v0.10.1

Release v0.10.1

v0.10.0

Release v0.10.0

v0.9.9

Fixed

• viz PNG/PDF rendering — render_to_file() for PNG and PDF formats was silently broken since 0.9.8: the png and pdf features of ns-viz-render were not propagated through the native-render feature gate in ns-py and ns-cli, causing #[cfg(feature)] guards to compile out all raster/PDF code paths. SVG output was unaffected.

v0.9.8

Added

• ns-cli-py package — standalone nextstat CLI Python package (bindings/ns-cli-py) with pyproject.toml and __main__.py entry point.
• nextstat-nlp remote workflow runner — scripts/nlp/nlp_workflow_remote.sh for reproducible NLP pipeline verification on remote hosts (ONNX cache, regimen export, backend cache, internet mocks).
• nextstat-nlp --skip-nextstat-demo — flag to skip demo pipeline in automated runs; --quiet mode for suppressing verbose output.
• nextstat-nlp regimen fallback tests — tests/test_regimens_fallback.py for graceful degradation when backend unavailable.

Fixed

• Validation-pack pharma determinism — PQ references stabilized for reproducible cross-run validation.
• pip extras — added missing [all] extra, documented all pip install extras in README and quickstart.

Changed

• Python 3.11+ requirement — documented minimum Python version in installation guide.
• Benchmark artifact sanitization — replaced internal hostnames and server paths with generic placeholders in all public benchmark docs and JSON artifacts.

