estimate_aipw() keeps the AIPW score inline; the propensity arm picks up WeightIt automatically. Rd cross-references AIPW::AIPW as the canonical SuperLearner-based alternative.
• morie_estimate_g_computation() delegates the standardisation step to stdReg::stdGlm() when available; falls back to the inline glm() + treatment-flipped counterfactual contrast.
• morie_estimate_late() already used ivreg::ivreg(); now also recognises AER::ivreg() as a second delegation arm, falling back to manual 2SLS otherwise.
• morie_estimate_double_ml() / morie_estimate_irm() continue to delegate to DoubleML::DoubleMLPLR / DoubleMLIRM when available; fall back to the cross-fit ridge implementation.
• morie_e_value() delegates to EValue::evalue() when available; falls back to the inline closed-form RR -> E-value formula.
• morie_sensitivity_rosenbaum() delegates to rbounds::psens() when available; falls back to the inline sign-score bound.

Four new extender functions are introduced for CRAN dependencies that previously had no morie_* entry point:

• morie_causal_impact(data, pre_period, post_period, model_args) -> CausalImpact::CausalImpact() (Brodersen et al. 2015 Bayesian structural time-series intervention analysis).
• morie_causal_weighting(data, treatment, covariates, method, estimand, ...) -> WeightIt::weightit() (full method palette: glm / cbps / ebal / ps / energy / optweight).
• morie_causal_robust_se(model, type, cluster, ...) -> sandwich::vcovHC() / vcovHAC() / vcovCL() (HC0-HC5, HAC, one-way cluster-robust variance).
• morie_causal_mediation(y, d, m, x, trim, boot, ...) -> causalweight::medweight() (semiparametric IPW direct / indirect effect decomposition).

The four extenders hard-error on missing packages (no inline fallback) since each upstream implementation is too large to re-implement compactly. New tests live in tests/testthat/test-causal-extenders.R and are guarded with skip_if_not_installed().

DESCRIPTION Suggests now lists AIPW, CausalImpact, causalweight, rbounds, sensitivitymv, and stdReg so the extenders and the delegation arms can find their upstream packages when available.

R/bootstrap_methods.R rewrite - delegate to boot / bootstrap / resample / rsample / simpleboot / coin / ipred / sandwich

The bootstrap / resampling subsystem (~867 LOC) has been re-routed through the canonical CRAN packages while preserving the rmorie API and the morie_bootstrap_result / morie_jackknife_result / morie_permutation_test_result / morie_cv_result S3 return shapes so that stat_commands, the print methods, and MRM analyses keep working unchanged.

Thin-wrapped (with requireNamespace-guarded delegation arm + inline fallback so the wrapper keeps working on minimal installs):

• bootstrap() – nonparametric bootstrap now delegates to boot::boot (stratification via strata=) and boot::boot.ci for percentile / basic / normal / bca CIs. The studentized CI path and the cluster resampling path remain inline because boot::boot.ci(type = "stud") requires a precomputed variance estimator and boot::boot does not expose cluster-of-clusters resampling for the rmorie statistic(data) -> scalar signature.
• parametric_bootstrap() – delegates to boot::boot(sim = "parametric") with an MLE / ran.gen pair built from the requested distribution.
• block_bootstrap() – moving / stationary blocks delegate to boot::tsboot(sim = "fixed" / "geom"); the circular-block path stays inline because boot::tsboot does not expose a circular sim mode.
• jackknife() – delegates to bootstrap::jackknife when installed (rmorie reconstructs pseudovalues / influence values around the returned jack.values).
• permutation_test() / paired_permutation_test() – inline shuffle loops retained (the rmorie API returns the full null distribution which downstream MRM code consumes); coin’s oneway_test(distribution = "approximate") and symmetry_test(distribution = "approximate") are cross-referenced as the canonical CRAN equivalents.
• bootstrap_632() – inline .632 / .632+ retained because the rmorie API takes naked model_fn / score_fn callables; ipred::errorest(estimator = "632plus") is cross-referenced.
• repeated_cv() / leave_one_out_cv() – delegate to rsample::vfold_cv for fold construction when no stratification or grouping is requested; caret::trainControl and rsample::loo_cv cross-referenced.

