NEWS

---

sandtablR 0.0.0.9000

• Standardize the named v5 package logo across README, documentation pages and available application headers, with matching light and dark artwork.

First development version.

• R 4.5.0 or later is required. Suggested modelling packages need it (glmmTMB depends via pbkrtest and doBy on Deriv, which requires R 4.5).

Check and portability fixes

• keras3 and parallelly are declared in Suggests. Both were already used when present.
• The documentation and NAMESPACE are regenerated. The transformer engine (sdt_transformer() and its fit and predict plumbing) is now documented, and sdt_read_gold_standard(), sdt_apply_gold_standard(), sdt_triage_accuracy(), sdt_fit_triage_error(), sdt_simulate_triage() and sdt_error_fidelity() are exported, as their documentation already declared.
• The format() and print() methods for sdt_simulated_triage and sdt_triage_error_model objects, and the tibble::tbl_sum() methods for sdt_error_fidelity, sdt_gold_standard and sdt_triage_accuracy tables, are now registered, so these objects print with their own headers everywhere, not only from inside the package.
• sdt_export_simulation() returns the archive path as it was passed, with only ~ expanded. It returned a normalised spelling of the same file, which differs from the caller's on macOS (/private/var) and Windows.
• sdt_workflowset() skips the transformer, and names it in skipped, when keras3 is installed but has no working backend; the warning says how to install one. Its fit used to abort and take every other engine's row down with it. The transformer is now fitted after the other engines, so Python is not loaded before their parallel workers start. The workbench's skip notice gives each skipped engine its own reason.

Realignment with the source publication

The reference distributions were re-read against the open-access paper and extended from four pinned quantities to the full published set. Several package defaults that contradicted the patient card reproduced in the source's Fig. 1 were corrected. These are breaking changes; the package is pre-1.0 and the previous values were never published ones.

• Life points now run 1–6, the scale printed on the version 2.0 patient card, rather than 1–12. Band boundaries move with them.
• sdt_therapy_levels() gains "none", the card's "no operation" option. A casualty carrying it is never routed to theatre.
• New sdt_diagnostics() vocabulary: "efast", "xray", "ct". X-ray was absent, though the source reports its share.
• sdt_board_areas() gains "green_area", the board's green treatment area. Green casualties are cared for there rather than being sent to a ward bed.
• sdt_reference() gains version, serving both evaluated versions of the exercise, and a design part carrying the exercise as specified: 90 minutes, 20 patient cards, four card rounds, the team roles, and the nine-area board with its layout.
• summary.game_run() and sdt_plot_end_state() gain version, and now fold the board onto the eight locations the source reports rather than comparing three and lumping the rest. The two foldings that are modelling choices rather than published facts are recorded in sdt_reference("defaults")$end_state_map.
• Casualties arrive in four tranches at fixed points in the game, as the game leader plays the cards, rather than as a smooth stream. A smooth stream empties the screening area long before the horizon and cannot reproduce the published share standing there at the end.
• The engine was recalibrated against all eight published locations by data-raw/calibrate.R, a re-runnable parallel search. The previous calibration matched three locations while holding roughly twice the published share of casualties in the emergency department; the error was inside the cell that absorbed it.

Bugs this surfaced

• summary.game_run() dropped casualties that generated no events, shrinking the denominator and making the source's "not yet seen" category unreachable. Every casualty is now accounted for in every replication.
• sdt_decision_paths() dropped the same casualties. They now carry a single-step "not_seen" path.
• A casualty queueing for a screening slot recorded no event, so it counted as never seen. The screening area is now entered on arrival, before the seize, which is also what makes the recorded time there include the wait.
• sdt_simulate(horizon = Inf) produced NaN arrival times.
• A cohort smaller than the number of arrival tranches aborted.

The board, drawn

• New sdt_plot_board() draws the nine board areas laid out as in the source's Fig. 1, with every casualty as a point in the area it had reached at a given moment. Pass interactive = TRUE for \pkg{ggiraph} tooltips carrying the patient card.

The data model

• sdt_patient_card(), sdt_cohort() and sdt_scenario() form the data model: a casualty, an ordered set of them, and a board to run them across.
• sdt_cohort() returns a validated tibble subclass. Validation is vectorised over columns, so a bad import reports every offending row and column at once instead of aborting on the first.
• Controlled vocabularies for sighting category, therapy concept, board area, injury region and sex, all from one source of truth.
• format() methods for patient_card, scenario, game_run and sdt_model, with print() delegating to them so the two cannot drift.
• as_tibble() and autoplot() methods for a scenario; sdt_as_patient_cards() for the reverse direction.

Reference distributions

• sdt_reference() returns the marginals the package calibrates against, loaded lazily from inst/extdata and cached.
• The list deliberately separates published (attributed to Achatz et al. 2023) from defaults (modelling choices made where the source pins no value). A calibration test that mixed the two would be checking the model against its own assumptions.
• sdt_reference_scenario() and sdt_reference_horizon() build the board the engine was calibrated on, so the calibration claim can be re-run.