v0.9.7

Added

• PD models Python API — emax_predict(), emax_nll(), sigmoid_emax_predict(), sigmoid_emax_nll(), idr_simulate(), idr_nll() exposed to Python. Emax and Sigmoid-Emax for dose-response, IDR Types I–IV for indirect response simulation and likelihood. IDR type specified as string ("inhibit_production", "stimulate_loss", etc.).
• EGARCH/GJR-GARCH Python API — egarch11_fit() and gjr_garch11_fit() exposed to Python via nextstat.volatility.egarch() and nextstat.volatility.gjr_garch(). Full EGARCH(1,1) and GJR-GARCH(1,1) MLE with conditional variance and sigma output.
• nextstat-nlp 0.2.0 — GLiNER2-based text-to-stats plugin for extracting structured data from clinical text. Three extraction pipelines: survival (extract_survival_records), informative priors (extract_prior_candidates), and dosing regimens (extract_regimens). Three backend adapters: GLiNER2 Torch, GLiNER2 ONNX, and zero-dependency heuristic (regex). Deterministic time/event/numeric parsers with unit normalisation (days, weeks, months, years, hours). Frozen dataclasses with SHA-256 text hashes and ProvenanceBundle audit trail. Covariates: age, dose, sex, stage, ECOG PS, weight. Golden-file reproducibility tests (57 tests). Benchmark harness and paper scaffold for pharma publication.
• nextstat-nlp 0.2.1 — Added optional mlx backend (Apple Silicon) via a small Swift CLI using GLiNER2Swift (MLX/Metal). Added device= (Torch) and providers=/num_threads= (ONNX Runtime) passthroughs to pipelines and benchmark harness. Benchmark harness now reuses a single backend instance to measure warm throughput.
• Per-subject dosing for FOCE/SAEM/VPC/GoF — all population PK estimation and diagnostic functions now accept doses (per-subject dose vector) instead of a single scalar dose. Enables proper mixed-effects modeling of clinical data with varying doses across subjects. Python API: doses=[4.0] (broadcast) or doses=[4.0, 3.5, 5.0] (per-subject).
• CDISC .xpt (SAS Transport v5) reader/writer — read_xpt() / write_xpt() for FDA-submission SDTM/ADaM datasets, pure Rust with IBM 370 floating-point conversion. xpt_to_nonmem() auto-detects SDTM/ADaM columns (USUBJID, AVAL, ATPTN, etc.) for direct PK modeling.
• NONMEM parity whitepaper and benchmark suite — formal validation of FOCE/FOCEI and SAEM against NONMEM reference values on Theophylline (12 subjects, real data), Warfarin (32 subjects, synthetic), and Phase I IV bolus (24 subjects, 2-cpt synthetic). SAEM recovers all parameters within 7% of reference. Whitepaper: docs/whitepapers/nonmem-parity.md. Rust tests: cargo test -p ns-inference --test nonmem_parity. Python benchmark: scripts/benchmarks/bench_nonmem_parity.py.
• NUTS Pathfinder dense metric initialization — init_strategy="pathfinder" with metric="dense" now uses the full MLE covariance matrix (from the Hessian) to initialize a DenseCholesky metric, giving warmup a head-start on the dense preconditioner. Falls back to diagonal when Cholesky fails or metric="diagonal".
• NUTS dense metric reporting — sample() result now includes metric_type ("diagonal" or "dense"), mass_diag, and inv_mass_matrix (full inverse mass, row-major, only for dense) in sample_stats.
• NUTS vs CmdStan benchmark suite — scripts/benchmarks/bench_nuts_vs_cmdstan.py comparing ESS/s, R-hat, and divergence rates on StdNormal 10D, Eight Schools, and GLM Logistic models.
• LAPS GLM model expansion — GPU sampler now supports "glm_linear", "glm_poisson", "glm_negbin" (negative binomial), and "glm_composed_logistic" (logistic + random intercept NCP) in addition to the existing "glm_logistic". Full CUDA (f64) and Metal (f32) kernel support. Poisson and NegBin accept optional offset in model_data. NegBin estimates dispersion parameter log_alpha. ComposedGlm uses non-centered parametrization for group-level random intercepts.
• SCM Python API — nextstat.scm() exposes Stepwise Covariate Modeling (forward selection + backward elimination) via Python. Accepts observation-level covariate matrices, auto-centers at median, supports power/proportional/exponential relationships. Returns full audit trace with per-step ΔOFV, p-values, coefficients, and final model parameters. TypedDict return types (ScmResult, ScmStepResult).
• IQ/OQ/PQ validation protocol v2.0.0 — docs/validation/iq-oq-pq-protocol.md with 85 numbered test cases across 11 OQ subsections (PK analytical, MLE, FOCE, SAEM, GOF, VPC, SCM, Error Models, LLOQ, NONMEM I/O, Artifacts), PQ cross-validation with NONMEM on Theophylline/Warfarin reference datasets, traceability matrix, and deviation report template. 21 CFR Part 11 / EU Annex 11 / ICH E6(R3) compliant.
• 3-compartment PK models — IV bolus and oral absorption with analytical gradients.
• Bioequivalence testing — TOST (average BE), reference-scaled (RSABE), power analysis, and sample size calculation.
• MAP estimation — Maximum A Posteriori estimation with Hessian-based standard errors.
• Monte Carlo clinical trial simulation — simulation engine with PTA (Probability of Target Attainment).
• Dose optimization via PTA grid search — systematic dose finding using PTA target grids.
• ODE-based PK models — transit compartments, Michaelis-Menten elimination, and TMDD (target-mediated drug disposition).
• Multi-compartment FOCE/SAEM — 2-compartment IV and oral population PK estimation.
• Allometric scaling and covariate modeling for FOCE — weight-based allometric exponents and covariate effects in population PK.
• IQ/OQ/PQ validation pack — GxP-ready validation package for pharma environments (21 CFR Part 11 compliant).
• Pharma tutorials — PK modeling guide, NONMEM migration guide, and survival analysis guide.
• Pharma benchmark suite — NS vs NONMEM/Monolix/nlmixr2 comparative benchmarks.
• FOCE omega_fixed parameter — nlme_foce(omega_fixed=[False, False, True]) fixes specific random effects to their initial values during estimation (e.g. near-fix Ka inter-individual variability).
• FOCE diagonal_omega parameter — nlme_foce(diagonal_omega=True) forces a diagonal Ω matrix, disabling off-diagonal correlation estimation.
• 3-compartment SAEM — nlme_saem(model="3cpt_iv") and nlme_saem(model="3cpt_oral") for 6- and 7-parameter population PK estimation via SAEM. Model dispatch also supports "2cpt_iv" and "2cpt_oral".
• FO / ITS / IMP estimation methods — nlme_foce(method="fo"|"its"|"imp") adds First Order, Iterative Two-Stage, and Importance Sampling MC-EM estimation for all PK models (1/2/3-cpt × IV/oral). IMP uses true IS-EM: MAP + importance sampling from N(η̂, scale²Ω), weighted Ω M-step, ESS/max-weight diagnostics.
• SAEM covariance step — nlme_saem() now computes the covariance step automatically (R/S matrices, sandwich SE, RSE%, condition number) in the returned result dict.
• NPDE diagnostics — pk_npde() Python API for simulation-based Normalized Prediction Distribution Errors. Returns per-observation NPDE/PD, summary statistics, Shapiro-Wilk normality test.
• 3-compartment multi-dose superposition — DosingRegimen::conc_3cpt_iv() and conc_3cpt_oral() with analytical eigenvalue decomposition for multi-dose 3-compartment PK.
• SAEM covariate modeling — nlme_saem(covariates=[...]) accepts per-subject covariate dicts with power, exponential, proportional, and categorical relationships for all model types.
• SAEM convergence diagnostics — return_theta_trace=True stores per-iteration θ trace, Geweke z-scores, and relative parameter change in the saem sub-dict.
• Bootstrap NLME — bootstrap_nlme() computes nonparametric bootstrap SE and percentile/BCa confidence intervals for SAEM population PK parameters.
• Pharma IQ/OQ/PQ auto-runner — python tests/pharma_validation/runner.py auto-executes 84 test cases (5 IQ, 54 OQ, 25 PQ) across all pharma verticals: PK analytical correctness, error models, LLOQ handling, MLE, FOCE, SAEM, GOF, VPC, SCM, NONMEM I/O, bioequivalence, trial simulation, Theophylline reference fits, reproducibility, and timing benchmarks. Integrated into validation-pack/render_validation_pack.sh for one-command GxP validation pack generation.

v0.9.6

Release v0.9.6

v0.9.5

Release v0.9.5

v0.9.4

Release v0.9.4

v0.9.3

Release v0.9.3

v0.9.2

Release v0.9.2