Kept as in-house implementations (no clean CRAN drop-in for the rmorie API):

• subsampling() – Politis-Romano-Wolf rate scaling over a user-supplied statistic(data); np::npsubsample is closest but is kernel-specific.
• delete_d_jackknife() – generalised jackknife with custom max_subsets enumeration cap; resample::jackknife is cross-referenced.
• wild_bootstrap() – returns the full resampled coefficient distribution; sandwich::vcovBS (variance only) and fwildclusterboot::boottest (p-value only) cross-referenced.

Added four new extender entry points (thin pass-through to the canonical packages):

• morie_boot_run(data, statistic, R, strata, ...) – direct boot::boot bridge with rmorie-style statistic(data) -> scalar adapter.
• morie_boot_basic_ci(boot_obj, type, conf) – direct boot::boot.ci bridge returning tidy list(perc = c(lo, hi), bca = c(lo, hi), ...).
• morie_rsample_bootstraps(data, times, ...) – direct rsample::bootstraps bridge returning an rset.
• morie_simpleboot_two(x, y, statistic, R, ...) – direct simpleboot::two.boot bridge for two-sample bootstrap of a scalar statistic.

DESCRIPTION: adds boot, bootstrap, coin, ipred, resample, rsample, simpleboot to Suggests (sandwich was already present).

Function inventory is preserved: 11 prior exports kept (bootstrap, parametric_bootstrap, wild_bootstrap, block_bootstrap, jackknife, delete_d_jackknife, permutation_test, paired_permutation_test, subsampling, bootstrap_632, repeated_cv, leave_one_out_cv) plus 4 new (morie_boot_run, morie_boot_basic_ci, morie_rsample_bootstraps, morie_simpleboot_two). All 20 (15 prior + 5 new) tests/testthat/test-bootstrap_methods.R test_that blocks pass (73 expectations, 2 conditional skips when rsample or simpleboot is not installed).

R/effects.R rewrite - thin-wrap over emmeans / marginaleffects / broom / stdReg / rbounds / EValue (phase 1.j)

The treatment-effect / marginal-effects module has been thin-wrapped around its canonical CRAN extender packages and gained a new family of morie_effects_* wrappers over Vincent Arel-Bundock’s marginaleffects API plus the emmeans and broom ecosystems.

Thin-wrapped (existing API preserved; requireNamespace-guarded delegation arm + inline fallback so the wrapper keeps working on CRAN-only installs):

• estimate_ate() – guarded HC3 SE via sandwich::vcovHC + lmtest::coeftest; falls back to naive model SE with a warning.
• estimate_plr(), estimate_pliv() – already DoubleML-aware; retained as is with style cleanups (snake-case locals, stats::qnorm / stats::pnorm qualification, brace style).
• estimate_ate_gcomputation() – delegates to stdReg::stdGlm() for the canonical regression-standardisation backend (analytic SE on the contrast scale, no bootstrap needed); falls back to the legacy inline bootstrap implementation when stdReg is absent.
• sensitivity_rosenbaum() – delegates to rbounds::psens() when rbounds is installed and the matched-pair count is sufficient; falls back to the inline normal-approximation Wilcoxon signed-rank bounds otherwise.
• e_value() – delegates to EValue::evalues.OLS() when both EValue is installed AND the caller supplies an outcome standard deviation via the new sd_y argument; falls back to the closed-form continuous-scale RR proxy so the Python and R ports stay numerically aligned by default.

New extender wrappers (each a thin pass-through; the underlying package’s native object is returned verbatim so downstream code keeps working with the canonical API):

• morie_effects_emmeans(model, specs, ...) -> emmeans::emmeans(). Use emmeans::pairs() / emmeans::contrast() on the returned emmGrid for pairwise or custom contrasts.
• morie_effects_predictions(model, newdata, ...) -> marginaleffects::predictions().
• morie_effects_comparisons(model, variables, ...) -> marginaleffects::comparisons().
• morie_effects_slopes(model, variables, ...) -> marginaleffects::slopes().
• morie_effects_tidy(model, ...) -> broom::tidy() (with a summary()-based fallback frame for the lm / glm classes rmorie’s MRM pipeline ships, so downstream tidy callers work even on minimal-Suggests CI runs).

