Added end-to-end support for the unidimensional three-parameter logistic model with logistic scaling constant D = 1. The shared model engine, reliability reducers, item generation, EQC, SAC, response simulation, and S3 result methods now carry item lower asymptotes.
Added fixed, beta, and uniform guessing generators through guessing_params; custom forms use custom_params$guessing. Values must satisfy 0 <= guessing < 1.
Clarified parameter names: item guessing is stored as guessing (mathematical g_i), while c/c_star is the global discrimination scale. Calibration applies lambda_i(c) = c * lambda_i0 and never rescales guessing.
The guessing = 0 3PL path is computationally identical to the 2PL path. Probability and Fisher-information calculations use stable log-domain kernels for extreme linear predictors.
Calibration and estimand contracts
EQC now scans the complete user-supplied interval on log(c), records all resolved roots and local extrema, and applies an explicit root policy. EQC remains a fixed-form, average-information (info/tilde) calibrator.
SAC now runs a topology preflight before Robbins–Monro updates, constrains the trajectory to the selected branch, and performs an independent final evaluation. New controls are appended after the historical argument prefix: root_policy, preflight_controls, and evaluation_controls.
SAC explicitly distinguishes resample_items = FALSE (fixed-form estimand) from resample_items = TRUE (item-superpopulation estimand). Results store calibration/evaluation design metadata, achieved distributions and standard errors, branch diagnostics, and RNG provenance.
EQC result warm starts are accepted only for matching metric/model/item count and the same fixed-form contract. Same-seed SAC replay and caller RNG restoration are enforced by independent random-number streams.
check_feasibility() and rho_curve() use the shared topology scanner rather than assuming global monotonicity. Multiple roots remain possible for adversarial or poorly covered designs.
Reliability, migration, and performance
Reliability utilities now accept 3PL guessing vectors and optional quadrature weights, and return numerical/estimand diagnostics on request. The default scalar/list shapes remain backward compatible.
New result schema fields include guessing_vec, schema_version, item_scope, estimand_signature, and design_signature. Legacy Rasch/2PL results without guessing are interpreted lazily as zero-guessing designs and are not mutated. Incomplete legacy spc_result objects now fail with an actionable instruction to rerun sac_calibrate().
spc_calibrate() and compare_eqc_spc() remain deprecated forwarding aliases. The historical 18-argument prefix of sac_calibrate() is unchanged; new controls were appended to preserve positional calls.
Reliability reduction automatically chunks transient item-information matrices while preserving exact scalar results. Matrix-returning internal kernels retain their previous full-output contract.
Validation boundaries and known limitations
Analytic identities, finite-difference derivatives, extreme-predictor cases, independent holdouts, stochastic sensitivity runs, and external TAM probability/information calculations were used to validate the D = 1 3PL implementation.
TAM 3PL validation is an EAP-only diagnostic in the validated environment. TAM::tam.wle() did not support the fitted tam.mml.3pl object, so IRTsimrel does not claim 3PL WLE support. The public compute_reliability_tam() helper continues to provide WLE and EAP output for its documented Rasch/2PL paths.
Analytic information/MSEM reliability and fitted-score EAP reliability are different estimands; equality is neither assumed nor used as an acceptance criterion.
Population MSEM reliability is non-integrable for the package’s built-in Student-t heavy-tail distribution. SAC raises a classed error for that combination; use information reliability or explicitly define a finite, truncated empirical estimand.
The 3PL release remains unidimensional. Multidimensional models are outside the 0.3.0 scope.
eqc_calibrate() now supports "info"/"tilde" targets only. Direct MSEM-based targets ("msem"/"bar") are rejected because the MSEM objective can be non-monotone under EQC root-finding; use sac_calibrate() for direct MSEM targeting.
simulate_response_data() first parameter renamed from eqc_result to result and now accepts both eqc_result and sac_result objects.
Default item_source for eqc_calibrate() and sac_calibrate() changed from "irw" to "parametric" so core workflows do not require the optional external IRW integration.
New Features
sac_calibrate() now stores calibrated item parameters (beta_vec, lambda_base, lambda_scaled, items_base, items_calib, theta_quad) in the result object, enabling direct use with simulate_response_data().
simulate_response_data() now records calibration and simulation provenance, including result class, metric, calibration status, item source/design, calibration call, seed, sample size, and latent settings.
compare_eqc_sac() now returns richer diagnostics, including achieved reliability differences, metric/model/test-length metadata, and EQC/SAC status fields.
Completed a vignette overhaul, expanding from 6 vignettes (~3,400 lines) to 12 source vignettes (~10,200 lines) organized as getting-started material, two applied/methodological tracks, and a reference article:
Saved objects from IRTsimrel 0.1.x whose class is literally "spc_result" may not dispatch every S3 helper method directly. New calls to spc_calibrate() return "sac_result" objects through the migration alias, and simulate_response_data() plus comparison helpers still accept legacy "spc_result" objects. Recreate legacy calibration objects with sac_calibrate() if direct summary(), plot(), coef(), or predict() dispatch is needed.
Testing
Expanded the testthat suite across core functions, S3 methods, edge cases, and error handling.
Infrastructure
Fixed .Rbuildignore to exclude dev/, log/, pkg.lock, and release-process helper files from source package builds.
Fixed DESCRIPTION: removed LazyData, Remotes; added graphics, grDevices, utils to Imports; added MASS to Suggests.