Synthetic data

• sdt_synthetic_cohort() gains a simstudy backend drawing correlated age, injury severity and life points, with the sighting category derived from acuity by rule. The base backend is the always-available reference implementation and is pinned by a golden test at seed 1.
• sdt_synthesise() produces a releasable replica of a real seed cohort with synthpop's sequential CART models. Identifiers are regenerated rather than synthesised, since they carry row identity rather than structure.
• sdt_synthesis_utility() scores a replica against its seed per variable, so it can be judged before release.
• sdt_add_decisions() simulates the category a participant assigns, using the published per-category hit rates, with miscalls weighted toward neighbouring categories.

Simulation

• sdt_simulate() routes casualties by sighting category and therapy concept rather than sending everyone through theatre, with diagnostics, theatre and critical-care probabilities from the reference defaults.
• It records a per-area event log built from monitored entry attributes. Release timestamps would have misplaced every casualty still being treated when the clock stopped, which is precisely the population the end-of-game distribution is about.
• events carries both duration and service_time. They differ whenever a casualty is finished but waiting for a bed downstream, and conflating them made theatre times look like whatever the downstream bottleneck was.
• Three dispositions are terminal, and that is what reproduces the published end state: critical care is where a casualty stays, a ward bed is the end of the line, and a casualty worked up but not taken to theatre is still held in the emergency department. Routing them onward put roughly 70% of every cohort in the ward against a published 35%.
• summary() on a game_run reports the end-of-game location distribution, optionally beside the published reference.
• Replications can run concurrently via future.apply behind parallel = FALSE.

Decision paths

• sdt_decision_paths() returns the full ordered pathway per casualty per replication, in long or state-sequence-wide form.
• sdt_decision_consistency() is rebuilt on distance rather than modal agreement. The previous measure compared each casualty's single recorded area with itself and was therefore always exactly 1. It now offers mean pairwise edit distance across replications, or normalised longitudinal entropy. A single path scores NA, not a perfect 1: there is nothing to compare it with.
• sdt_seqdef() bridges to TraMineR over a fixed board-area alphabet, so runs stay comparable; sdt_seq_cluster() groups pathways by optimal-matching distance and partitioning around medoids.
• sdt_plot_sequences() draws index and state-distribution views in ggplot2 rather than base graphics, so they theme like every other plot.

Machine learning

• sdt_fit() is a full tidymodels path: preprocessing recipe, stratified resamples, tuning grid, and a final fit on a held-out split. The split and resamples are kept, so reported numbers can be traced to what produced them.
• sdt_workflowset() screens every installed engine onto one leaderboard and names what it skipped.
• New sdt_engines(), sdt_metric_set(), sdt_metrics(), sdt_predictions(), sdt_roc() and sdt_consistency_score().
• Models can target the participant's call as well as the card's true value. The true category follows from life points by construction, so every engine scores near 1 on it and the glass-box versus black-box comparison is empty.

Explainability

• sdt_importance(), sdt_profile() and sdt_shap() wrap DALEX, all returning ggplot2 objects and all degrading with an actionable message when DALEX is absent.

The workbench

• sdt_run_app() launches a six-panel workbench: scenario, simulate, annotate, model lab, explain and sequences, plus a one-click self-contained HTML report of the session.
• The app's modules are ordinary package functions in R/ rather than scripts under inst/. Code under inst/ is invisible to R CMD check, lintr and covr, so an app kept there is the one part of a package nothing verifies. Every module is now driven headlessly by shiny::testServer().
• Annotation edits are validated against the vocabularies and written back into the shared scenario; an invalid edit is refused and explained.
• Every optional package is guarded, and each panel falls back to a static plot or a plain table rather than failing.

Fixes to the initial scaffold

• sdt_simulate() passed at = to simmer::add_generator() as a named argument rather than as at(), so no simulation ran at all.
• Two cli format strings would have aborted at format time: one pluralising with no quantity, one interpolating a dot-prefixed expression that cli reads as a style name. Both sat on error paths, where they would not have been noticed. Static tests now guard both patterns, along with line length, library() calls and bare T/F.
• ventilated is logical, and recipes counts logicals as neither nominal nor numeric, so it slipped past both selectors and reached glmnet and xgboost as a column they could not put in a numeric matrix.
• tune() markers must appear literally in a model specification: passed through a helper they evaluate to the same object, but parsnip inspects the expression, so the parameter set came back empty and the placeholder reached the engine.
• DALEX::loss_cross_entropy() looks the observed class up by name in the probability matrix, and tidymodels names those columns .pred_red; every lookup was out of bounds until the prefix was stripped.
• Annotating a row updated the scenario, which fired the observer that reset the annotation count, so the "rows annotated" badge always read zero.