The effects and margins packages are added to Suggests for the cross-reference path – users who want Fox’s effect() / predictorEffects() or Leeper’s Stata-style margins::margins() can call them directly on a model fitted via rmorie, without an intervening wrapper.

DESCRIPTION: adds broom, effects, emmeans, marginaleffects, margins, performance, rbounds, stdReg to Suggests.

Net: R/effects.R grows from 461 to ~640 LOC because the legacy treatment-effect functions now carry both a CRAN-delegation arm AND an inline fallback (CRAN policy does not allow Suggests to be hard required), plus 5 new extender wrappers + 1 shared helper. All 6 legacy exports preserved (estimate_ate, estimate_plr, estimate_pliv, estimate_ate_gcomputation, sensitivity_rosenbaum, e_value); 5 new exports added (morie_effects_emmeans, morie_effects_predictions, morie_effects_comparisons, morie_effects_slopes, morie_effects_tidy).

R/multiple_testing.R rewrite - delegate to poolr / qvalue / harmonicmeanp / gMCP / mutoss

The multiple-testing-correction subsystem (~916 LOC) has been rewritten to forward to the canonical CRAN / Bioconductor packages where one exists. Every wrapper preserves the rmorie API and the morie_multiple_testing_result / morie_rich_result S3 shape so that the stat_commands dispatcher, the print.morie_multiple_testing_result method, and MRM analyses keep working unchanged.

Thin-wrapped (with requireNamespace-guarded delegation arm + inline fallback so the wrapper keeps working on CRAN-only installs):

• bonferroni(), holm(), hochberg(), hommel(), benjamini_hochberg(), benjamini_yekutieli() – already stats::p.adjust wrappers, retained as-is.
• sidak(), holm_sidak() – inline closed-form math; mutoss cross-referenced in Rd as the canonical step-down equivalent.
• storey_q() – delegates to qvalue::qvalue (Bioconductor) when installed; falls back to the inline Storey cutoff otherwise.
• estimate_pi0() – delegates to qvalue::pi0est for storey / bootstrap methods when installed; falls back to inline estimators.
• fisher_combined() – delegates to poolr::fisher when installed.
• stouffer_combined() – delegates to poolr::stouffer when installed and no weights are supplied (weighted Stouffer stays inline because poolr does not take per-test weights).
• tippett_combined() – delegates to poolr::tippett when installed.
• harmonic_mean_p() – delegates to harmonicmeanp::p.hmp when installed.
• n_effective_tests() – delegates to poolr::meff (Galwey / Li-Ji / Nyholt) when installed.
• fixed_sequence(), fallback_procedure() – gMCP cross- referenced in Rd as the canonical graphical-MCP equivalent; inline implementations retained to preserve the rmorie stage-list return shape.

Kept as in-house implementations (no clean CRAN drop-in for the rmorie API):

• cauchy_combination() – Liu and Xie 2020; ACAT is GitHub-only.
• hierarchical_bonferroni() – rmorie-specific stage-list return shape; gMCP covers the concept with a different graphical API.
• local_fdr() – Efron empirical-Bayes KDE shape that locfdr does not match in return structure.
• permutation_fwer(), permutation_fdr() – step-down max-T and empirical-null-p FDR over user-supplied null matrices; no CRAN function exposes the same API.
• adjust_p_values() – front-end dispatcher across the rmorie wrappers; consumed by stat_commands.

DESCRIPTION: adds poolr, qvalue, gMCP, harmonicmeanp, multcomp, mutoss to Suggests.

Net: R/multiple_testing.R grows from 916 to 1118 LOC because each thin-wrap function now carries both a CRAN-delegation arm AND an inline fallback (CRAN policy does not allow Suggests to be hard required, and qvalue is Bioconductor only). Function inventory is unchanged: 25 exports preserved (bonferroni, sidak, holm, hochberg, hommel, holm_sidak, benjamini_hochberg, bh, benjamini_yekutieli, by_fdr, storey_q, fisher_combined, stouffer_combined, tippett_combined, simes_combined, harmonic_mean_p, cauchy_combination, fixed_sequence, fallback_procedure, hierarchical_bonferroni, estimate_pi0, adjust_p_values, n_effective_tests, local_fdr, permutation_fwer, permutation_fdr, plus the print.morie_multiple_testing_result S3 method). All 85 tests/testthat/test-multiple_testing.R assertions pass on a fallback-only install.

