Refreshed the downloaded engine binary (new release 0.1.4). Adds lavaan’s y ~*~ y (delta / unit-liability) operator for ordinal models: the model-implied covariance is diagonal-standardised on the marked variables so their thresholds become free — the smooth equivalent of OpenMx’s mxConstraint(diag2vec(Vb) == 1), compatible with the trust-region optimiser and both variance parameterisations. This makes the strictly-contingent multi-group ordinal cascade (e.g. a three-stage CCC twin model) identifiable where fixed thresholds left the liability scale free. Standardisation runs on the F64 GPU kernel (−2logL matches CPU exactly); the analytic gradient carries the exact VJP (verified against finite differences).
Multi-group ordinal fits now report standardised-estimate SEs (stdSE, previously returned as zeros) via the delta method over the identified subspace, so scale-invariant standardised estimates get finite, meaningful SEs even when raw-scale parameters sit on an unidentified scale ray; and the scale-free Newton-decrement convergence value is now reported on this path (previously NaN).
Standard errors under the direct-symmetric variance parameterisation are no longer NaN on ill-conditioned-but-positive-definite information matrices (nearest-SPD Cholesky inverse). Estimates and SEs continue to match OpenMx. Run fastsem_update() after upgrading to pull the new binary.
fastsemR 0.1.3
Engine update (2026-09)
Refreshed the downloaded engine binary (new release 0.1.3). Fixes an intermittent crash in the definition-variable algebra-FIML path (a thread-safety bug in the parallel gradient) and improves the optimiser’s convergence test: it now stops on a scale-free Newton decrement (√λ² ≈ remaining move in SE units, invariant to how variables are scaled) rather than a unit-dependent gradient norm, so well-conditioned models converge in fewer iterations while weakly-identified directions are reported honestly. Fit objects now carry newtonDecrement, a convergence-quality value. Estimates and standard errors continue to match OpenMx. Run fastsem_update() after upgrading to pull the new binary.
fastsemR 0.1.2
Engine update (2026-09)
Refreshed the downloaded engine binary (new release 0.1.2) with the latest fastsem engine. Ordinal FIML now runs through a single estimation path shared by single- and multi-group models, so features compose: vartype: direct, definition variables, and grouping now all run on the GPU for ordinal models (multi-group ordinal previously forced definition variables onto the CPU and rejected vartype: direct). Estimates and standard errors continue to match OpenMx. Run fastsem_update() after upgrading the package to pull the new binary.
fastsemR 0.1.1
Engine update (2026-09)
Refreshed the downloaded engine binary (new release 0.1.1) with the latest fastsem engine. High-dimensional FIML now parallelises to models with up to 64 observed variables (previously 32) for both the fit and the standard-error Hessian; the per-pattern gradient’s memory is bounded, so large models that previously exhausted memory now complete; and a latent crash in the standard-error step for models above 32 observed variables is fixed. The optimiser dispatches a second-order trust-region step scaled to the parameter count. Estimates and standard errors continue to match OpenMx. Run fastsem_update() after upgrading the package to pull the new binary.
The downloaded engine is now cached under a release-versioned filename, so a future package upgrade fetches the matching binary automatically on first use — no manual fastsem_update() needed. Any legacy version-less cache file is superseded.
fastsemR (development version)
Engine update (2026-09)
Refreshed the downloaded engine binary (release 0.1) with the fixed fastsem engine: multi-group ordinal FIML now works (twin ACE / CCC models were previously fit as continuous-on-codes, collapsing the genetic contrast into C); Newton-TR is the primary optimiser for ordinal and SSM/Kalman fits; and the Linux R library is now GPU-accelerated by default (OpenCL, with graceful CPU fallback). Estimates and standard errors match OpenMx. Run fastsem_update() to pull the new binary.