R/did.R rewrite - delegate to did / DRDID / fixest / HonestDiD / bacondecomp / DIDmultiplegt

The DiD subsystem (~1,719 LOC) has been rewritten to forward to the canonical CRAN packages instead of carrying ~700 LOC of base-R fallback code:

• morie_did_panel_fe() is now a thin wrapper over fixest::feols(y ~ d | unit + time) with cluster-robust SE and hard-errors if fixest is not installed (the base-R two-way within-transform fallback has been removed).
• morie_did_event_study() is a thin wrapper over fixest::feols with fixest::i(rel_time, ref = -1) relative-time dummies plus unit and time fixed effects; hard-errors if fixest is missing.
• morie_did_group_time_att() is a thin wrapper over did::att_gt for the Callaway-Sant’Anna group-time ATTs and hard-errors if did is not installed (the base-R bootstrap fallback has been removed).
• morie_did_doubly_robust() is a thin wrapper over DRDID::drdid_rc (Sant’Anna-Zhao 2020 repeated-cross-section doubly-robust DiD) and hard-errors if DRDID is missing (the hand-written GBM-or-logistic + linear-or-GBM bootstrap fallback has been removed). The ps_model / or_model arguments are retained for back-compat but ignored; DRDID uses logistic propensity and linear outcome regression internally.
• morie_did_bacon_decomposition() is a thin wrapper over bacondecomp::bacon and hard-errors if bacondecomp is not installed (the base-R timing-pair enumeration fallback has been removed).
• morie_did_chaisemartin_dhaultfoeuille() is a thin wrapper over DIDmultiplegt::did_multiplegt and hard-errors if DIDmultiplegt is not installed (the base-R switcher-comparison bootstrap fallback has been removed).
• morie_did_synthetic() continues to delegate to synthdid::synthdid_estimate (no change; synthdid is the only R implementation).
• morie_did_sensitivity_analysis() keeps its delta-bound CI sweep but the Rd now cross-references HonestDiD::createSensitivityResults_relativeMagnitudes as the reference implementation of Rambachan-Roth 2023 for event-study estimates.

The OLS-based wrappers (morie_did_2x2, morie_did_repeated_cross_section, morie_did_triple_difference, morie_did_continuous_treatment, morie_did_fuzzy) continue to use the in-package .morie_did_ols_robust_se helper because the specs are simple OLS / 2SLS regressions and a CRAN dependency for trivially-short OLS would be a regression. The same helper is reused by morie_did_wild_cluster_bootstrap, which remains base-R by design (fwildclusterboot is GitHub-only; see 0.9.5.12 NEWS).

Aggregators that consume DiD output and produce rmorie-specific tables (morie_did_aggregate_gt_att, morie_did_staggered, morie_did_parallel_trends_data, morie_did_test_parallel_trends, morie_did_placebo_test_*, morie_did_heterogeneous, morie_did_diagnostics) are unchanged.

Net: R/did.R shrinks from 1,719 to 1,463 LOC (-256, -15%); all 22 morie_did_* exports preserved; result-list shape is unchanged so downstream callers see the same fields.

DESCRIPTION: adds DRDID, HonestDiD, DIDmultiplegt to Suggests (fixest, did, bacondecomp, synthdid were already listed).

R/matching.R rewrite - delegate to MatchIt / cobalt / WeightIt

The matching subsystem (~2,183 LOC) has been rewritten to forward to the canonical CRAN packages instead of carrying ~950 LOC of base-R fallback code:

• morie_matching_nearest_neighbor(), morie_matching_exact(), morie_matching_cem(), morie_matching_mahalanobis(), morie_matching_optimal_pair(), morie_matching_full(), morie_matching_subclassify(), morie_matching_variable_ratio() are now thin wrappers over MatchIt::matchit() and now hard-error if MatchIt (or its optional optmatch back end) is not installed.
• morie_matching_genetic() is a thin wrapper over Matching::GenMatch() + Matching::Match() and hard-errors if Matching is missing.
• morie_matching_entropy_balance() is a thin wrapper over WeightIt::weightit(method = "ebal") (or ebal::ebalance as a fallback) and hard-errors if neither is installed.
• morie_matching_balance(), morie_matching_balance_table(), morie_matching_love_plot_data() are kept (they return the morie_balance_result shape downstream MRM code consumes) but the Rd cross-references now point users at cobalt::bal.tab() / cobalt::love.plot() for richer balance reporting.
• morie_matching_cardinality() keeps its iterative-caliper heuristic; the Rd now cross-references designmatch::cardmatch() for the exact mixed-integer-programming alternative.

The carceral-domain helpers (morie_matching_att_matched / ate_matched / atc_matched, morie_matching_abadie_imbens_se, morie_matching_rosenbaum_bounds, morie_matching_doubly_robust, morie_matching_multi_treatment, morie_matching_longitudinal, morie_matching_quality, morie_matching_overlap, morie_matching_estimate_propensity / _trim_propensity / _common_support) are unchanged - they encode rmorie-specific output shapes (morie_match_result, morie_te_result) that the MRM / SIU / OTIS code paths depend on.

Net: R/matching.R shrinks from 2,183 to 1,586 LOC (-597, -28%); all 27 morie_matching_* exports preserved; behaviour-compatible for callers that already have MatchIt installed (which is the case for all matching tests in the rmorie suite).

DESCRIPTION: adds cobalt, designmatch to Suggests.

Breaking - CRAN-equivalent functions removed

To reduce code duplication with established CRAN packages and address rOpenSci feedback on fn_call_network_size, the following functions have been removed in favour of their well-maintained CRAN equivalents:

Removed from rmorie	Use instead
cohens_d / morie_cohens_d	effectsize::cohens_d
cramers_v / morie_cramers_v	effectsize::cramers_v
eta_squared / morie_eta_squared	effectsize::eta_squared
hedges_g / morie_hedges_g	effectsize::hedges_g
morie_effective_sample_size	posterior::ess_basic or coda::effectiveSize
morie_find_project_root	here::here() or rprojroot::find_root()
fleiss_kappa	irr::kappam.fleiss
kruskal_wallis	stats::kruskal.test (base R)
shapiro_wilk	stats::shapiro.test (base R)
anderson_darling	nortest::ad.test
jarque_bera	tseries::jarque.bera.test

Install the replacements with install.packages(c("effectsize", "irr", "nortest", "tseries", "here")).

morie_two_sample_t_test() / morie_chi_square_test() / morie_anova_one_way() now return list fields named cohens_d / cramers_v / eta_squared (was morie_cohens_d / morie_cramers_v / morie_eta_squared); the computation is inlined and unchanged. Internal callers of morie_find_project_root() now go through a private .morie_project_root() wrapper around here::here().

DESCRIPTION: adds here to Imports; adds effectsize, irr, tseries to Suggests.

rmorie 0.9.5.12 - 2026-05-25

CI: drop fwildclusterboot (pak recursive Remotes unreliable) (3MMM.40c)

• Removed fwildclusterboot from Suggests and removed the .morie_did_have_fwildboot() helper + the if (.morie_did_have_fwildboot()) { fwildclusterboot::boottest(...) } branch in morie_did_wild_cluster_bootstrap(). The function now goes straight to the base-R Rademacher/Webb wild-cluster bootstrap (which already existed as the fallback and mirrors the Python implementation; no math change for any caller).
• Reason: pak’s resolver does not reliably recurse through a Remote’s own Remotes. 3MMM.40 added s3alfisc/fwildclusterboot and 3MMM.40b added s3alfisc/summclust, but the resolver still reported summclust: Can't find package called summclust – so the recursive-Remote pattern is structurally fragile. Following the same “drop optional CRAN-archived/GitHub-only deps” pattern used for rdd in 3MMM.40.
• Remotes: now lists only synth-inference/synthdid, which has no GitHub-only transitive Imports.